The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV [1 ed.] 0128189487, 9780128189481

Table of contents :
Cover
The Interdisciplinary Handbook of Perceptual Control Theory
Copyright
Contributors
Author biographies
Preface
Section A: Why do we need perceptual control theory?
1 - The world according to PCT
2 - Understanding purposeful systems: the application of control theory in engineering and psychology
Understanding control: control theory
Doing reverse engineering from a forward engineering perspective
Perceptual control theory: control of perception
Reverse engineering a robot
Testing for controlled variables: reverse engineering living control systems
Conclusion
References
3 - The crisis in neuroscience
Paradigm and crisis
Behavior is not solely determined by neural output
Calculation problem and control
Control of input
Misunderstanding control
Observer's bias
Input/output analysis and behavioral illusion
Case study 1: Sherrington's analysis of the reflex
Case study 2: sensorimotor transformations: from flies to monkeys
The proper study of behavior
Conclusions
References
4 - When causation does not imply correlation: robust violations of the faithfulness axiom
Introduction
Preliminaries
Causal inference
Zero correlation between a variable and its derivative
Control systems
Example 1
Example 2
Example 3
A digression on disturbances
Example 4
Example 5
Summary
Current causal discovery methods applied to these examples
The fundamental problem
Conclusion
Appendix: Sufficient conditions for zero correlation between a function and its derivative
References
Section B: Models of brain and behavior
5. - Unraveling the dynamics of dyadic interactions: perceptual control in animal contests
Introduction
Lessons from fighting
Some thoughts on methodology
Some thoughts on interpretation
Problems needing novel solutions
Conclusion: where to go from here?
Acknowledgments
References
6 - How the brain gets a roaring campfire: Structuring for perceptual results
Introduction
Features of perceptual control
How to structure a goal
Organizational rationale
Modeling with falsifiability in mind
Predictive proposals of PCT
Hierarchical classes of perception
How to maintain commensurate matching
Getting a workable address
Cortical implementation
PCT reference signals and HTM “name” cells
Neocortical layers in action
Modeling the temporal flow of experience
How to build contextualized beliefs
First chunk: detect co-occurrences
Second chunk: notice lateral sequences
Third chunk: compile into a “name”
Fourth chunk: pool the evidence
Fifth chunk: provide a timing gate
Sixth chunk: output the interpretation
Compatible cortical mind-sets
Constructing invariants: a summary
Addressing reference signals
Status of the brain's campfire
Remaining challenges
Acknowledgments
References
7 - How the brain gets a roaring campfire: Input and output functions
Introduction
Prolegomenon: The what and the how
A hierarchy of perceptions
Open-loop methods to study closed-loop features
Constructing visual input functions: This way in
Signal-to-noise sensitivity
Early forms of contrast
On and off signals
Figure-ground contrasts
Motion and transition
Modeling with spatial frequencies
Handling stark edges
Objects recognized by harmonic composition
Sensorimotor coordination: This way out
Perceiving objects and their relational flux
Blended frames of reference
Summary and implications
So how does the brain get that campfire?
Acknowledgments
References
8 - The phylogeny, ontogeny, causation and function of regression periods explained by reorganizations of the hierarchy of perc ...
Introduction
Evolution
Development
Causation
Function
Discussion
The field of ethology is ready for a paradigm shift toward PCT
PCT informs human developmental studies beyond the sensorimotor stage
PCT and understanding the evolution of human cognition
Summary of the discussion
References
Section C: Collective control and communication
9 - Social structure and control: perceptual control theory and the science of sociology
Overview of my argument
A control-theory analysis tool kit
Atenfels: physical components of feedback loops
Matching atenfels to perceptions
Atenfels and the facilitation of feedback paths
The mirror world
Collective control processes
Cooperation and conflict
Giant virtual controllers: the social power of numbers
Human activities as feedback paths
Comparing human activities to physical artifacts
Collective control and levels of perception
Protocols: structural frameworks for dyadic interactions
Atenfels and human interaction
Symbols and meaning from the PCT perspective
Associative memory and the organization of the brain
Meaning, symbols, language, and culture
Collective control networks and social groups
Scale and stability of collective control networks
Multiple overlapping collective control networks
The anatomy of social structures
The four main types of collective control networks in social structures
Embedding and interleaving of social structures
A conceptual map of a social Structure's collective control networks
Three mechanisms of social stability
Collective control and social structural levels
Stabilization of physical environments
Cultural environments and social structural levels
High and low culture and layers of perception
The two faces of social structure
The dynamics of social structures
Work and social structures
Creating and maintaining stable feedback paths
Work and resources
Socialization of new members of social structures
Redundancy of feedback paths and reorganization of perceptual hierarchies
Learning by imitation and play
Differences among types of new members of social structures
Innovation and change in social structures
Mismatches between self and living environment
Competition, innovation, and social structures
Consequential and inconsequential innovations
Other sources of social change
Discussion
Acknowledgments
References
10 - Perceptual control in cooperative interaction
Introduction
History: Layered Protocol Theory (LPT)
Control and perceptual control
Elements of control
The Powers hierarchy of control
Perception of uncertainty
Protocols
Classes of protocol
Feedback loops and control loops
Generalized feedback loops
Atenfels
Four-element loops
Protocol representation
A protocol example
The General Protocol Grammar: introduction
Extending the GPG: error correction
Representing problems
Protocols proper
Protocol function: control of belief
Protocol function: three propositions
Protocol function: R-Display and interrupt
Protocol function: moving through the GPG
The test for the controlled variable in a protocol
Triadic protocols
Protocol levels
Protocol as communication
Summary
References
11 - Language and thought as control of perception1
Introduction
Perceptual Control Theory (PCT)
Imagination
Memory
Levels of the perceptual hierarchy
Collective control
Methodological summary
Language perceptions
Phonemic distinctions
Disturbing the control of pronunciation
Words and morphemes
Word dependencies
Word selection
Word order
Variant shapes of words
Repetition and the construction of objective information
Knowledge in a subject-matter domain
Operator Grammar
Operator-argument dependencies
Selection and likelihood differences
Linear order
Reductions
Control of linguistic information
Common knowledge and discourse coherence
Information is constructed by parallel repetition
Sublanguage
Nonverbal perceptions and subjective meanings
Language variation and change
Linguistic information and information theoretic ‘information’
Consequences
Some consequences of empirical linguistics for PCT
Uses of information theory with PCT
Status of categories
Some consequences of PCT for empirical linguistics
Status of acceptability judgments
Status of metalinguistic sameness
Latency period
‘Mainstream’ views
Cognitive Psychology and Generative Linguistics
Connectionism versus computationalism
Computer languages and robots
Some epistemological considerations
Conclusion
References
Section D: Applications
12 - Perceptions of control theory in industrial-organizational psychology: disturbances and counter-disturbances
Organizational psychology and perceptual control theory
Established views and their discontents
Current state of PCT in IO
A history of PCT in organizational psychology
The emergence of PCT
Conceptual papers
Empirical work
A backlash
The 1990's
Conceptual work
Empirical work
The 2000s to the present
Conceptual work
Empirical work
The self-efficacy studies
Another backlash
Enter computational modeling
Thinking and learning with PCT
Discussion
The permeation challenge
Self-regulation in IO
Conclusion
References
13 - Method of Levels Therapy
Method of levels and perceptual control theory
Perceptual control theory (PCT)
A brief history of MOL therapy
Doing the MOL therapy two-step
Learning MOL
Why learn PCT and MOL therapy?
What is the evidence for MOL therapy effectiveness?
Future directions for MOL therapy research
Scientific validity
Conflict with assumptions
Accessibility
Conclusions
Therapy manuals on method of levels
Key links
References
14 - Robotics in the real world: the perceptual control theory approach
Introduction
Perception-based robotics
Models and Control
Purpose
Perceptual control
Hierarchical control
The power of hierarchies
What is behavior?
Modeling behavior
Comparing approaches
Model-based predictive control
Memory-based prediction
Behaviour-based robotics
Modeling dynamics
Discussion
References
15 - PCT and beyond: toward a computational framework for ‘intelligent’ communicative systems
Introduction
Good old-fashioned artificial intelligence
Behavior-based robotics
Artificial cognitive systems
Agent based modeling
Contemporary intelligent systems
Whither perceptual control theory?
Classic automatic control
Hierarchical perceptual control
The way forward?
Toward ‘intelligent’ communicative systems
Discussion
Summary and conclusion
Acknowledgments
References
Section E: Synthesis
16 - Ten vital elements of perceptual control theory, tracing the pathway from implicit influence to scientific advance
What was unique about PCT in 1960?
The dissemination and divergence of PCT
Psychotherapy
Reality therapy
Grawe's Psychological Therapies
Method of Levels (MOL)
Social, personality and occupational psychology
Carver & Scheier's self-regulation theory
Lord and Levy's (1994) Control theory
Vancouver's self-regulatory theory
Sociology
Affect control theory
Identity control theory
Perceptual control theory
Summary
Is there a convergence of other theories with PCT, or is it unique?
The ecological approach
Event coding
The empirical strategy of vision
Controlled versus automatic behavior
Free energy principle
Society of Mind
Interactive cognitive subsystems
Conway's model of autobiographical memory
Summary
Empirical tests of the unique features of PCT
Behavior is the control of perception
The test for the controlled variable
A perceptual control hierarchy
Reorganization of the perceptual control hierarchy
Rerouted perceptual memory as imagination
Demonstration of several specific levels of perceptual abstraction
A definition of consciousness
The use of symbols as ‘order-reduction representations’
Conflict
Resolution of conflict
Summary
Addressing the critiques of PCT
Summary
Limitations, open issues and future direction for PCT
The components of control: how are they formed?
Intrinsic systems and evolution
Neural components
Muscular mechanisms
Physical properties of anatomy, including sensors and effectors
Understanding feedback paths through the environment
Developing input functions
The operation of specific levels in the hierarchy
What is the role for ‘internal modeling’?
Potential extensions of the PCT architecture
Future research design
Chapter summary
Acknowledgments
References
Online materials
17 - How the brain gets a roaring campfire: Structuring for perceptual results
Introduction
Features of perceptual control
How to structure a goal
Organizational rationale
Modeling with falsifiability in mind
Predictive proposals of PCT
Hierarchical classes of perception
How to maintain commensurate matching
Getting a workable address
Cortical implementation
PCT reference signals and HTM “name” cells
Neocortical layers in action
Modeling the temporal flow of experience
How to build contextualized beliefs
First chunk: detect co-occurrences
Second chunk: notice lateral sequences
Third chunk: compile into a “name”
Fourth chunk: pool the evidence
Fifth chunk: provide a timing gate
Sixth chunk: output the interpretation
Compatible cortical mind-sets
Constructing invariants: a summary
Addressing reference signals
Status of the brain's campfire
Remaining challenges
Acknowledgments
References
18 - How the brain gets a roaring campfire: Input and output functions
Introduction
Prolegomenon: The what and the how
A hierarchy of perceptions
Open-loop methods to study closed-loop features
Constructing visual input functions: This way in
Signal-to-noise sensitivity
Early forms of contrast
On and off signals
Figure-ground contrasts
Motion and transition
Modeling with spatial frequencies
Handling stark edges
Objects recognized by harmonic composition
Sensorimotor coordination: This way out
Perceiving objects and their relational flux
Blended frames of reference
Summary and implications
Inhibit what you don't want, disinhibit what you do
Get the timing right
So how does the brain get that campfire?
Acknowledgments
References
19 - How the brain gets a roaring campfire: Thalamus through a PCT microscope
Introduction
First, know your equipment
Multifunctional comparators
A bank of thalamic comparators
Electrochemical on-off switches
Tonic versus burst oscillation: signaling error, heightening gain
Bi-directional comparators
Thalamic routing
Summary: thalamic control loop signals
Make it a roaring campfire
A final survey through that PCT microscope
Acknowledgments
References
Appendix
1 - Key websites for further reading on perceptual control theory
Appendix
2 - Leading figures in perceptual control theory
A note from the editor
Bruce Abbott
Tom Bourbon
Tim Carey
Gary Cziko
Philip Farrell
Ed Ford
Dag Forssell
Perry and Fred Good
Wayne Hershberger
Fred Nickols
Richard Pfau
Mary Powers
Richard Robertson
Shelley Roy
Phil Runkel
Sara Tai
Further reading
Index
A
B
C
D
E
F
G
H
I
L
M
N
O
P
R
S
T
U
V
W
Z
Back Cover

Citation preview

The Interdisciplinary Handbook of Perceptual Control Theory Living Control Systems IV

Edited by Warren Mansell CeNTrUM (Centre for New Treatments and Understanding in Mental Health) Division of Psychology and Mental Health School of Health Sciences Faculty of Biology Medicine and Health University of Manchester Manchester Academic Health Science Centre Manchester, United Kingdom

Academic Press is an imprint of Elsevier 125 London Wall, London EC2Y 5AS, United Kingdom 525 B Street, Suite 1650, San Diego, CA 92101, United States 50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom Copyright © 2020 Elsevier Inc. All rights reserved. Exception to the above: Chapters 10 & 11: Copyright © 2020 Published by Elsevier Inc. All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions. This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein). Notices Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary. Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility. To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein. Library of Congress Cataloging-in-Publication Data A catalog record for this book is available from the Library of Congress British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library ISBN: 978-0-12-818948-1 For information on all Academic Press publications visit our website at https://www.elsevier.com/books-and-journals

Publisher: Nikki Levy Acquisitions Editor: Joslyn Chaiprasert-Paguio Editorial Project Manager: Tracy I. Tufaga Production Project Manager: Selvaraj Raviraj Cover Designer: Victoria Pearson Esser Artist of the Line Art Drawing in the Biography Section: David Bygott

Typeset by TNQ Technologies

Contributors Heather C. Bell, Behavior and Evolution, University of California at San Diego, La Jolla, CA, United States David M. Goldstein, Moorestown, NJ, United States Erling O. Jorgensen, Riverbend Community Mental Health, Concord, NH, United States Richard Kennaway, University of East Anglia, School of Computing Sciences, Norwich, United Kingdom Warren Mansell, CeNTrUM (Centre for New Treatments and Understanding in Mental Health), Division of Psychology and Mental Health, School of Health Sciences, Faculty of Biology Medicine and Health, University of Manchester, Manchester Academic Health Science Centre, Manchester, United Kingdom Richard S. Marken, Antioch University, Los Angeles, CA, United States Kent McClelland, Department of Sociology, Grinnell College, Grinnell, IA, United States Roger K. Moore, Department of Computer Science, University of Sheffield Regent Court, Sheffield, United Kingdom Bruce Nevin, The Endangered Language Fund New Haven, CT, United States Sergio M. Pellis, Department of Neuroscience, University of Lethbridge, Lethbridge, AB, Canada Frans X. Plooij, International Research-institute on Infant Studies (IRIS), Arnhem, the Netherlands William T. Powers, Lafayette, Colorado, United States M. Martin Taylor, Martin Taylor Consulting, Toronto, ON, Canada Jeffrey B. Vancouver, Ohio University, Athens, OH, United States Henry Yin, Department of Psychology and Neuroscience, Duke University, Durham, NC, United States Rupert Young, Perceptual Robots, Windsor, United Kingdom

xv

Author biographies

William T. Powers Independent Scientist and Inventor, 1926e2013 Bill Powers was a scientist, engineer and a polymath. He began his work on what became known as perceptual control theory (PCT) in the 1950s, and he continued to develop, illustrate and apply the theory until his death. Bill trained as a physicist and engineer, becoming well versed in the field of control engineering. As a young boy of seventeen fresh from high-school, Bill joined the Navy to train and serve as an electronics technician, where he learned about control devices. Bill has been responsible for inventing a number of control devices, including a curvetracer for plotting isodose contours in the beam of radiation from a Cobalt-60 therapy machine, the automatic all-sky photometer for use on the moon (for Apollo 18), and he won the Marshall Field Award for his microcomputer system for receiving, formatting, and typesetting satellite-broadcast stock tables in real time. Bill applied his insights from control engineering to the behavior of living organisms and started to share his theory from the mid 1950s onwards. He published his first paper in 1960 with colleagues Robert Kenley Clark and Robert MacFarland. His undergraduate mentor Donald T. Campbell paved the way for a degree in psychology, yet ultimately, Bill chose instead to continue his work on designing electronic systems while pursuing PCT ‘on the side’. In 1973, Bill published Behavior: The Control of Perception (B:CP) e a full exposition of PCT that is now the key reference on the theory. On publication, B:CP was welcomed with positive reviews, for example by such notable figures as Carl Rogers, the creator of person-centred counselling, and Thomas Kuhn the philosopher of science. Since 1985, Bill facilitated the Control Systems Group e an international group of individuals who utilize, study and discuss PCT within a huge range of disciplines. This organisation continues to grow and provide the catalyst for developments in PCT. In 1998, Bill published Making Sense of Behavior: The Meaning of Control e an accessible ‘arm-chair’ introduction to PCT. There are now a wide range xvii

xviii Author biographies

of books explaining PCT, including Living Control Systems I and II e edited volumes of Bill’s essays and papers on PCT. Bill’s last book - Living Control Systems III: The Fact of Control e is a tour de force in psychology, philosophy, mathematics and computing. Accompanied by a CD of thirteen sophisticated computer demonstrations of PCT, it provides the best exposition yet of what PCT can offer. Warren Mansell University of Manchester, United Kingdom I discovered PCT from Gary Cziko’s marvelous book, Without Miracles, during the late 90s. Working as a researcher in cognitive behavioural therapy, I was frustrated by the disjunct between cognitive theories and behavioral theories. Surely there must be a way to link higher level thoughts with moment-to-moment action? Nothing in what I had read up to this point actually provided a mechanism for how they were linked. When I came across Perceptual Control Theory, I saw that not only were these higher and lower level processes linked, but Bill Powers had worked tirelessly to try to understand the successive layers of controlled perceptions that were involved. I did not expect what followed from my literature searchesdpapers building on models from this theory and using it in diverse fields of the social and life sciences, and beyond. The full realization that all behavior is a means to control perception then sunk in, and illuminated for me a whole range of discrepant findings in psychology. I have worked mainly on using PCT to understand the processes of psychological change in people seeking help for mental health problems, and on how to enhance this process in therapy, using Method of Levels. More recently I have “returned to the laboratory” and begun to develop computerized tasks that test PCT directly, both within the mental health field and wider, such as research on motor control. Rick Marken Antioch University, Los Angeles, United States I discovered PCT in 1974, right after getting my Ph.D. and waiting to start my first teaching job at Augsburg College in Minneapolis. I was browsing through the behavioral science section of the main library at UCSB when a book with a strange title caught my eye. It was Behavior: The Control of Perception. Four years later the same book appeared in the library at Augsburg as did a paper by its author, William T. Powers, in the journal, Psychological Review. So my

Author biographies

xix

career as a PCT researcher began in 1979 shortly after reading the Psychological Review article, and testing its claims using interactive programs that I had written for the newly available Apple II personal computer, which I acquired for the Psychology Department’s research lab. Since my main focus in psychology had been research methodology, my PCT-based research was mainly aimed at demonstrating the revolutionary implications of PCT for how research should be done in psychology. And, indeed, this is still the main focus of my work in PCT, as demonstrated by my contribution to the present collection of papers. I’ve developed several on-line computer programs that demonstrate various aspects of PCT; in particular, 1) what control is, 2) how control can’t be explained by the causal model that is the basis of all current research on psychology, 3) how understanding behavior is a matter of understanding what perceptions an organism is controlling and 4) how the test for the controlled variable (TCV) can be used to reveal what perceptions a person is controlling. I have also done research to show that the TCV can be used to determine what perceptions are being controlled in "real world" behavior, such as when a person (or dog) runs to catch a moving object, such as a baseball (or Frisbee). Henry Yin Duke University, United States I study the neural basis of behavior in mice. My work has focused on a set of brain nuclei called the basal ganglia, which are particularly important for volitional behavior. Our work is beginning to reveal the importance of control theory in the analysis of neural function. I first became aware of perceptual control theory in 2001, when I was a graduate student at UCLA. Reading Wayne Hershberger’s intriguing paper on approach behavior in chicks, I noticed many citations of William T. Powers. At about the same time, I was also reading O. H. Mowrer’s Learning Theory and the Symbolic Processes (1957), which discussed Powers’ work in progress, a few years before the publication of the first paper by Powers, Clark, and McFarland in 1960. Although I quickly read Powers, I was not able to understand him fully at the time. My second attempt to understand Powers was in 2008, just after starting my own laboratory. This time progress was rapid, and I also discovered, to my surprise, that their author was still alive. In 2009 I contacted “Bill” (as he was known to all who knew him), and started a regular correspondence which continued until shortly before his death in 2013. Even though his health and memory were failing, his analytical powers remained unsurpassed, as I was fortunate enough to witness in person when I met him in the summer of 2011. He always replied to my emails promptly and called me often when he had something important to say. I was incredibly fortunate to have known Bill when he was still alive. My appreciation

xx Author biographies

of his greatness has steadily increased since his death. I believe that perceptual control theory will provide the first solid foundation for a scientific psychology, much as Mendel’s work provided the foundation for genetics. Richard Kennaway University of East Anglia, United Kingdom Unusually for those drawn to this area, my professional background is in mathematics and computing, rather than the social and life sciences. After studying at the universities of Edinburgh and Oxford, I have been engaged in research in those areas since then, currently in collaboration with biologists studying plant morphogenesis. My contact with the ideas of Perceptual Control Theory began in the mid-90s, when by chance I encountered William Powers’ book, Behavior: The Control of Perception. Soon after, I sought out and joined the online community of CSGNET. Where I can, I have endeavored to contribute to the mathematical and computational analysis of some of the concepts of PCT, with respect to both control systems that are designed by engineers and those found in living organisms. Control systems pose challenges to current methods of discovering causality by statistical methods, which are explored in my paper in this volume. Sergio Pellis University of Lethbridge, Canada I first encountered PCT in 1977 when I was searching for methods by which to analyze social interactions. In this quest, I came across a book chapter by Ilan Golani (1976: Homeostatic motor processes in mammalian interactions: a choreography of display. In P. P. G. Bateson & P. H. Klopfer (Eds.), Perspectives in ethology (Vol. 2, pp. 69e134). New York, NY: Plenum Press), an Israeli ethologist who described the use of the Eshkol-wachman Movement Notation EWMN), a choreographic system originally developed for dance notation, for non-human animal social interactions. The EWMN tracks the co-movements of animals in time and space and so is able to describe the animals’ movements in a dynamic context. In this book chapter, Golani explicitly placed this methodology in a cybernetic context, citing Bill Powers’ book (Behavior: The Control of Perception). I read Bill’s book and learned how to use EWMN, which together gave me a theory and a method for the study of behaviour. Over the years, I have used the EWMN and the PCT framework to study play, aggression, sex, predation, righting and food handling in a variety of animals, and the reaching and grasping of people with Parkinson’s disease.

Author biographies

xxi

A key aim of my research has been to understand how complex behaviour is produced and transformed during development and evolution. Heather Bell University of California at San Diego, United States I was introduced to PCT in 2008 by my supervisor, Sergio Pellis, toward the end of my M.Sc. in Behavioral Neuroscience at the University of Lethbridge. I was intrigued by PCT’s ability to explain the behavior of organisms in a way that is both physiologically plausible, and that does not force the observer to parse behavior into discrete and arbitrarily-defined packets. After reading William Powers’ Behavior: The Control of Perception (1973), I decided to test PCT for my PhD project, using different animals and behaviors. Since then, I have used PCT to describe food protection in rats and crickets, and to explain some facets of male-male combat in cockroaches. I also developed an agent-based model of food protective behavior using the cybernetic rules extracted from the rats and crickets. The model with its simple cybernetic architecture effectively recreated the behavioral elements seen in the living organisms, despite the fact that none of those “behaviors” were indicated in the code for the program. Currently, I am working at UCSD, where I am using PCT to understand communication in honeybees. Erling Jorgensen Riverbend Community Mental Health, Concord, United States I first encountered what came to be called Perceptual Control Theory in the stacks of the Oberlin College library, around 1979. I was immediately struck by the clarity and incisiveness of this author, William T. Powers. After some time following other pursuits, I chanced upon the Control Systems Group, and was able in the early 1990’s to attend several of their conferences. I was delighted to discover a vibrant and maturing science of PCT, offering a rigorous mechanism behind goal-seeking behavior applicable across numerous degrees of scale. Over the years, I have become a regular participant on the CSGnet listserv, as well as contributing website and annual conference papers. A particular insight gleaned from PCT has been the importance of temporality, as a marker for different hierarchical levels of meaning, and this informed my doctoral dissertation. I have also had an abiding interest in neurophysiology and how the brain may be implementing the functional connections of control loops. This forms the basis of my contributions to the present volume.

xxii Author biographies

Frans Plooij International Research-institute on Infant Studies (IRIS), The Netherlands Having completed our studies in Educational Psychology, Physical Anthropology and Behavioral Biology, and just married, my wife Hetty van de Rijt and I left for the Gombe National Park, Tanzania, East-Africa to study chimpanzees with Jane Goodall in 1971. There was no place on Earth where one could observe free-living newborn chimpanzee babies at such close range. We did not have any theory or hypothesis at hand for testing, but we were trained in systematic, direct observation of animal behavior in the field, in the tradition of Nobel Laureate Niko Tinbergen. So that is what we did for nearly two years. When we returned to Europe in 1973 to work in Robert Hinde’s Medical Research Council-unit on the Development and Integration of Behavior, University Sub-department of Animal Behavior in Madingley, Cambridge, England, we had to analyze reams of data. Out of this analysis emerged the notion of regression periodsddifficult periods where the baby clings more closely to the mother. The results of the data analysis also supported the idea that in the course of early ontogeny a hierarchical organization emerges in the central nervous system that underlies the behavioral development of the freeliving chimpanzee babies and infants. It was only then, in 1977, that our friend and colleague Lex Cools, a neurobiologist, suggested that we compare our findings about the capabilities of infants at the different stages of development to the levels of perception spelled out by Hierarchical Perceptual Control Theory (PCT) developed by William T. Powers. PCT turned out to explain our findings very well. Once we had earned our Ph.D. degrees in Cambridge, England and Groningen, the Netherlands, we moved on to study human babies in their home environments and showed that they, too, go through difficult, age-linked regression periods. We wrote a parenting book about it that became an immediate bestseller in 1992. It has been translated in many languages all over the world and is still going strong after all those years. The title in English is “The Wonder Weeks.” Our original research in The Netherlands has been replicated and confirmed by other researchers in Spain, Britain and Sweden. For information about the research upon which The Wonder Weeks is based, and about editions in various languages, see www.thewonderweeks. com.

Author biographies

xxiii

Kent McClelland Grinnell College, United States My discovery of PCT happened in the summer of 1984 when, as a young faculty member at Grinnell College, I attended a foundationsponsored workshop on “engineering, technology, and the liberal arts” and found myself bored with presentations by the engineers who led the workshop. Skipping sessions, I holed up instead in a university library reading a book I’d seen references to, which happened ironically to be written by an engineer: Behavior: The Control of Perception. The book intrigued me, and although I couldn’t decide quite what to make of it, I kept thinking about it and resolved in 1989 to spend my sabbatical leave reading everything available that was written by William T. Powers or cited his work. My graduate training had been in statistical sociology and methods of survey research, but I had grown disillusioned with that approach and was eager to find some new way of understanding social life. During my sabbatical semester, I wrote the first draft of a long paper that reviewed PCT literature, introduced the theory to sociologists, and discussed social power from a PCT perspective. In the paper, which was eventually published as “Perceptual Control and Social Power,” I invented the label Perceptual Control Theory to distinguish the Powers approach from other control theories, and this has become the name by which the theory is known. At meetings of the Control Systems Group in the 1990s, Bill Powers showed me how to do simulation modeling of control systems, and my work since then has explored theoretical implications of control-system models of conflict and collective control. In my chapter for this volume, I offer my current understanding of how collective control can be seen as the foundation of social order, and conflict, the engine of social change. Martin Taylor Martin Taylor Consulting, Toronto, Canada After a 1956 B.S.E. in Engineering Physics at the University of Toronto, I considered graduate work in control, but instead got a 1958 Masters in Operations Research and a 1960 Ph.D. in Experimental Psychology, both at the Johns Hopkins University, followed by a career as a “Defence Scientist” in Toronto. In 1972, with two colleagues, I published a three-level control model to explain the difference between passive

xxiv Author biographies

tactile touch perception and active haptic object perception. Shortly thereafter, I read Power’s paper on perceptual control theory in Science, and tore it out of the journal for retention, but it made no lasting impression at that time. Starting in 1984, in trying to construct a multimodal (voice, keyboard, pointing) interface to a relational database, I began to develop what became the Layered Protocol Theory of Human Computer Interaction, which soon turned into a general theory of human communication. Around 1990, I was pointed to CSG-L, the mailing list for PCT. It soon became clear that LPT, although independently developed, was actually an application of PCT to a situation that involved two interacting people. My Chapter in this book describes a special component of Layered Protocol Theory. At the 1993 annual PCT meeting, Powers claimed that I was the only person he knew who claimed that his own theory was actually based in PCT. Since then, most of my PCTrelated interests have involved discovery of the properties of multiple perceptual control loops interacting within and among individual organisms or social groups. Bruce Nevin Edgartown, MA, United States I began learning PCT in 1991. Twenty years after commencing fieldwork on a Native American language, I was finishing course work and working on my dissertation. Beginning with my thesis in 1970, I had been exploring the psychological and computational implications of structured repetition as the means of constructing information in discourse, which led to considerations of the form and psychological basis of a knowledge base, and questions about subjective meanings with objective linguistic information. It was immediately evident to me that PCT was the right framework for formulating and investigating these questions. Discussions on CSG-net in the ’90s amounted to an online PCT seminar taught by Bill Powers. I studied Behavior: The Control of Perception and read other papers and books as they became available to me. In time, I described how PCT predicts what speakers would do if we could disturb their auditory perception of their own speech in real time, and when researchers who were uninformed about PCT did exactly that - I designed a PCT model explaining their result. My abiding interest is the constraints on word dependencies that constitute the objective information in language. The correlation of form with information is most clearly seen in the specialized language of a science subfield.

Author biographies xxv

Jeffrey Vancouver University of Ohio, United States Since my graduate school days in the late 80s, I have had an interest in negative feedback loop processes and control theory conceptualizations specifically. My initial exposure was via snippets of control theory concepts in various contexts and under various labels (cybernetics; self-regulation). I especially appreciated the comprehensiveness and rigor of Perceptual Control Theory and Bill Powers’ work after exposure to CSGnet in the early 90’s. Since graduate school, I have been trying to examine negative feedback loops processes in a way that my discipline e Industrial-Organizational (I-O) Psychology e would appreciate while still being true to the dynamics of the theory. In particular, this has involved understanding where well-accepted psychological constructs (e.g., goals, self-efficacy beliefs) relate to perceptual control theory mechanisms. In the process, my work has demonstrated the value of dynamic and computational approaches to the science of I-O Psychology. David Goldstein Moorestown, NJ, United States I first heard about Bill Powers’ ideas in a graduate course on Theories of Perception at the University of Connecticut. Soon after, I was teaching in a Developmental Psychology course at Stephen F. Austin State University (SFA). My graduate students had learned about Powers’ ideas from Dr. Tom Bourbon and they regularly enquired about how his work compared to Piaget’s work on child development. So, I invited Bill Powers to give a talk at SFA, which he did. From that time on, I was “hooked”on what we now call Perceptual Control Theory (PCT). I organized a mini-conference (The Edgemoor Conference in 2007) for my colleagues who worked with me at the residential center where I was Clinical Director. The following year, I organized a CSG conference in Cherry Hill, NJ. My interest in PCT is mostly clinical (self-image, psychotherapy), but also academic (tracking, synesthesia, biofeedback). I have always been interested in a conceptual frame for psychology and I believe that PCT is such a theory. I have been privileged to have Bill Powers as a mentor and friend.

xxvi Author biographies

Rupert Young Perceptual Robots, Windsor, United Kingdom I am an independent researcher and technologist. I received my degree in Computing with Artificial Intelligence from the School of Cognitive and Computing Sciences at the University of Sussex, UK and his PhD in Robotics Vision from the Centre for Vision, Sound and Signal Processing at the University of Surrey, UK. I am passionate about understanding the world around us and in particular the nature of humans and other living systems. With regards to specific research I am interested in building real-world robotic models of perception and behavior as a means to investigate and understand purposive living systems, and in order to produce useful artificially intelligent systems. After a number of years working as a Land Surveyor and travelling around the world I returned to university as a mature student to study Artificial Intelligence, and was particularly interested in Cognitive Psychology and Philosophy of The Mind. I was so absorbed that I went on to complete the PhD in Robotic Vision and Perception, where I came across Perceptual Control Theory. I had not been convinced that conventional approaches to Artificial Intelligence gave much insight into real intelligent systems and quickly realized that PCT had the potential to provide that insight and unify the different strands of behavioral sciences. As there were no opportunities to continue academically with PCT, I moved into the commercial IT arena and have over fifteen years experience leading the design, development and implementation of the software and architecture for large, enterprise projects. Building on this experience, and returning to PCT, I have developed a methodological framework and software libraries for executing and visualizing perceptual control systems on real-time robots. Roger K. Moore University of Sheffield, United Kingdom I have over 40 years’ experience in speech technology R&D and, although an engineer by training, much of my research has been based on insights from human speech perception and production. One of my early publications (Russell, M., Moore, R., & Tomlinson, M. (1983, April). Some techniques for incorporating local timescale variability information into a dynamic time-warping algorithm for automatic speech recognition.

Author biographies xxvii

In ICASSP’83. IEEE International Conference on Acoustics, Speech, and Signal Processing (Vol. 8, pp. 1037e1040). IEEE.) revealed how local timescale variability in speech could be modelled as a continuum with varying degrees of phonetic contrast. I subsequently realised that this was a special case of a more general principle published in 1990 by Bjorn Lindblom known as the “H&H” (hyper & hypo) theory of phonetic variation in speech, in which he argued that the apparent variability of speech was actually the inevitable consequence of a negative-feedback control mechanism that balances the effectiveness of communication against the effort involved in communicating. This fit perfectly with my own ideas, so I continued to incorporate concepts of closed-loop control into my research until, in 2005, I finally came across Bill Powers’ work on perceptual control theory (PCT). In some sense, PCT completed the picture, and I was able to integrate the core ideas into to a novel architecture for spoken language processing (by mind or machine) known as “PRESENCE” - PREdictive SENsorimotor Control and Emulation. PRESENCE was published in 2007, and since that time I have been investigating the implications, particularly in the context of vocal interaction in-and-between humans, animals and robots.

Preface On Friday 24th May 2013, William T. Powers, the founder of perceptual control theory (PCT), sadly passed away. This book is dedicated to him. I owe him a debt of gratitude. Without his scientific discoveries - my work, my world, and my life, would not make quite as much sense as they do today. Many of my colleagues have had a similar loss. And because Bill continued to be so generous with his time, patient in the face of persistent questions, and crystal clear in his reasoning, the loss was that much greater for all of us. There will no doubt be a story of Bill Powers’ incredible life as an engineer, inventor, scientist, mentor, friend, father, brother, and son, but this book is not that book. This is the book that he wanted written by the many people with whom he had shared his vision. Each of them had gone on to share, test, or use his ideas in very different ways, in diverse academic fields, all with Bill’s support and of course, feedback. I must particularly thank Alice Powers McElhone, Bill’s sister, who through her business, Benchmark Publications, ensured the continued publication of Behavior: The Control of Perception, Bill’s first book, as well as the publication of his subsequent works, Making Sense of Behavior and Living Control Systems I, II, and III. From 2013 to 2018, Alice nearly brought the publication of this final book in the Living Control Systems series to completion. We are extremely grateful to Elsevier for taking her good work to the finish line in time for her to witness the fruit of her labors. Whilst this book is not a biography, I am going to begin this Preface with some context about how Bill’s journey along this road of discovery may have begun. I will then explain the components of his theory, and go through how each chapter in this book utilizes the theory to construct an interdisciplinary account of behavior. On one of Bill Powers’ visits to the Control Systems Group conferences we held here in the UK, he told me that the dropping of the first atomic bomb was the turning point for him. Watching the horrific loss of life unfold, Powers realised that people had made the choice to let loose this catastrophic act of destruction on other people. Using the science of physics, we e the human race - had discovered how to destroy each other, but not how to build bridges with one another and heal conflict in our society. The same is of course true today. Bill told me that he chose, at that point, to turn his intellect and knowledge of physics and engineering to understanding people and, critically of course, what is involved in their purposeful actions, motives, and intentions. And so, in PCT, we see the xxix

xxx Preface

detailed blueprint of an engineer, but wholly, and uniquely, from the perspective of the “machine” itself - the living organism - and ultimately from the perspective of each one of us. Bill stated that control is essential for life, and it requires a system of parts, organized in a ‘loop’ uniting the individual and the environment. This loop keeps things at a state that is “just right” for the organism. Goldilocks would have approved of PCT, adjusting the temperature of her porridge by adding dashes of cold milk, and then hot milk, until it met her standard. The means of control should not be fully specified in advance. In fact, I’m sure that if Goldilocks were around today, rather than firing up the stove, she would pop her tepid porridge in the microwave for 20 seconds to get it to the temperature she desired. So, how technically, does control work? Each chapter will explain further. In brief, however, here is an overview, followed by an explanation of the components. Inside the organism, incoming signals from aspects of the environment (e.g. heat of the porridge) are compared with an internally specified reference value (e.g. 60 degrees celsius) for the controlled variable (porridge temperature). These comparisons drive actions (e.g. pour in hot milk) to act against disturbances (e.g. air convection) to those signals. This is done as continuously as possible to keep the inputs at their desired values. Voila, perfect porridge! This is PCT in a very small nutshell. The next section of this Introduction, and the following chapters of this book, including Bill’s own, describe PCT in ever larger nutshells, with some that describe the whole fruit, the tree, and in some cases, the wider forest and whole ecosystem of perceptual control. Fig. 1 shows the basic unit of PCT in more detail, with each element defined. There are features of this diagram that are important, and need to be explained in order to introduce the chapters in the Handbook. It is important to note the location of the border between the organism and the environment. The input function, output function and comparator are situated inside the organism. Only the feedback function, the disturbance, and the aspects of the environment being controlled, are outside the organism. The input function is responsible for coding the aspect of the environment to be controlled. It transforms energy from the environment (e.g. heat conducted from the porridge to one’s lips during a sip) into a continuous perceptual signal. This signal is proportionate to the current amount of the controlled variable e e.g. the current temperature of porridge. The input function is therefore carried out by the relevant sense organ(s) and further processing may continue within the neurons that send signals upwards into the nervous system. The perceptual signal is subtracted from the reference signal, to create the error signal by the comparator. This signal is transformed by the output function. For example, it may be amplified (by a gain value), and ultimately sent to muscles in the body (although see the section below about the lower level control units). This signal is then transformed by a feedback function in the environment. Part of the feedback function, in the example of Goldilocks, is the cold milk she might use to cool

Preface

To higher systems

xxxi

From higher systems REFERENCE SIGNAL r

COMPARATOR

PERCEPTUAL SIGNAL Represents magnitude of one dimension of environment

Specifies intended or desired magnitude of perceptual signal

Measures mismatch between reference and perceptual signals e=r–p

p

ERROR SIGNAL e

Indicates amount and direction of difference between reference and perceptual signals

INPUT FUNCTION

OUTPUT FUNCTION

Converts state of input quantity into magnitude of perceptual signal p = Ki Qi

Converts magnitude of error signal into state of output quantity Qo = Ko e

CONTROLLING SYSTEM ENVIRONMENT INPUT QUANTITY

FEEDBACK FUNCTION

Physical variable that affects sensory inputs of controller (may be multiple)

Physical properties that convert action or behavior into effect on input quantity

Qi

Qo

Measure of system’s physical output action or observed behavior

Qi = Kf Qo + Kd D

DISTURBANCE Physical variable that affects input quantity (may be multiple)

OUTPUT QUANTITY

D

FIG. 1 A model of a negative feedback control unit as specified by PCT. Definitions, notations and formulae for key terms are provided within the diagram. Diagram redrawn by Dag Forrsell from a drawing by William T. Powers.

down the porridge if it is too hot. In this way, the controlled variable is shifted towards the desired direction. This change is immediately sensed by the input function, and checked, inside the organism, by the comparator. There are some further features that are critical to a correct understanding of PCT. First, although it is necessary to explain the passage of signals round the control loop in a step-by-step way, like I have done, this is not how it works in real life. In real life, the functions are constantly, simultaneously processing signals. The loop does not wait for a specific signal to go all the way round the loop before it processes the next signal. Rather, like the circular motion of water in a whirlpool, the signals are continuously changing, and present in the whole loop all of the time. This property means that a PCT model does not simply execute a response (or fixed action pattern) and then occasionally receive feedback from the environment to learn from the effects of its actions. The dynamic nature of the loop means that action is unfolding continuously to keep

xxxii Preface

an aspect of the environment controlled, and can do so even without learning, and without having to stop to evaluate the effects. A second feature of PCT is the hierarchy. Fig. 1 shows that the living organism actually achieves control through a branching hierarchy of control units. Many of the chapters in this book will explain and elaborate on the hierarchy. But for now, it is important to make it clear that the hierarchy explains both why, and how, a particular variable is controlled. It explains why, because the reference signal for any variable is set by the output signal from a control unit at the level above. For example, when Goldilocks comes home after a day playing out in the snow, she detects this discrepancy from her usual body temperature, and she so she wants her porridge hotter than usual. The hierarchy explains how, because the output signal for a unit specifies the reference value of lower level systems, with only the lowest level of the hierarchy interfacing directly with the environment. For example, to adjust the temperature of porridge, Goldilocks may want to perceive cold milk pouring into her bowl, which she achieves by wanting to perceive a bottle of milk tipping into the bowl, which she achieves by wanting to perceive a bottle of milk in her hand, etc. At each level down, we have a desired perception, rather than an action. The environment itself also needs unpacking. It represents the wider system outside the nervous system e the body, the physical world, and other people. Those aspects called feedback functions are those that allow the organism to control. The disturbances are those features of the environment that interfere with control and are counteracted by the organism acting through the feedback functions. So, for Goldilocks, the feedback function includes her body and the various objects she uses in her kitchen. The disturbances would include convection currents and heat conduction in the room, and objects in the room that might get in the way e the door of the fridge where the milk is located for example. A key question that arises is e what sets the reference values at the highest level? First to note are the intrinsic systems. These are biological control systems. They include the homeostatic systems controlling body temperature, blood glucose levels, pain thresholds, and other essential variables for survival and reproduction. The reference values for intrinsic systems e termed intrinsic references e are set within the organism prior to development; they have become specified through evolution over generations rather than learned during the organism’s lifetime. According to PCT, when intrinsic systems are in error (intrinsic error), they drive a process known as reorganization. Reorganization is responsible for learning and adjusting the various functions involved in perceptual control, and reorganization does so in a random, trial-and-error fashion, until it ‘stumbles’ upon the change in function that reduce the intrinsic error. According to PCT, reorganization is also vital to resolve conflict. Conflict is defined as the control of a variable based on two or more opposing reference values for the same variable. For example, Goldilocks may want her porridge at 70 to warm herself up quicker, but also wants it at 50 to stop her mouth from

Preface

xxxiii

feeling burned. Because living organisms, especially humans, attempt to control so many aspects of themselves and the world, conflict is inevitable. The longterm solution to resolving conflict is to allow the control system whose outputs set the incompatible reference values for the variable to reorganise and therefore reduce the intrinsic error. Powers proposed that the spotlight of our conscious awareness provides an indicator of where reorganization is operating. Therefore, in order to resolve long-standing conflicts, people need to shift and sustain their awareness to the system above the conflict. Method of Levels is the psychological therapy, and counselling and coaching method based on PCT, that helps a person to resolve conflict. Thus, the highest-level systems, and their reference values, have been ultimately been formed through reorganization in order to minimise conflict and, in turn, reduce intrinsic error. A further feature of PCT that helps to explain some of the more sophisticated features of human behavior is the role of memory. Powers proposed that memory is distributed to every control unit, such that it is addressed by the outputs of the level above, and is used as a reference value for control. For example, Goldilocks remembers the satisfaction of a lovely hot bowl of porridge that her mother had made her, and tries to recreate that same temperature. Memories can be recirculated as input within the hierarchy within a short-circut that is known as the ‘imagination mode’. This allows us to exert control over our thoughts, mental imagery, and plans inside our heads. Some of the chapters in this book elaborate on the details of this mode, and explain its role. Lastly, we have the phenomenon of collective control, which emerges when two or more individuals are linked via the same environment and attempt to control some of the same aspects of this environment. For example, Goldilocks and Mother Bear both need the cold milk to warm their porridge. Goldilocks opens the fridge, but it turns out to be on the top shelf and only Mother Bear can reach it for her. Thus, getting the cold milk from the fridge is managed by collective control. There is a wide range of different kinds of relationships that can emerge in these kinds of collective networks, and several chapters elaborate on these. Now the theory is briefly summarised, I will briefly introduce each chapter in the book. Bill Powers wanted this edited volume to show how the world could be different if people were to take the perspective of perceptual control theory. This is the sentiment behind the first chapter, which is his own rallying call for a new scientific perspective on behavior. In this chapter, he speaks to the target audience for this planned book which includes both academics and practitioners who already use PCT in their work, and also those new to PCT who want to find out more. PCT provides a fresh perspective, and in many cases, may have the capacity to transform their discipline. Bill had hoped, with each of his earlier publications, that his discovery would be adopted by mainstream science. In particular, his collection of previous works, Living Control Systems I, II and III, were aimed at paving the way. However, this moment - as is often the case with work that promises a paradigm shift and change of course in one’s thinking - is

xxxiv Preface

now saved for some time in the future. Bill anticipated this with the publication of the Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV, in that presenting PCT from the perspective of such a wide range of fields of study cannot help but speak to this theory’s all-encompassing reach. The first four chapters, including Bill’s, each serve to explain to a new acquaintance to PCT why its unique perspective contrasts with the established approaches in the behavioral sciences. In doing so, it unavoidably exposes the limitations of other scientific theories, and provides a rationale for why PCT is needed to explain behavior. The idea that PCT is the ‘correct’ approach to studying behaviour is sometimes regarded as a sensational claim, and certainly not all the authors of this volume share the same view. But these initial, controversial chapters are nonetheless required reading. Bill Powers sets the stage by explaining how the insights of early control engineers stood at odds with the newly developing science of behaviour at the start of the 20th Century. Rick Marken builds upon this perspective in his chapter by spelling out how the engineer’s knowledge of how a control system is constructed (forward engineering) contrasts with the psychologist’s goal of trying to understand the mechanism involved in a behavior (reverse engineering). Reverse engineering based on PCT is known as the Test for the Controlled Variable (TCV), and Marken explains this original method of experimentation in detail. Henry Yin further builds upon Marken’s case. He uses popular examples within behavioral neuroscience e Sherrington’s analysis of the reflex, and sensory transformation in animal studies e to illustrate the fallacy of the established stimulus-response account that takes the observer-perspective on behaviour. Yin contrasts this with the face validity of a perceptual control model, which takes the behaving organism’s own perspective. In a final chapter in this section, Richard Kennaway uses a mathematical proof to show that a fundamental assumption behind how we analyse cause-and-effect systems does not apply to perceptual control systems. Typically, it is assumed that if two variables correlate then there may be some kind of causal relationship between them. Yet, perceptual control systems have a known a causal relationship e behaviors control perception - in the absence of a correlation between behaviour and perception. Moving beyond the clear cases to be made for PCT when contrasting it with established theories and methods, the next step is to show its capacity to integrate and explain natural observations. The chapters in this section focus on animal behavior within contests, brain anatomy and neurophysiology, and chimp and human development. Sergio Pellis and Heather Bell utilize PCT to explain how dynamic interactions emerge within a contest through attempts by rival animals to attack specific target areas on their opponent, and defend specific areas of their own body. Erling Jorgensen proposes the neural substrates of perceptual control. His first chapter describes the neural circuitry between the layers of the cortex, in order to propose how memories are accessed as neural specifications for incoming signals. His following chapter identifies the evidence for neurones that specify the input functions (the coding of sensory signals from the environment)

Preface xxxv

for each perceptual level. Frans Plooij tells the story of his discovery that the stage-by-stage profile of infant development in chimps and humans fits neatly into the PCT architecture and he illustrates how reorganization is implemented to restore control. Plooij found that every few months after birth, the input functions for a new level of perception are constructed. This is first experienced as uncontrollable, entailing distress and conflict with the caregiver. However, as the infant’s brain reorganizes, control is restored until the next stage begins. The following three chapters were all written in close collaboration because they describe, model, and illustrate the implications of collective control. Kent McClelland, a sociologist, describes how organisations of individuals of increasing size can be understood as interconnecting systems that control perceived aspects of the environment in common with one another. Martin Taylor explains in detail how two or more individuals can control in an arrangement known as a protocol, in which the behaviour of one person serves as the controlled perception of another person. Examples include the greetings, favours, and rituals appropriate to the culture and status of the individuals involved. Bruce Nevin combines these perspectives on collective control to present an explanation of how PCT can be used to provide a working model of how subjective meaning relates to the objective information within language and thought. With the ramifications of PCT across the biopsychosocial domains fully described, the next chapters in the book describe how PCT is transforming methodologies for research, professional practice and technological innovation. Jeffrey Vancouver recounts the journey of Powers’ work through several decades of theory and research within organisational psychology, culminating in the testing of computational PCT models developed against data from behavior in the workplace. David Goldstein and myself describe the history, research and practice of Method of Levels (MOL) as an individual psychotherapy. We start with Powers’ early use of the approach to discover the perceptual levels when describing his theory, through Tim Carey’s development of MOL as a therapy, and forge through to its empirical study in recent years. MOL has certain features e client appointment scheduling, the client’s own choice of what problem to focus on, the lack of a requirement for session plan e that make it a particularly flexible solution to providing therapy in challenging contexts such as schools, inpatient wards and prisons. Rupert Young echoes the earlier chapters of the book in claiming the unique, transformative potential of PCT. He uses a variety of illustrative examples to make the case that PCT is the only viable approach to building a truly autonomous robot, and he describes his own creations. In the next chapter, Roger K Moore provides a vision for even more complex robotic devices, that utilize PCT to infer the intentions of people with whom the device needs to communicate with and, thereby, utilizing language in a truly adaptive and purposeful manner. At the end of this Handbook, as a way to reflect on the long, interdisciplinary journey through the chapters, I take a breath to review what PCT has achieved in

xxxvi Preface

science, its key influences, its current empirical status, and the directions for future research and application. Right now, over 60 years since the first publications of Powers’ ideas, I propose that it has no parallel among other interdisciplinary theories of behavior in terms of its explanatory power, breadth of application, and its impact on research methodology. Yet, these are early days in terms of the future acceptance and use of PCT, spurred on by this first ever collection of reviews across a spectrum of scientific domains. I hope that you witness the bold spirit of Bill Powers in each of these chapters, and that these accounts build and refine your understanding of the theory so that your own work within these professional and academic disciplines can be transformed and flourish. In keeping with the theory, just take your own pace with the Handbook, drop a line to ask the authors any questions that come to you, try new ideas out in your own practice and everyday life, get feedback from other people continuously, and allow any fresh perspectives you have to emerge naturally. And maybe, just maybe, Powers’ legacy will be the antidote to the conflict in our human society that he had hoped for. Warren Mansell CeNTrUM (Centre for New Treatments and Understanding in Mental Health), Division of Psychology and Mental Health, School of Health Sciences, Faculty of Biology Medicine and Health, University of Manchester, Manchester Academic Health Science Centre, Manchester, United Kingdom

Acknowledgment Thank you to Vyv Huddy and Alison Powers for their feedback and advice on this Preface.

Chapter 1

The world according to PCT William T. Powers Lafayette, Colorado, United States

Beware of the man who works hard to learn something, learns it, and finds himself no wiser than before . He is full of murderous resentment of people who are ignorant without having come by their ignorance the hard way. d Kurt Vonnegut, “Cat’s Cradle”.

In the 1930s, electrical engineers discovered how to build systems of a kind that could do the following things: 1. Act on the world outside them to cause changes of temperature, speed, position, force, rate of rotation, altitude, and acidity, to mention at random only a few possible variables. 2. Bring any of those aspects of the world to a desired goal-state, starting from any previous state, by acting directly on it. 3. Keep any of those aspects in the goal-state (or it bring it back to the goalstate) even if unwanted variations occur so that different amounts and directions of action are required each time to achieve the same result, and do this even if the variations are unpredictable, and even if the causes of the variations are invisible and unknown. 4. When the specification for the goal-state of a variable is changed, alter the action on the physical world however necessary in the current environment to bring the sensed magnitude to the new value and keep it there. Devices of this kind have been called, ever since the 1930s, negative feedback control systems. They are goal-seeking, intentional, purposive organizations of matter and energy. However, by the time the nature of such devices was discovered, scientists who were concerned with a strictly scientific study of the behavior and functioning of living systems had reached a consensus that for any physical system, the behavior just described is impossible. And although their descendants in cognitive psychology and kindred fields, unlike their predecessors, may sometimes use words like purpose and goal, they are still uncomfortable with them because there seems to be no principled way to account for such phenomena. Crucial aspects of their theories and methods remain rooted in the old ‘strictly scientific’ view. The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00001-0 Copyright © 2020 Elsevier Inc. All rights reserved.

3

4 SECTION | A Why do we need perceptual control theory?

How was this counter-factual conclusion reached? Observably, cause precedes effect in physics, so it was assumed that, since living things must obey the laws of physics, behavioral actions are likewise preceded by causes (‘stimuli’) and followed by effects (‘responses’) in strict temporal order. It was further reasoned that since the causes of behavioral actions had to precede them, and since the actions in turn cause the observed results, there is no way for a goal d a specific result of the action d to reach back through time as a cause of the action from which it results. There was no way for a goal-state to ‘stimulate’ the behaving system into acting in just the way needed to produce that very goal-state as a ‘response’. It was concluded that the appearance of goal-directed, intentional, or purposive behavior had to be an illusion, and it was the duty of science to put that illusion aside along with all the other myths and illusions of an earlier time, and to search soberly for the correct explanation for these misleading appearances of behavior. Many behavioral scientists nevertheless could not reconcile this conclusion with their experience. They rejected ‘reflexology’ and ‘stimulus-response psychology’ and studied behavior in a naturalistic way, accepting even unexplainable appearances at face value. They tried to adhere to the basic tenets of scientific research as much as they could, using statistics and mathematical modeling with some competence, but there remained an apparently unbridgeable gap between the empirical, functional understanding of the ‘soft sciences’ (as they came to be called) and the ‘hard sciences’ based on fundamental principles of biophysics and biochemistry. This discontinuity between the levels of analysis of the two camps persists today. The soft sciences couldn’t point to physical stimuli causing specific behaviors, but in the effort to be more scientific they could at least claim to study cause and effect by looking at patterns in the relations between environmental influences and general kinds of behavior. They studied phenomena that the stricter behavioral scientists could not, but their conception of linear causality from environment to behavior was the same, and their methodology was highly similar. They could study the relationship of childhood poverty to a subsequent life of crime and show that a link existed (for a significant portion of the population studied), even if the resultant ‘general linear model’ could give no insight into the form of the mechanisms involved. In this way, the same basic model of linear causation that was adopted by behaviorists has persisted in fields like cognitive psychology, sociology, and personality theory. Only prior causes can generate future effects, and goals or purposes represent future effects, therefore they cannot affect behavior. Either goals are illusory, or they exist in a subjective ‘mental’ world with little or no explainable connection to the physical world. In the course of searching for the correct explanation to replace the illusion of purpose, behavioral scientists have proposed many theories, but first they had to change the way behavior is described. Since the appearance of goalseeking is an illusion, a person who wants to be considered a scientist can’t

The world according to PCT Chapter | 1

5

speak of what an animal or a person is ‘intending’, what it ‘wants’ or is ‘trying to accomplish’. We observe that the environment stimulates the organism’s sensors, generating neural signals that are routed by various known paths to the muscles and glands where they produce movements and secretions like those to which we give the names of behaviors. When the muscles contract, we see not simply actions, but conditioned responses to stimuli that produce effects radiating out into the environment. This, it was agreed by large numbers of scientists, is the only scientific way to describe behavior. Observable effects always come after observable causes. Goals are easily understood: they are simply outcomes of behavior determined by prior stimuli. If a person doesn’t happen to know of the existence of negative feedback control systems, this line of reasoning and its subsequent influence on interpretations may well seem inescapable. If one wishes to investigate living systems, and also to do this scientifically, environmental determinism seems inevitable. There is apparently no other path to take. But in ordinary experience we do seem to have goals, intentions, wishes, desires, objectives, preferences, and targets for a great many things that we haven’t achieved yet. Such goals and so on appear to exist; it is a very convincing illusion. When we read a description of what a negative feedback control system can do, it not only seems quite reasonable and familiar, but it sounds just like the things we can do ourselves. We can easily imagine needing milk, driving a car to the grocery store, buying milk, bringing it back home, and stowing it in the refrigerator. That looks very much like carrying out a series of intentions, like having a goal or goals and then achieving it or them. It’s hard to imagine that this is an illusion. It seems completely real.1 It’s even harder for a scientist to portray what happens without making it sound goal-directed and intentional. Yet that is exactly what large numbers of scientists have tried to do. Somehow, they had to describe what was observed so that the responses were always being caused by stimuli from outside, and show that while the outcome of an action might be some result beneficial to the organism, the action was produced without reference to the benefit, certainly not in order to get a benefit that did not yet exist. The very term ‘response’ carries the presupposition that it results from some prior stimulation. By careful use of language, scientists could bypass the illusion of purpose. This was not easy to do but they did it d or thought they had done it. And all in vain. It was a futile exercise, a waste of effort, simply because negative feedback control systems are real physical systems subject to all the laws of physics and chemistry, yet they can do all those impossible things I mentioned, just as all organisms do them. All the arguments saying that such things can’t be done are swept aside when we see them being done d more easily, of course, when we understand how they are done. We can stop trying

1. Imagining also appears really to occur.

6 SECTION | A Why do we need perceptual control theory?

to imagine how to describe behavior in any way but the right way, which d cross your fingers and knock on wood d we now know with as much certainty as today’s science permits. In tracking down the place where the sciences of life went astray in the opening decades of the 20th Century, we are led to discoveries about the nervous system made early in the 1800s. It was discovered that there were two kinds of nerves, one carrying signals from sensory organs inward into the brain and the other carrying signals from the brain outward, from higher centers to lower centers and finally to the muscles (and glands). As the 20th Century started, the great neurologist Sherrington saw the outgoing signals as general commands formulated in the cerebral cortex and then elaborated, level by level, into the detailed commands necessary to cause complex and coordinated muscle actions. Also level by level in reverse order, the incoming signals were seen as stimulating the brain to formulate those commands. We can be almost certain now that Sherrington was mistaken about the functions of those signals, particularly about the idea that the signals going from higher centers toward the muscles are commands to produce actions. The signals are not commands. Instead, they are specifications not for actions but for sensory perceptions of the results of actions d ‘reference signals’ in the language of control engineers. The message they carry is, “Make the perception you are sending to me look like this.”2 The lower systems then, level by level, immediately alter their actions as required, the lowest system changing the environment (which includes muscles) so that the perceptual signals rising from sensory neurons toward the spinal cord are adjusted to match the lowest level of reference signals. At each more general or abstract level, perceptual signals are made to match the reference signals descending from the next level above.3 So it is not a future event that acts at any level as a goal. Real neural signals that exist here and now define a hierarchy of outcomes to be achieved, just as a real blueprint, here and now, specifies via a general contractor what the subcontractors and their employees are to build. Each system continuously compares perceptual signals (derived from a combination of perceptual signals from lower systems) with the reference signal it is receiving from above. They continue to act until the difference between 2. Sherrington, a hundred years ago, realized that “the brain thinks in terms of movements, not muscles.” He found that brain stimulation in a particular location could cause movement of a limb to a specific place from any starting position. Different efforts and even different muscles were required. and only the final position repeated. But he couldn’t explain how that could happen. Control theory can. 3. At the top level, the reference signals are from some genetic or other fixed source. The values of fixed reference signals can also result from learning, a topic which is beyond the scope of this introductory survey. In broadest terms, learning progresses from ability to perceive a variable, through ability to influence its value, to establishing a preferred state for it and controlling the perception.

The world according to PCT Chapter | 1

7

reference and perception is close to zero. Most electronics engineers would recognize that organization as what they call a ‘cascade’ of negative feedback control systems. We call it a hierarchy of control. To see the nervous system as a collection of control systems is thus no more controversial than it is to see the heart as a pump. The heart is a pump because it pumps. The nervous system is a (multidimensional, multilevel) control system because it controls. Of course a pump is easier to understand for someone who knows how pumps work than for someone who doesn’t. But control systems have also been easy to understand since engineers figured out how they work. Unfortunately, they didn’t figure it out until well after life scientists had committed themselves to some serious theoretical mistakes. This book is about those mistakes and how we can fix them. Before any ‘fixing’ can be done, however, it has to be permitted to happen. Here is the barrier that prevented the discoveries by those control-system engineers of the 1930s from immediately revolutionizing all the life sciences. Even those who quickly approved of what the engineers had done failed to understand the disruption this would cause to established careers and the perimeters of their work. All of us who hopped onto the cybernetics bandwagon and later took off in still newer directions naively expected these new ideas to be appreciated and adopted by traditionally open-minded scientists who preferred knowing the truth to being right. The massive opposition from some quarters and the passive resistance from others came as a disappointing surprise, but perhaps it shouldn’t have. Science has a social as well as an intellectual aspect. Scientists are not stupid. They can look at an idea and quickly work out where it fits in with existing knowledge and where it doesn’t. And scientists are all too human: when they see that the new idea means their life’s work could end up mostly in the trash-can, their second reaction is simply to think “That idea is obviously wrong.” That relieves the sinking feeling in the pit of the stomach that is the first reaction. Being wrong about something is unpleasant enough; being wrong about something one has worked hard to learn and has believed, taught, written about, and researched, is close to intolerable. All scientists of any talent have had that experience. The best of them have recognized that their own principles require them to put those personal reactions aside or at least save them for later. They know that any such upheaval is going to be important, and ignoring it is not an option. But those who recognize and embrace a revolution in science are the exception. Most scientists practice ‘normal science’ within a securely established – and well-defended – paradigm. That is what we are up against here, and have been struggling with since before most of you readers were born. We have spent that time learning more about this new idea and getting better at explaining it, but no better at persuading others to change their minds in a serious way when their career commitments are threatened by it. What we had thought would be a nutritious

8 SECTION | A Why do we need perceptual control theory?

and deliciously buttered carrot has proven to function like a stick. The clearer we have made the idea, the more defenses it has aroused. We are now facing reality. This is going to be a revolution whether we like it or not. There are going to be arguments, screaming and yelling or cool and polite. It’s time to sink or swim. As the father, or by now the grandfather, of Perceptual Control Theory, I have invited the colleagues you meet in this book to join me in making the case for PCT. And as there are many more equally qualified Perceptual Control Theory experts, I hereby ask members of the Control Systems Group (now the International Association of Perceptual Control Theory), and the distinguished authors of The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV, to seek others all around the world, who are wellacquainted with some conventional branch of the life sciences as well as with PCT d who understand why the older theories were persuasive and even useful, and also what has to be changed about them if the realities of negative feedback control are to be introduced into them. Part of the plan for this book has been to encourage cooperative authorship and peer review of the work as it develops. Many of us have found that group editing of a common document is an eye-opening experience that changes competition into cooperation, and self-interest into a desire that the document itself be better than one contributor could make it. If this part of the introductory chapter is still here to be read, you will know that this experiment worked. The reference sections list a substantial collection of books and articles explaining PCT. In this book you will see it explained in specific contexts. I may add brief comments here and there; aside from that, this is good-bye for now. Bill. March 2013. Lafayette, Colorado.

Chapter 2

Understanding purposeful systems: the application of control theory in engineering and psychology Richard S. Marken Antioch University, Los Angeles, CA, United States

Both the engineer and the psychologist have an interest in knowing how purposeful systems work. Purposeful systems are devices that act to bring variable aspects of the environment to goal states and keep them there, protected from the effects of unpredictably varying disturbances. This process is called “control”, and so purposeful systems are also known as control systems. The theory that explains how these systems work is known, appropriately enough, as control theory. Engineers use control theory to help them build control systems from scratch, a process called forward engineering. Psychologists use control theory to help them understand how existing control systems have been “built”, a process called reverse engineering. The same control theory is used in both cases but how it is (or should be) used is quite different. This chapter will explain how control theory has been used in engineering and psychology and describe the proper way to use control theory in psychology to reverse engineer living control systems.

Understanding control: control theory Engineers started building control systems well before they had a detailed understanding of how those systems worked. Indeed, engineered control systems date back at least to 300 BCE when Ktesibios of Alexandria designed a system that controlled the rate of flow of water into a water clock (Lewis, 1992). It did this by maintaining a constant water level in the tank from which the water flowed into the clock. An understanding of how these artificial control systems worked began to emerge only in the latter half of the 19th

The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00002-2 Copyright © 2020 Elsevier Inc. All rights reserved.

9

10 SECTION | A Why do we need perceptual control theory?

century in the form of control theory. The basic idea of control theory is that control is the result of a closed-loop negative feedback process, such as that diagrammed in Fig. 2.1. The control system diagram in Fig. 2.1 represents a closed-loop of cause and effect: the error variable, e(t), is the cause of the output variable, y(t), which, along with a reference variable r(t), loops back to cause the error variable itself. It is a negative feedback loop because the output causes a decrease in (has a negative effect on) the value of the error that is the cause of that output; so increases in error cause outputs that decrease the error that is causing those same outputs. When appropriate functions are selected for the system and plant e YH and YP, respectively e this negative feedback loop will vary its actions, o(t), appropriately so as to keep y(t) under control in the sense that the value of y(t) will be kept close to the value of the reference variable, r(t), protected from the effects of environmental disturbances, d(t). Control theory provides the mathematical basis for determining the functions YH and YP that will result in y(t) being kept under control. A familiar example of the negative feedback control process diagrammed in Fig. 2.1 is the home thermostat that controls room temperature. The room temperature that is being controlled corresponds to the variable y(t) in Fig. 2.1. The setting of the desired room temperature corresponds to the variable r(t). The thermostat (the System in the diagram) continuously computes the difference between the desired and actual room temperature (r(t)-y(t)), which corresponds to the error variable, e(t). The thermostat converts this error variable into a signal, o(t), that turns a heater and/or air conditioner– the Plant– on or off depending on whether the actual room temperature, y(t), is above or below the goal or reference temperature, r(t). The actions of the Plant join with the effects of environmental disturbances, d(t), to produce the actual room temperature. The result is that room temperature, y(t), stays nearly exactly equal to the reference temperature, r(t); if the reference temperature is set to 68
F the room temperature stays at 68 F, even in the face of large disturbances such as changes in the temperature outdoors or the number of people in the room.

FIG. 2.1 Diagram of a negative feedback control system.21

Understanding purposeful systems Chapter | 2

11

Although the variable y(t) is called “Output” in Fig. 2.1, it is more appropriately called the controlled variable since it is the variable that is being kept matching a reference value while being protected from disturbances. This is clear from the example of the thermostat, where y(t) corresponds to the temperature of the room, the variable that the thermostat is controlling. In Fig. 2.1, y(t), is called an “Output” because the Figure is labeled from a forward engineering point of view. Engineers build (forward engineer) things to produce some product or output for users; tractors are built to produce lifted dirt; radios are built to produce music, and thermostats are built to produce a constant room temperature. So engineers think of y(t) as “Output” because they build control systems to produce an output – a controlled result – for the benefit of a user of the system.

Doing reverse engineering from a forward engineering perspective Unfortunately, control theory was introduced into psychology from the forward engineering perspective shown in Fig. 2.1, a perspective that happened to be compatible with the prevailing cause-effect view in psychology of how the behavior of living organisms “worked”; behavior was seen as an output that was caused by stimulus input. So the “System” in Fig. 2.1 was seen as being equivalent to a living organism that produces behavioral output, y(t), in response to stimulus input, e(t). Psychologists used this forward engineering view of control theory to determine the nature of the “transfer function” (the function YH in Fig. I) that represents the causal path through the organism e usually a human e from stimulus input to behavioral output (Craik, 1947, 1948; Howell, 1971; McRuer & Krendel, 1959; Sheridan & Ferrell, 1974). The aim was to determine how inputs cause the outputs that people use to control various machines, such as aircraft (Arents, Groeneweg, Borst, van Paassen, & Mulder, 2011; Wickens, Hollands, Parasuraman, & Banbury, 2012). This information was used to build input displays that would help people control these machines more effectively. One could say that forward engineering techniques were being used to reverse engineer living control systems so that the performance of these systems could be improved via forward engineering. The problem with reverse engineering control systems from a forward engineering perspective is that doing so ignores the first question that should be asked when trying to understand the behavior of these systems: What is the system controlling? This question never arises when forward engineering is done by engineers; engineers know exactly what variables their control systems are controlling because they have built those systems to control those variables; the engineer knows that the thermostat controls room temperature

2. Based on Fig. 14.2, p. 160 of Jagacinski & Flach (2002).

12 SECTION | A Why do we need perceptual control theory?

and that the cruise control system controls car speed because the engineer builds those systems to control those variables. But the psychologist doing reverse engineering is in quite a different position. The psychologist doesn’t know what variables are being controlled by the control systems they study because, unlike engineers, they have not built those systems to control certain variables. Unless you know what a control system has been built to control, it is difficult to tell what variable(s) a control system is controlling by simply looking at what the system is doing. This is because control systems control their inputs, not their outputs. For example, the thermostat doesn’t really control room temperature as an output. It controls room temperature as a sensed input. The sensor is often a coiled bimetallic strip that expands and shrinks as the air temperature increases and decreases. The state of this sensor is the input to the thermostat; the thermostat system acts to keep this input matching the reference set by the user of the thermostat. The thermostat, like all control systems, controls the temperature it senses e its input. This fact about how control systems work has probably been noticed by many astute observers who have tried to control the temperature in one room by setting the thermostat in another; the thermostat only controls the temperature in the room where the thermostat’s sensor is located. So the way a control system is represented in Fig. 2.1 e as a system that controls its output – gives a misleading picture of how control systems actually work. This way of representing a control system makes sense to engineers doing forward engineering but it gives the false impression to psychologists doing reverse engineering that it is possible to know what a system is doing without knowing what variables it is controlling. This problem can be solved by representing a control system in a way that makes it clear that knowing what the system is controlling is essential. Such a representation exists in the form of Perceptual Control Theory (PCT), an application of control theory that is particularly relevant to reverse engineering control systems in general and living control systems in particular (Powers, 1973; 2005).

Perceptual control theory: control of perception The PCT model of a control system is shown in Fig. 2.2. The control organization shown in that figure is functionally identical to that shown in Fig. 2.1; the difference is mainly one of point of view. While Fig. 2.1 shows a control system from the “outside” point of view of an the engineer who is building a system to control a particular variable, Fig. 2.2 shows a control system from the “inside” point of view e that of the control system itself. The inside and outside views of a control system differ mainly in terms of how the controlled variable, y(t), is represented. In the outside view (Fig. 2.1) the controlled variable is outside the system and is being compared to a reference specification, r(t), that is also outside the system. So the outside view

Understanding purposeful systems Chapter | 2

13

FIG. 2.2 PCT model of a control system; control theory for psychologists.

implies that the controlled variable exists only as a variable in the system’s environment and that the reference for the state of that variable is also in the environment. The comparison between controlled variable and reference specification is then shown as happening outside the system so that the input to the system is the difference between these two variables e the error variable, e(t) e which drives the actions, o(t), that keep the controlled variable matching the reference. But the reference specification for the desired state of a controlled variable exists inside a living control system, not in the environment. So the comparison of controlled variable to reference specification must be happening inside the control system. And what is compared to the reference specification cannot be the controlled variable itself but, rather, some analog of that variable inside the system. This fact is made explicit in the inside view of a control system. The inside view in Fig. 2.2 shows the controlled variable, y(t), in the system’s environment with the reference specification for the state of that variable inside the control system where it belongs. The controlled variable is also inside the system in the form of a perceptual signal, p(t), which is produced by the system’s input function, Yi. The input function, YI, defines the controlled variable in the sense that is determines the aspect of the environment that the system controls. The result is a perceptual signal that is an analog of y(t), the aspect of the environment that is defined by the input function, I. So the inside view in Fig. 2.2 shows that the control system compares the reference specification, r(t), to a perceptual analog of the controlled variable, p(t). The result is an error signal, e(t), that drives actions, o(t), that affect the controlled variable, y(t), and, thus, the perceptual analog of this variable. The result is that the system keeps the perceptual signal, p(t), and the controlled variable, y(t), in the state specified by the reference.

14 SECTION | A Why do we need perceptual control theory?

The inside view of a control system shows that the behavior of the system is organized around the control of its own perceptions; perceptions that correspond to aspects of the environment that are defined by the perceptual input functions of the control system. This means that in order to understand the behavior of a control system an observer has to be able to figure out what aspect of the environment the control system is controlling; that is, the observer would have to determine what perceptual variables the system is controlling. This is equivalent to determining the perceptual functions that define the aspects of the environment that correspond to the perceptions the system controls. Engineers know what aspects of the environment their control systems control because they build the perceptual functions that produce the perceptual variables that their systems control. Engineers know what aspect of the environment their thermostats control because they build the perceptual functions that produce a signal that is an analog of air temperature; and they know what aspect of the environment their cruise control systems control because they build the perceptual functions that produce a signal this is an analog of the speed of the car. So the perceptual function can be left out of the forward engineering view of a control system in Fig. 2.1. For the psychologist who studies control systems, however, the determination of what variables a system controls is central to understanding the behavior these systems; the perceptual function is, therefore, an essential component of the reverse engineering view of a control system. So Fig. 2.2 shows that the first step in reverse engineering a living control system must be the determination of what variable(s) it is controlling. This is equivalent to trying to figure out the nature of the perceptual function, YI, in Fig. 2.2.

Reverse engineering a robot One way to illustrate the process of reverse engineering a living system is to show how you might go about reverse engineering a simulated living system. An example of such a system is the “mobile inverted pendulum” (MIP) robot that was developed as a class project by students at UCSD (Bewley, 2013) and can be seen in a video available at YouTube.1 One frame of the video is shown in Fig. 2.3. A robot is a nice system to use to illustrate the process of reverse engineering because it is an artificial control system that acts a lot like a living one and, unlike a living one, we know what perceptions it was built to control, and how it implements this control, so we can check to see if the result of the reverse engineering process is correct. The reverse engineering process begins by observing the system under study to see if it seems to be controlling anything. A brief look at the video of

1. http://www.youtube.com/watch?v¼_Q4kLxzaHBg.

Understanding purposeful systems Chapter | 2

15

FIG. 2.3 One frame of the YouTube video of the balancing MIP robot.

the MIP robot would suggest that the robot is controlling its balance in the sense that it is keeping itself standing upright on two wheels; and this upright stance is maintained even in the face of disturbances, such a push, that would be expected to knock it over. So the balancing behavior of the MIP robot fits the definition of control: maintenance of a variable in a reference or goal state, protected from disturbance. While the variable that seems to be controlled by MIP is its upright stance, we know from Fig. 2.2 that what the robot is actually controlling is a perceptual variable that is an analog of some aspect of the robot’s environment that corresponds to upright stance. So the next step in the reverse engineering process is to come up with some guesses e hypotheses e about what that perceptual variable might be. This requires some understanding of how upright stance might be maintained. What we can see is that the robot is keeping itself upright by moving forward or backward, as necessary. So the perceptual variable that is being controlled must be some analog of upright stance that is affected by forward and backward movement. And there are at least two possibilities. One is that the robot is controlling a perception of the spatial deviation of the body from plumb; in this case the robot would have gyros that sense its orientation in space. Or the robot might be controlling a perception of itself relative to visual vertical; in this case the robot would have a camera that produces a perception of the angle between its body angle and visual vertical. The next step in reverse engineering is to test your hypotheses about the perceptual variable that the robot is controlling. We do this by applying disturbances that would affect the hypothesized controlled variable if it were not under control. For example, to test whether the robot is controlling a perception of its spatial orientation relative to vertical we have to think of things we could do that would change this variable if the robot were not controlling it. The simplest thing to do would be to push on the robot in a way that would change its orientation relative to vertical if it were not controlling

16 SECTION | A Why do we need perceptual control theory?

that perceptual variable. And we can see from the demo that pushing did not affect the robot’s orientation relative to vertical; it remained nearly perfectly upright. So that is evidence that this robot is controlling its spatial orientation and it is likely built with a gyro that produces a perception of that variable. However, as with any research, when we do experiments we have to be aware of the possibility of confounding variables. When you push the robot you are disturbing not only its spatial orientation but also its visual alignment with vertical. So the push experiment confounds two perceptual variables that might be controlled: spatial and visual orientation. One way to solve this confound is to find a way to disturb one of these variables without disturbing the other. One obvious way to do this is by applying a disturbance to the perception of visual orientation; such a disturbance can be applied in a way that has no effect on the perception of spatial orientation. This can be done by placing the robot in a room with vertical lines projected on the wall. If the robot is controlling a perception of visual orientation, shifting the lines forward or back should cause it to fall as it tries to compensate for this disturbance by realigning itself with those lines. If movement of the lines does not cause the robot to fall, then you can conclude with some confidence that the robot is not controlling its orientation relative to visual vertical; the disturbance caused by the movement of the lines is completely effective. And since the robot remains upright you can also conclude that the robot is likely controlling its spatial orientation. In fact, the MIP robot is controlling a perception of spatial orientation provided by gyros and it is not controlling a perception of its orientation with respect to visual vertical. This would have been confirmed if the MIP robot had been placed in the room with moving vertical lines; moving the lines would not have caused MIP to fall over.

Testing for controlled variables: reverse engineering living control systems The process of reverse engineering a living control system is the same as that used to reverse engineer a robot simulation of such a system. It starts only after we have determined that the system under study is, indeed, a control system e that it is controlling some variable(s). We did this with the MIP robot by looking at its behavior “through control theory glasses” (Marken, 2002). That is, we looked to see if the system was doing anything that was surprisingly consistent; surprising in the sense that is not expected based on our understanding of how the physical world works. In the case of the MIP robot we saw that it balanced on an axel, maintaining an upright stance even when it was pushed, the surprise being that such a push should have resulted in the robot falling over. In the case of living control systems, we can see them producing many surprisingly consistent results all the time. An example is catching a ball, a result that is achieved with surprising consistency given the fact that on

Understanding purposeful systems Chapter | 2

17

each catch the ball moves in a different trajectory and the catch is made by running toward the ball over different terrains (Marken, 2001). Once we know that the system seems to be controlling something, we can go about the process of trying to figure out what it is controlling; that is, we can start testing to see what perceptual variable(s) it is controlling. The method we use is the control theory-based “Test for the Controlled Variable” or TCV (Powers, 1979; Runkel, 2003). It is the method we used to reverse engineer the MIP robot. Formulating Hypotheses about Possible Controlled Variables. The TCV starts with the formulation of a hypothesis or guess about what aspect of the environment e what perceptual variablee the system might be controlling. These hypotheses come from observation of the system’s behavior e in particular the results that the system seems to be producing “on purpose” – as well as some understanding of how various aspects of the system’s environment might be perceived. For example, in the case of the MIP robot, hypotheses about what it might be controlling came from observation of the fact that one result the robot produced “on purpose” was balancing itself on an axel; the robot was controlling its upright stance. An understanding of how upright stance might be perceived led to two hypotheses about the perception the robot might be controlling: a perception of its orientation in space and its orientation relative to visual vertical. The development of hypotheses about the perception(s) controlled by a control system is probably the most difficult step in the TCV because it requires considerable knowledge of how things work as well as a bit of creativity in order to come up with alternative hypotheses when the current hypothesis doesn’t pan out, as it often doesn’t. This point is illustrated nicely by the “Coin Game”, a version of the TCV described by Powers (2005, p. 236). In this game, one participant is asked to determine the arrangement of a set of four coins that is being maintained in a particular condition by the other participant. The “condition” being maintained can be any perceived aspect of the coin arrangement, such as a particular pattern (like an “X”) or a relationship between the coins, such as “increasing size from left to right”. Determining what condition is being maintained is equivalent to testing for the perceptual aspect of the coins e the perceptual variable e that is being controlled. And there are many possible hypotheses about the aspect of the coins that is being controlled, many of which could be quite esoteric and, thus, hard to think of, such as “heads if the mint date is even and tails if the mint date is odd”. Looking for Compensation for Disturbance. Once we have a hypothesis about a perceptual variable that is being controlled we can test this hypothesis by applying disturbances that should have an effect on this variable if it is not under control. For example, when our hypothesis was that the MIP robot is controlling a perception of its orientation with respect to visual vertical, we did something similar to what Lee & Aronson (1974) did to test for control of upright stance in toddlers: they moved the walls surrounding the toddler so that

18 SECTION | A Why do we need perceptual control theory?

vertical lines on the wall shifted forward or backward relative to the child. If the toddler were controlling a perception of a visual perception of its orientation then the movement of the walls would cause the toddler to move to realign itself with the vertical lines. Since the lines were moved quickly, this should cause the toddler to lose its balance and fall. And indeed, it did; Lee and Aronson showed that you could get a toddler to fall without even touching it. Unlike the MIP robot, human toddlers seem to control upright position by controlling a perception of their visual orientation with respect to visual vertical. An important characteristic of the TCV is that it is an iterative process. If you apply a disturbance to a hypothetical controlled variable and find that it does have the expected effect e or very close to it e then you know that that variable is not under control. But things don’t just end there. The next step is to come up with a new hypothesis about the variable under control and test again to see if disturbances that should affect this variable actually do. If they do, then you still have not found the variable that is under control and again you have to come up with a new hypothesis about the variable being controlled. You continue this process of hypothesizing and testing until you have found a definition of the controlled variable that passes the test; that is, a definition of the controlled variable that is not affected by disturbances that should have an effect if that variable were not under control. TCV as Art and Science. The steps in the TCV are these: hypothesize what variable is under control, apply disturbances that should affect the variable if it is not under control, look for lack of effect of the disturbances via compensating actions, and if there is an effect start the testing over with a new hypothesis about the controlled variable. These steps are repeated until the researcher has a pretty good idea of what perceptual variable is under control. A more detailed description of the steps involved in doing the TCV can be found in Runkel (2003). But the TCV should not be considered a recipe to be followed by rote but, rather, a set of principles that can serve as the basis for various approaches to identifying the perceptions controlled by a living control system. Different methodologies can be used to carry out the TCV depending on the “real life” circumstances in which the controlling is observed. In the lab, the TCV can be done using controlled experimentation (Marken, 1989). Even in the lab, however, it may be difficult to “see” the controlled results that correspond to the perceptual variables that are under control because these results occur at the sensory surface of the control system. This is the case, for example, in laboratory studies of how humans intercept moving objects, such as baseballs and Frisbees (McBeath, Shaffer, & Kaiser, 1995; Shaffer, Krauchunas, Eddy, & McBeath, 2004). The controlled variables in this case are aspects of the optical trajectory of the object as the pursuer moves to intercept it; the pursuer is controlling on optical result that exists only on the sensory receptors of the retina. In this case, the manipulation of disturbances in

Understanding purposeful systems Chapter | 2

19

the lab situation must be supplemented with computer modeling (Marken, 2005; Shaffer, Marken, Dolgov, & Maynor, 2013). In computer modeling, the hypothesized controlled variable is put into a computer simulation of a control system. The simulated control system operates in the same environment as the real one being tested in the lab. The behavior of the simulation controlling the hypothesized controlled variable is then compared to that of the real control system (usually a human but sometimes another organism, such as a dog). If there is a good fit of model to data then that is evidence that the variable controlled by the model is, indeed, the same as the perceptual variable controlled by the real control system. If the fit is poor, then a new hypothesis regarding the controlled variable is inserted into the model and the behavior of the model is again compared to that of the real system. The process continues until the researcher finds a definition of the controlled variable that gives the best fit of model to data (Marken, 2014a,b). In real life, the perceptions controlled by living control systems e particularly human control systems e can be quite complex. The perceptions controlled by artificial control systems, on the other hand, are usually rather simple in the sense that they can be easily measured and monitored by testing equipment, which makes the process of testing for controlled variables much more precise. But humans control perceptions like the degree of honesty of a communication or the degree of the intimacy of a relationship, which are currently impossible to measure and monitor with lab equipment. This makes reverse engineering of human control systems a particular challenge since the only known device that can measure variables like “honesty” and “intimacy” is the human brain itself. This means that reverse engineering human control systems to determine the kinds of complex perceptions they control is going to be somewhat of an art that relies on “subjective” methods, such as having expert human observers evaluate the degree to which a variable like honesty has been affected by disturbances. An example of this approach to reverse engineering human control systems was described by Robertson, Goldstein, Mermel, & Musgrave (1999) who used collections of 3  5 cards with self-descriptive adjectives to represent the state of a complex perceptual variable e the state of one’s self-concept e and non-self-descriptive adjectives to serve as disturbances to this variable. Finally, an important consideration when using the TCV is that living control systems control many perceptual variables at the same time and the reference specifications (goals) for the state of some of these perceptions are varied as the means of controlling other perceptions (Powers, 2005). For example, the reference or goal level for intimacy with another person will vary, depending on whether a higher-level goal is to get a job from, or to marry, that person. The implication of this for reverse engineering using the TCV is that a single failure to protect a hypothetical controlled variable from disturbance is not enough to rule out that variable as being under control. For example, person A may fail to compensate for the stand-offish behavior of person B, not

20 SECTION | A Why do we need perceptual control theory?

because person A is not controlling for intimacy, but because his or her desired intimacy with person B is very low, so there is no need to compensate for the stand-offishness. Changing the disturbance from “stand-offish” to “enticing” might lead to compensatory actions that would suggest that something like “intimacy” is, indeed, under control.

Conclusion The engineer and the psychologist have two very different reasons for using control theory in their work. The engineer uses control theory as the basis for understanding control systems in order to build systems that control better; the psychologist uses control theory as the basis for understanding control in order to see how these systems control. The difference in these approaches to using control theory is equivalent to that between the processes of forward and reverse engineering, respectively. Psychologists who have used the forward engineering approach to understanding living control systems have bypassed the first question that should be asked when trying to reverse engineer these systems: what is the system designed to control? PCT, which provides a guide to reverse-engineering control systems, shows that control systems are designed to control certain perceived aspects of their environment. So the answer to the question of what a living system was designed to control is the identity of the perceptual variable(s) it controls. Such identification can be found using the Test for the Controlled Variable (TCV), which is actually a set of methodologies that allow the psychologist to make inferences about a control system’s subjective experience (its controlled perceptions) using methods that involve applying disturbances to hypothetical controlled perceptions and watching for lack of effect. The TCV provides the tools for a revolutionary new approach to understanding the purposeful (control) behavior of living organisms; an approach that uses reverse engineering to determine the perceptual variables around which their behavior is organized.

References Arents, R. R. D., Groeneweg, J., Borst, C., van Paassen, M. M., & Mulder, M. (2011). Predictive landing guidance in synthetic vision displays. The Open Aerospace Engineering Journal, 4, 11e25. Bewley. (2013). Mobile inverted pendulum class project for MAE 143C. UCSD. Craik, K. J. W. (1947). Theory of the human operator in control systems. British Journal of Psychology, 38, 56e61. Craik, K. J. W. (1948). Theory of the human operator in control systems. British Journal of Psychology, 39, 142e148. Howell, W. C. (1971). Engineering psychology: current perspectives in research. New York: Appleton-Century-Crofts.

Understanding purposeful systems Chapter | 2

21

Jagacinski, R., & Flach, J. (2002). Control theory for humans: quantitative approaches to modeling performance. NJ: Erlbaum. Lee, D. N., & Aronson, E. (1974). Visual proprioceptive control of standing in human infants. Perception & Psychophysics, 15, 529e532. Lewis, F. L. (1992). Applied optimal control and estimation. Prentice-Hall, 1992. Marken, R. S. (1989). Behavior in the first degree. In W. Hershberger (Ed.), Volitional action: conation and control (pp. 299e314). Amsterdam, The Netherlands: Elsevier. Marken, R. S. (2001). Controlled variables: Psychology as the center fielder views it. American Journal of Psychology, 114, 259e281. Marken, R. S. (2002). Looking at behavior through control theory glasses. Review of General Psychology, 6, 260e270. Marken, R. S. (2005). Optical trajectories and the informational basis of fly ball catching. Journal of Experimental Psychology: Human Perception and Performance, 31, 630e634. Marken, R. S. (2014a). Testing for controlled variables: A model-based approach to determining the perceptual basis of behavior. Attention, Perception, & Psychophysics, 76, 255e263. Marken, R. S. (2014b). Doing research on purpose: a control theory approach to experimental psychology. St. Louis: MO: New View. McBeath, M. K., Shaffer, D. M., & Kaiser, M. K. (1995). How baseball outfielders determine where to run to catch fly balls. Science, 268, 569e573. McRuer, D. T., & Krendel, E. S. (1959). The human operator as a servo system element. Journal of the Franklin Institute, 267, 381e403. Powers. (1973). Behavior: the control of perception. New York: Aldine-De Gruyter. Powers, W. T. (1979). The nature of robots. Pt. 4: Looking for controlled variables. Byte, 4, 96e118. Powers. (2005). Behavior: The control of perception. In New canaan, conn (2nd ed.). Benchmark Publications. Robertson, R. J., Goldstein, D. M., Mermel, M., & Musgrave, M. (1999). Testing the self as a control system: Theoretical and methodological issues. International Journal of HumanComputer Studies, 50, 571e580. Runkel, P. (2003). People as living things. The psychology of perceptual control. Hayward, CA: Living Control Systems Publ. Shaffer, D. M., Krauchunas, S. M., Eddy, M., & McBeath, M. K. (2004). How dogs navigate to catch Frisbees. Psychological Science, 15, 437e441. Shaffer, D. M., Marken, R. S., Dolgov, I., & Maynor, A. B. (2013). Chasin’ choppers: Using unpredictable trajectories to test theories of object interception, attention. Perception & Psychophysics, 75, 1496e1506. Sheridan, T. B., & Ferrell, W. R. (1974). Man-machine systems: information, control, and decision models of human performance. Cambridge: MIT Press. Wickens, C. D., Hollands, J. G., Parasuraman, R., & Banbury, S. (2012). Engineering psychology & human performance (4th ed.). Upper Saddle River, NJ: Pearson.

Chapter 3

The crisis in neuroscience Henry Yin Department of Psychology and Neuroscience, Duke University, Durham, NC, United States

In the last century, rapid progress in our understanding of the nervous system led to the emergence of the field now commonly called neuroscience. The appeal of neuroscience is obvious. As the nervous system is necessary for behavior, a fact known since antiquity and now familiar even to school children, neuroscience is expected to provide an answer to the question of how human behavior can be explained. But it is precisely this question that still eludes us. For most progress in neuroscience has been limited to delineating how different neurons are connected to each other and how they communicate with electrical and chemical signals. Such work has produced a massive literature, impressive in its scope and depth, but it has not yet revealed how the brain works as a whole or how it produces behavior. In fact, contrary to popular belief, neuroscientists still do not understand how the nervous system of any organism produces any type of behavior. Even the behavior of the roundworm C. elegans remains unexplained. Although neuroscience textbooks are full of colorful illustrations and diagrams, they do not contain working models of behavior. They do not contain system equations that specify how the different parts of the brain are related functionally. Other than biophysical models such as the HodgkinHuxley model of the action potential (Hodgkin & Huxley, 1952), neuroscience lacks the kind of model found in the physical sciences. The failure to explain behavior is seldom acknowledged by neuroscientists, and apparently unknown to the lay public. Even when acknowledged, opinions differ as to why, despite our detailed knowledge of neural structure and signaling, we still cannot explain behavior. Some think that there is too much noise in the sensory input, or that behavior is too variable (Osborne, Lisberger, & Bialek, 2005); while others think that mechanisms of behavior could be fundamentally non-deterministic, invoking the ubiquitous uncertainty principle from quantum mechanics (Penrose, 1999; Neuringer, 2004; Glimcher, 2005). Still others think that, because the brain is so complex, to understand it we must first map the connections of every neuron and record all the signals,

The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00003-4 Copyright © 2020 Elsevier Inc. All rights reserved.

23

24 SECTION | A Why do we need perceptual control theory?

generating ‘big data’ that can be processed with powerful computers (Markram, 2006; Alivisatos et al., 2012). While the complexity of the brain appears to be the best excuse for the failures of neuroscientists to explain behavior, a similar excuse can be found for any system that we do not understand. For example, suppose an alien visiting earth for the first time found a toaster and wished to understand it. The toaster, too, can appear exceedingly complex. One could, for example, study its slot width, its mechanisms of heat production, the chemical composition of its paint, and so on. To those of us familiar with the toaster, however, these approaches seem ridiculous. What is missing is a consideration of the toaster’s function. A toaster also contains many parts, if we go down to the atomic level, but knowledge of that level does not tell us that or how it toasts bread. Even the so-called connectome, a connectivity map of the whole nervous system, tells us very little about function, since the map does not tell us much about the actual signals traveling between the different parts. Not surprisingly, for a simple organism like C. elegans, for which a connectome does exist, this information has not produced any explanation for its behavior.

Paradigm and crisis Mired in details, neuroscience has focused exclusively on what can be studied in light of existing tools and conceptual frameworks from physics and chemistry, like the proverbial man looking for his wallet under the streetlight, not because he lost his wallet there, but because that is where the light is. What has not been considered is the possibility that the current paradigm in neuroscience is wrong. A paradigm, in the familiar sense introduced by Kuhn, represents a set of unquestioned assumptions in any field of science. A paradigm determines the questions asked, as well as the methods used to answer them. A crisis arises from perceived incompatibilities between the current paradigm and facts that do not appear to fit this paradigm (anomalies in Kuhn’s terminology). When more and more incompatible facts appear, a crisis is developed, to be resolved only by a scientific revolution, which establishes a new paradigm (Kuhn, 1962). Neuroscience, I shall argue, is now in the middle of such a crisis, because its current paradigm rests on shaky foundations. What are the most fundamental assumptions that form the current paradigm in neuroscience? I would suggest the following: Behavior is a function of what happens to the organism. What happens to the organism is registered by its sensory receptors and transmitted through its sensory systems. The mind is assumed to be what takes place between stimulus and response, consisting of complex processes, often associated with information processing, that can ultimately generate observable behavior. As Skinner stated: “Eventually a science of the nervous system based upon direct observation rather than inference will describe the neural states and events which immediately precede instances of behavior. We shall know the

The crisis in neuroscience Chapter | 3

25

precise neurological conditions which precede, say, the response “No, thank you.” These events in turn will be found to be preceded by other neurological events, and these in turn by others. This series will lead us back to events outside the nervous system and, eventually, outside the organism” (Skinner, 1953). Skinner’s statement is considered self-evident by most neuroscientists. But its underlying assumption represents a major obstacle to progress. Because Skinner is often considered a spokesperson for behaviorism, a school of thought that has been replaced by more enlightened ideas as a result of the “cognitive revolution,” it must be pointed out that that his basic assumption is shared by most cognitive scientists, the professed enemies of behaviorism. It would be a mistake to believe that, simply by inserting variables between stimulus and response, i.e. by focusing on the internal structure as most cognitive scientists did, one can move beyond the behaviorists. The underlying assumption, which remains unchanged, is that there are causes and effects: Sensory input are classified as causes and behavioral output as effects. Just as force applied to a rock causes it to move, so “information” or sensory input sent to the organism somehow causes behavior. What counts as cause and what effect might vary, so long as causes precede effects. This assumption of linear or unidirectional causation is unquestioned among most students of behaviordbe they reflexologists, psychophysicists, ethologists, Hullian or Skinnerian behaviorists, cognitive scientists, or systems neuroscientists. None of them ever imagined that it could be wrong, but it is. The linear causation paradigm naturally leads to the input/output approach in experimental work: manipulate input and measure output. Neuroscientists assume that, in studying the neural basis of behavior, they are studying sensorimotor transformation, “the process by which sensory stimuli are converted into motor commands” (Pouget & Snyder, 2000). I shall argue, however, that there is no sensorimotor transformation, that the function of the nervous system is to not to respond to stimuli, and that behavior is neither a function of incoming stimuli nor of internal states. Rather the nervous system functions to control some set of perceptual inputs (Powers, 1973b). Unfortunately, the concept of control has been so widely misunderstood, both by engineers who designed the first control systems and by cyberneticists who attempted to apply control theory to the study of living organisms, that it is a formidable challenge to reveal the common misunderstandings and to explain what control means. As Mark Twain is reported to have said: “It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t true.” In contemporary neuroscience, the list of assumptions that just ain’t true is long indeed, so patience is required as I expose each in turn. I shall first examine the relationship between observable behavior and control, and outline the basic tenets of control theory. I shall then explain various misunderstandings of control theory and the resulting inappropriate applications to the study of

26 SECTION | A Why do we need perceptual control theory?

behavior. Whenever possible, I shall illustrate common conceptual confusions using specific examples from neuroscience.

Behavior is not solely determined by neural output To understand what control is, it is necessary to start with the question of what behavior is, because what we observe as behavior is merely the outward manifestation of the more fundamental process of control. We can easily observe behavior, but we cannot just as easily observe control, which is enabled by a specific organization inside the control system. The most common fallacy is found in the following statement from a recent review on behavior: “Behavior is “what animals do”. It can be defined as the muscular output of an organism or, alternatively, as the externally observable dynamical features of an organism” (Gomez-Marin, Paton, Kampff, Costa, & Mainen, 2014). The fallacy here is to equate the muscular output with the observable dynamical features. When we observe any organism behave, it is easy to attribute the observed behavior to the muscular or neural output. But this is an illusion. Neural output at the final common path from alpha motor neurons in the spinal cord to muscles is necessary for normal behavior as observed. But necessity does not equal sufficiency. Neural output is necessary but not sufficient. As Bernstein first pointed out, what we call behavior is not the sole result of neural output (Bernstein, 1967). Take the simplest example of standing: when standing the neural output sent to the muscles via the final common path is indeed producing muscle contraction and exerting forces. But there are other forces acting on the body (e.g. gravity, wind) that determine the behavior of standing as observed. Any behavior is the result of two types of influences, one from the organism’s nervous system, and the other from the environment. From the naı¨ve perspective, what is seen is the contribution of the organismdits output. What is not seen is the contribution of the environment, the forces that are acting upon the body just as neural signals are amplified to generate torque through muscle contraction. There is, in fact, no one-to-one mapping between muscle contraction and any posture or movement. The causal chain commonly envisioned, from neural output to muscle contraction to movement, is largely imaginary. We cannot tell by muscle output alone what the movement actually is, just as we cannot predict the speed or position of a car by measuring engine output. Even if we obtain a complete recording of every motor neuron or every muscular contraction, this recording, when played back again, would not reproduce the same behavior. The first step toward an understanding of behavior, then, is to appreciate that neural output is insufficient to determine behavior as observed. This can be called the “insufficiency” principle. Because behavior is not the sole product of neural output from the final common path, repeating neural output precisely is not sufficient to repeat the same behavior.

The crisis in neuroscience Chapter | 3

27

Calculation problem and control Behavior, which we take for granted, represents a remarkable achievement that is rarely noticed (Brunswik, 1943): In spite of environmental disturbances, behavior is still achieved successfully most of the time. The behavior is a controlled result, the visible portion of a hidden tug of war between invisible environmental disturbances and actual neural output. The neural output from the final common path must always vary according to the changes introduced by the unknown disturbances. How can the necessary variations be generated by the nervous system to cancel the effects of the disturbance exactly? If the source of disturbance is often unknown and its magnitude unpredictable, how does the brain know how much output to produce, and when? This is what I have called the calculation problem (Yin, 2013). Although the challenges presented by the calculation problem is widely appreciated, the usual solution betrays a key misunderstanding. It is often assumed that the nervous system contains internal models of physics and computes the inverse dynamics and kinematics needed to drive the effectors (McIntyre, Zago, Berthoz, & Lacquaniti, 2001; Green & Angelaki, 2010). It is also assumed that the brain can anticipate the effects of future disturbancesdsuch as exactly how the wind will change or how the fatigue of muscle fibers will alter the properties of the effectorsdto generate the exact motor commands sent to the muscles. In other words, it is assumed that the brain is a computing device with a complete knowledge of the physical environment as well as all the correct equations for the kinematics and dynamics, and that this is true for all animalsdthe human as well as the cockroach. If this is not an example of a faith in miracles, it will do until a better one comes along. But as illustrated by decades of failures in artificial intelligence and robotics, in any realistic environment these approaches cannot solve the calculation problem, which becomes overwhelming due to unexpected disturbances and high degrees of freedom. There is, however, a far simpler solution to the calculation problem. Negative feedback control systems can solve it without performing the inverse calculations, without knowing what the disturbances are or where they come from, without internal representations of the physics of the environment or feed-forward computations. The function of closed loop negative feedback control systems is to control. Control is the process that brings some variable to a desired value, or to maintain that value, despite disturbances that would otherwise produce deviations from the desired value. Regardless of the source and nature of the disturbance, the system generates the appropriate behavior to reduce the deviation. Consistent ends are the sensed variables to be controlled, and variable means are the outputs.

28 SECTION | A Why do we need perceptual control theory?

Control of input A control system can only control its input. That is, some input variable is defended against disturbances as it is maintained or brought to a new value by the actions of the control system (Powers, Clark, & McFarland, 1960). Coincidentally, misunderstanding or neglect of this principle is also the major source of confusion regarding closed loop control in the life sciences and the primary reason that negative feedback control has been rejected by so many neuroscientists as an explanation of behavior (Robinson, 1990; McIntyre & Bizzi, 1993; Shadmehr & Wise, 2005). In the control system (Fig. 3.1), the input comes from sensors, but it can also be derived from some lower order perceptual signal, which is more often the case in the nervous system. The main features of the control system are as follows (Powers, 1973a, b). (1) Control of input: The perceptual input signal is sent to a comparator, which also receives a reference signal. The difference between the two is called the error signal, which is used to produce the output.

FIG. 3.1 Top, the control loop as commonly illustrated, made popular by Wiener and control engineers. Bottom, the control loop according to Powers. The effect of the perceptual signal is shown to have a negative sign, but it is also possible for this signal to have a positive sign, i.e. having a net excitatory effect on the comparator units, so long as the reference signal is opposite in sign. What is required is that the comparator produces a signal representing the discrepancy between reference and perception.

The crisis in neuroscience Chapter | 3

29

(2) Internal reference: Control is possible because the system contains the equivalent of internal reference signals. Reference signal represents what some input signal should be for that control system. Because the system produces output that, via feedback, reduces the discrepancy between what the input variable should be and what it is, it is capable of reaching the desired or referenced perception. Disturbances are those effects that push the value of the controlled variable away from the value specified by the reference signal, generating the error signal. (3) Variable output: To match the input to the reference signal, the controller uses its output, which can affect the input variable through some environmental feedback path. Because the output is proportional to error, it can be highly variable even when the input or reference remains stable. In a hierarchy of control systems (e.g. living organisms), the variable outputs of the higher levels cannot be observed unless neural activity is recorded. Observable behavior, produced by muscles, represents the outputs of the lowest level of the hierarchy. (4) High loop gain: Negative feedback controllers are characterized by high loop gain. At steady state, after the controller is allowed to exert its effect, the effect of the disturbance is rejected, and how much error remains is determined by the loop gain. The higher the gain, the sooner will the system reduce error and reach steady state, and the smaller the remaining error. A high gain does not mean a lot of output will be produced given a small input, as in an input/output device; while this is true of the output function by itself, it is not true of the system as a whole when the loop is closed, as the output produced is reduced by itself. The high gain makes the system faster and more accurate. (5) Negative feedback: The net effect of the output on the input is described by the feedback function in the environment. But in a negative feedback controller error is always used to reduce itself, via the feedback function in the environment. Whatever the effect of the disturbance, that effect is canceled by the behavioral output. Whenever the feedback reduces the discrepancy, there is negative feedback. The control system pushes back whenever a controlled variable is pushed outside of its range as determined by the reference signal. By definition, it is only negative feedback that can achieve control, whereas positive feedback only increases the discrepancy between reference and input, so that the input will deviate further from the desired value, i.e. the opposite of control. (6) Circular causation: The transfer function and feedback function are simultaneous. If we examine the basic diagram of the negative feedback controller, the output (or observable behavior) is determined by two signalsdthe reference signal and the perceptual signal. It is therefore wrong to claim that either the reference or the perceptual signal can cause behavior. Control systems are characterized by simultaneous interactions in a loop with no beginning and no end. It is inappropriate to analyze the

30 SECTION | A Why do we need perceptual control theory?

control loop serially, like a reflex chain (Hull, 1934). The serial analysis assumes that during the time in which the output occurs, the input does not change, and during the time it takes for the input to change, the output does not change. But values in these paths do not take turns when they change. (7) Hierarchical organization: Living organisms contain multiple control systems, arranged hierarchically. The higher-level perceptions are functions of a subset of the lower-level perceptions. Only the lowest level uses muscles to act on the external environment. Negative feedback controller can only be commanded by altering its reference signal. Any attempt to alter its output directly, without adjusting reference at the same time, will make the input deviate from the reference signal, thus producing a corrective action that cancels the effect of the output manipulation. The higher system tells a lower system how much perception to create.

Misunderstanding control Although control is a simple concept, it has been persistently misunderstood (Rosenblueth, Wiener, & Bigelow, 1943; Craik, 1947; von Holst & Mittelstaedt, 1950; Ashby, 1956, 1958). Perhaps Wiener, the founder of cybernetics, was responsible for the major error (Rosenblueth et al., 1943; Wiener, 1948; Ashby, 1956; Powers, 1973a). Despite his enormous achievements, in his attempt to apply systems analysis to the study of living organisms, Wiener merely followed engineering conventions in his thinking on the control loop (Wiener, 1948; Powers et al., 1960). His illustration incorrectly describes the interaction between the organism and the environment. Wiener’s error can be appreciated when we compare his diagram with the diagram according to perceptual control theory shown in Fig. 3.1 (Powers et al., 1960; Powers, 1973b). The key question is which parts of the loop (input, reference, output, feedback) correspond to the organism. In Wiener’s illustration, the reference signal is labeled as an input, which is compared with the feedback from the output, producing the error signal that is transformed into output. Thus his model puts the comparator outside the organism. It assumes that the organism’s role in the closed loop is to convert the error signal into outputdwhich happens to be the role played by the output function in the control loop. The Wiener model, the controlled variable is not the input to the organism, but rather the output. The “output” is actually some effect of the actual output, as measured by some observer, e.g. the engineer who is designing and testing the system. What the engineer calls the output is his perception of the system output (sometimes called the “process variable”). He wants to achieve certain outputs, i.e. some perception of the system output. The design allows him to “command” the system by injecting certain reference signals (sometimes

The crisis in neuroscience Chapter | 3

31

called “control signals”). The error signal is not generated by comparing some internal reference with the perceptual input. Rather the reference is an input introduced by the user, and this signal is compared with the actual output of the system (labeled feedback), as measured independently by the observer. The convergence of these two signals generates the requisite error signal that is turned into the output. The input variable is affected by both disturbance and feedback, but in opposite directions. Like the thermostat in action, it does not need to sense the disturbance (outside temperature) directly. So long as the loop is closed, its input is protected by its output via the feedback function. Moreover, reference signals are not inputs to the organism, but are mostly found inside the brain. Consequently, the comparison function does not take place in the environment, but inside the organism. Wiener was under the influence of the linear causation paradigm. He treats the controller as an input/output device: error in, behavior out. Of course, any particular component of the control system, in isolation, is simply an input/ output device, and as such, it cannot control anything. Control can only be an emergent property of the entire closed loop, not a property of individual components of the loop. Engineers view the control system from the perspective of the designer or user. The goal of the engineer is to generate a driving signal that somehow produces the appropriate outputs. The user of the thermostat knows what kind of setting to use, because he can sense the errordthe current temperature is either too high or too low, for him. Consequently, the output is the effect that the engineer wishes to achieve. The input is the setting that the user can adjust, according to his preference. In the Wiener diagram, the comparator function compares the output of the system with the desired value. For example, the velocity of the actuator is measured and compared with the value desired by the experimenter or observer. The observer must then come up with the right driving signal to the motor plantdwhich represents the right reference signal to be injected into the system. The controller does not determine its own reference. It is up to the user to decide what it will do, based on the user’s perception of its feedback. It is still control of input, except that the user functionally serves as a higher level in the control hierarchy, and the controller in question becomes completely subservient. Not surprisingly, in engineering, negative feedback controllers are also called “servos.” This lack of autonomy is a deliberate part of the design. When neuroscientists talk about control theory, it is usually Wiener’s model that is used, which inevitably leads to either confusion or mistakes (Robinson, 1981; Camhi, 1984; Todorov & Jordan, 2002). Consequently, few considered the possibility that the nervous system can contain intrinsic reference signals or purposes, and that the comparison function takes place not in some ideal observer but in the organism itself. This confusion has a momentous consequence. What happens outside the organism is now placed

32 SECTION | A Why do we need perceptual control theory?

inside the organism. Inverse calculations on kinematics and dynamics are now attributed to the organism, especially the brain, and research efforts have been wasted on finding neural correlates of these computational processes (Wolpert & Kawato, 1998; Pouget & Snyder, 2000; Todorov & Jordan, 2002).

Observer’s bias The second obstacle to a correct understanding of control theory is a strong bias in the observer of behavior. In the paradigm borrowed from classical physics, science is assumed to be objective, in the sense that observations must be made from a third person perspective. This is how one would measure the motion of physical objects, as Galileo did when he rolled balls down inclined planes or dropped balls from the tower of Pisa. How things appear to the ‘object’ being measured is irrelevant, because the ball is assumed to have no perceptions. This assumption is correct in physics, but it leads to absurdity when applied to the study of living organisms. As a result of this bias, it is considered unscientific to take into account subjective perceptions. Whenever the question of subjective perception is brought up, life scientists tend to feel uncomfortable, feeling the onset of a “philosophical” discussion. But the third-person perspective has key limitations when applied to the study of living organisms. The observer does not have direct access to the perception of another organism. In the standard linear causation paradigm, the subjective perceptiondhow things appear to the organismdis either ignored as in behaviorism, or even worse, unwittingly imposed by the observer as in much of contemporary neuroscience. The latter error, equating the perception of the experimenter with the perception of the organism under study, is the observer’s bias that has persistently impeded progress in neuroscience. To understand control in living organisms, subjective perception is of paramount importance, because all controlled variables are perceptual variables. Leaving aside the question of conscious perception, the perceptual signals in any control system can be identified. To understand the system it is necessary to know what such signals represent.

Input/output analysis and behavioral illusion Suppose we observe that a stimulus precedes a behavior. When we vary the stimulus, the behavior will also vary, allowing us to establish the behavioral output as a function of the input. This type of input/output analysis is the basis for psychophysics and the study of the neural substrates of behavior (Stevens, 1975; Britten, Shadlen, Newsome, & Movshon, 1992). It is generally accepted as the only appropriate way to conduct scientific experiments on behavior, but it is precisely this approach I shall challenge here. Within this paradigm, the role of the scientist seems straightforward: (1) to identify the adequate stimulus that triggers the response; (2) to identify the

The crisis in neuroscience Chapter | 3

33

function that expresses the response as a function of the identified stimulus; (3) to elucidate the neural mechanisms that implement this function. But this approach, applied to control systems, produces a conceptual trap. This trap is particularly vicious as it seems so innocuous on the surface. As Powers first noted, the appearance of cause and effect in behavior can be a powerful illusion (Powers, 1973b). The observation is clear enough: a stimulus is presented, and a response is reliably generated. What is the illusion? First, what appears to the experimenter or third-person observer to be a stimulus event cannot be equated with the actual input to the organism. Often it is simply a disturbance to some unidentified controlled variable. In fact, the practice of choosing the “right” stimulus to evoke a reliable response, a common practice in neuroscience, virtually guarantees that the stimulus acts as a disturbance to some unidentified control system. That is, the physical stimulus will impact the perceptual input function of some control system in such a way as to generate a significant deviation from its internal reference value, thus generating corrective outputs. When a perceptual variable is being controlled, however, the output of a controller will not be related to the actual perceptual input to the organism. The input might be constant, like temperature in the air-conditioned room, but the output of the thermostat will vary according to the effect of the disturbance on the perceptual variable. Instead, the output will be correlated with the disturbance, precisely because it opposes the effects of the disturbance. What is hidden from view is the input variable being disturbed, the controlled variable that is simultaneously being disturbed by the “stimulus” and pushed back by the behavior. The more successful the control system, the more will the output resemble the disturbance. Meanwhile, the hidden input variable is protected from the effect of the disturbance. The disturbance does not get to “cause” much deviation because the output is reducing the error at the same time. Trained in the traditional paradigm, neuroscientists find it difficult, if not possible, to understand the closed loop circular causation just described. Suppose a control system is controlling a perceptual input, and the stimulus presented by the experimenter also introduces a disturbance on this controlled variable. The output will change the input accordingly, by opposing the effect of the disturbance. Thus the perceptual variable is simultaneously affected by the system output, through the feedback function, and by the disturbance introduced by the environment. The effect of disturbance on the input is a property of the environment. The feedback function, or the effect of the output on the input, is also a property of the environment. When the disturbance increases, behavioral output will increase accordingly. For the sake of convenience, suppose the reference signal for a particular controller is close to zero. If control is successful, the perceptual signal will be

34 SECTION | A Why do we need perceptual control theory?

kept close to zero. Assuming that both the feedback function and the effect on the input by disturbance is linear, then given 64 units of disturbance, the output will, via the feedback function, cancel the effect of the disturbance, forcing the input to approach zero. Because one unit of output, via the feedback function, is assumed to reduce one unit of input, the output is also 64. Correlation between disturbance and output will be high, so that the system behaves as if, for every unit of disturbance presented, one unit of output is presented (Table 3.1). The controller illustrated in Table 3.1 has a zero reference for the perception in question (e.g. pain, see discussion of Sherrington’s experiments below). The output will simply reduce the magnitude of perception. If the disturbance increases the value of the controlled variable, output will be generated to oppose that effect. Changing the feedback function, which is a property of the environment, can change the relationship between disturbance and output in a control system. Because in conventional input/output analysis, what the experimenter calls the “effective stimulus” for eliciting behavior is actually a disturbance, and the function relating input to output can simply reflect environmental properties, without revealing anything about the internal organization of the system being examined. All units are arbitrary. The neuroscientist might be satisfied by such a linear relationship: The nervous system appears to produce one unit of output for one unit of input. The remaining task, then, would be to figure out the neural mechanism underlying such a sensorimotor transformation. But suppose now the feedback function is changed, so that the feedback varies not linearly with the output, but as the square of the output. To keep the input close to zero, the output required will be 8 units. As the disturbance varies, the behavioral output will vary as the square root of the disturbance. Now the nervous system seems to have the capacity to take the square root of the inputs. Gone is the linear sensorimotor transformation, even though the organism has not changed at all. For the relationship between disturbance and output is determined by environmental properties, and shows nothing about the internal organization of the control system. It is only guaranteed by the existence of control, i.e. of outputs opposing the effects of the disturbance, precisely because these outputs are generated by the error signal. In short, negative feedback controllers can create the illusion that they are input/output devices. The input/output analysis, which applies only to linear causation systems, inevitably fails to reveal the properties of the control system. When the experimenter attempts to study behavior as a function of the stimulus, more often than not he is not describing the brain, but properties of the environment, due to the hidden feedback loop. What he believes to be the neural correlates of sensorimotor transformation are simply irrelevant environmental properties. This is the behavioral illusion. If the above analysis is correct, then a disturbingly large proportion of work on the neural substrates of behavior will have to be discarded.

TABLE 3.1 An illustration of how a conventional input/output analysis generates a property of the environment rather than the organism. Feedback (f)

Perception p¼d - f

Reference

Feedback function

Behavioral output (b)

Output as a function of disturbance

64

64

w0

w0

f¼b

64

b¼d

64

64

w0

w0

f¼b

8

b ¼ d1/2

64

64

w0

w0

f¼2b

32

b ¼ d/2

2

The crisis in neuroscience Chapter | 3

Disturbance (d)

35

36 SECTION | A Why do we need perceptual control theory?

Case study 1: Sherrington’s analysis of the reflex To illustrate the fallacies discussed above, I will use a few simple examples from neuroscience. A useful starting point is the work of Sherrington, an influential neuroscientist whose work represented the culmination of the reflex arc tradition started by Descartes (Sherrington, 1906). As Sherrington wrote: “the outcome of the reflex as expressed by the activity induced in the effector organ is a response appropriate to the stimulus imparted to the receptor. This due propriety of end-effect is largely traceable to the action of the conductor mediating between receptor and effector. Knowledge of the features of this ‘conduction’ is therefore a prime object of study” (Sherrington, 1906, p. 9). This is an example of the input/output analysis of behavior. To study the reflex Sherrington had to find the stimulus that would reliably elicit some behavioral output from the animal. What is employed as the “effective stimulus” in a traditional experiment is by definition an errorgenerating disturbance that causes a deviation of some undefined perceptual variable from the reference value. Its status as disturbance is contingent upon the reference signal, i.e. determined by the internal organization of the organism, not by the experimenter. To find a reliable relationship between the effective stimulus and the reflexive response, Sherrington had to use certain types of “preparations.” First, he focused on protective reflexes, for example, the scratch reflex which is elicited when a drop of acid is applied to the skin. But in any intact organism, even the protective reflexes are not reliable, especially when they are repeatedly elicited. Sherrington’s solution was to remove the descending influence of the brain by making a cut at the level of the midbrain, which allowed the animal to survive long enough for the experiments, but also removed nearly all of the inconvenient behaviors. The animal was then reduced to “decerebrate preparation,” staying alive but providing the appearance of linear causation. Despite the appearance of a stimulus-response sequence, reflexes still reflect the actions of control systems. By surgically removing the influence of the brain, Sherrington also removed any source of descending reference signals in the control hierarchy (Yin, 2016). The key question is not how the stimulus elicits the behavior, but how the behavior affects some controlled perception. In all cases, the reflex elicited reduces the disturbance. The protective reflexes are the outputs of control systems that protect the value of some lower level perception. They are protective because some excessive level of the perception in question is associated with injury to biological tissue. For example, in the scratch reflex that Sherrington studied, the key question is how much of a particular perceptual signal related to the acid is tolerated by the dog. Why should the behavior of scratching be so appropriate? What is missing from Sherrington’s analysis is the internal reference signal, a representation of how much of a particular type of perception should be accepted by

The crisis in neuroscience Chapter | 3

37

the sensors at any moment. Commenting on Sherrington’s work, Adrian observed, “The control may involve widespread movement: shading the eyes against too strong a light, turning the head to catch a sound, or sniffing to identify a smell. Even the flexion reflex of the spinal animal might be regarded as an example, for its aim is the removal of the pain receptors from the stimulus which has set them in action” (Adrian, 1957). Adrian realized that it is impossible to discuss reflexes without also mentioning their aim. Although he had the right intuition, he lacked the concept of the internal reference, which is literally the representation of the “aim” of behavior. When applying a drop of acid to the skin of a decerebrate dog, Sherrington did not pause to reflect how this stimulus would alter the perception of the dog (in this case the input to the lower spinal circuits). He neglected the feedback function that allows the scratching behavior to affect the perceptual input. Sherrington’s approach essentially forces the organism to act like some inputoutput system, so that some reliable relationship between input and output could be established. By treating behavior as a function of inputs, the correlation between stimulus and response appears to be a function of the nervous system, but in fact reflects the environmental properties, as illustrated above in the discussion on the behavioral illusion. Sherrington argued that the “due propriety of end-effect is largely traceable to the action of the conductor mediating between receptor and effector” (Sherrington, 1906). The action of the conductor is the illusory organism function, which can change when the reference signal or environmental properties change. Although Sherrington never produced a scientific model of any reflex, his work is regarded as a major achievement in understanding the neural substrates of behavior, with a profound impact on the subsequent development of neuroscience (Gallistel, 1980). Although modern investigators often believe they are moving beyond Sherrington by inserting internal states and other intervening variables between stimuli and responses, or by top-down modulation of stimulus-response paths, their models also share the basic assumption of linear causation (Miller & Cohen, 2001).

Case study 2: sensorimotor transformations: from flies to monkeys To illustrate the observer’s bias, I will use two examples from the study of sensorimotor transformations, both involving the visual modality and behaviors that are “triggered” by visual inputs. In both examples, control theory is incorrectly applied. The first example is optomotor behavior in the fly. The experiments discussed were performed by Mittelstaedt, and the analysis is taken from a leading textbook on the neural mechanisms of behavior (Mittelstaedt, 1964; Camhi, 1984). In these experiments, a restrained fly is placed on a platform inside a cylinder with a striped pattern. When the cylinder is rotated, the fly

38 SECTION | A Why do we need perceptual control theory?

will turn in the same direction. This behavior is interpreted in the framework of negative feedback. According to Camhi, the fly uses negative feedback in its turning behavior. The difference between the fly turning and the cylinder turning is called “slip speed.” It is considered to be the error signal in the control loop. Following Wiener’s illustration of the control loop, the slip speed error signal enters the organism and elicits the turning behavior. It is considered to be the stimulus that “causes” the turning behavior, betraying the underlying assumption of linear causation. What Camhi calls the input, the cylinder rotation, is the input from the perspective of the experimenter, which is not identical to the actual image velocity on the retina of the fly. This is an example of the observer’s bias discussed earlier. Repeating Sherrington’s blunder, Camhi never asks how the stimulus appears to the fly. Due to its turning behavior, the fly is not perceiving the same motion perceived by the experimenter. The experimenter does not share the fly’s view of the world, because he is not rotating with the cylinder. He is not rotating, in part because his internal reference signal is different from that of the fly. As a result of his bias, Camhi calls the disturbance the input to the system. The input entering the fly’s nervous system is the actual motion sweeping across the fly’s retina, or slip speed. The slip speed is not an error signal in the controller, but the actual input to the optomotor control system. The rotation of the cylinder is the disturbance. The actual motion perception is the controlled variable. When the environment is not moving, the slip speed is close to zero. When the cylinder rotates, this slip speed starts to increase. If the fly stays still, rather than turning with the cylinder, the slip speed would reflect the disturbancedthe speed of rotation introduced by the experimenter. But the fly turns with the cylinder, reducing the slip speed. Thus the slip speed appears to be related to the actual controlled variable. The cylinder rotation introduces a visual input that deviates from the value dictated by the internal reference of the fly, which indicates how much slip speed is to be reached or tolerated. The comparator is located somewhere in the fly brain; and it compares the perceived visual motion with the reference signal. Given the behavior, we can assume that the fly has an internal reference signal close to zero, which means that any perceived slip speed will create an error signal. The output generated is turning in the same direction as the cylinder, which reduces the value of the controlled variable to the level specified by the reference. For this optomotor control system, the slip speed is not the error signal created by the control system, but is proportional to the uncorrected portion of the effects of the disturbance. Camhi writes: “one surprising feature of the optomotor feedback loop is that it is actually impossible for the fly to keep up perfectly with the cylinder’s rotation; for to do so would. create a slip speed equal to zero. But a slip speed

The crisis in neuroscience Chapter | 3

39

of zero would produce zero behavior, so that fly’s turning would, of necessity, stop. Given this circumstance, the best that the fly can do is to keep its turning speed very close, but not equal, to the cylinder’s rotation speed.” Here Camhi assumes that feedback control requires the animal to generate sufficient error in order to generate behavior, but he ignores the fact that negative feedback systems, though relying on error to generate output, are also error reducing. Nor is he aware of integrators used in controllers to generate output as a function of error in the recent past, so that zero error does not necessarily mean zero output. Camhi’s confusion is exposed by his analysis of the control system. To calculate gain, he simply divides the output by input. But the output is the measured behavior, whereas the input used is the disturbance or cylinder motion, which is not input to the organism, but “input” from the perspective of the experimenter. Here is an example of how the traditional view of control systems introduced by engineers leads to a major conceptual confusion. To add to the confusion, Camhi also makes a distinction between the open loop gain and closed loop gain. To calculate “closed loop gain,” Camhi divides the turning speed of the fly (output) by the speed of the cylinder rotation (input). The ideal gain is 1, so that the output velocity matches the input velocity perfectly. But this calculation reflects the engineer’s bias discussed earlier, as the actual input to the organism is the slip speed rather than the cylinder rotation as perceived by the experimenter. To calculate the “open loop gain”, the fly is fixed to a rod to prevent its body from turning, abolishing the feedback function, but its movements are measured by a circular card held by its feet. The open loop gain tells us how much output is generated by a given amount of error. Using these measures, Camhi calculates that the closed loop gain measured typically ranges from 0.75 to 0.95, whereas the open loop gain ranges from 6 to 8, concluding that the open loop gain is much higher than the closed loop gain. But the actual input is not the cylinder rotation, which only becomes the input when the feedback is cut by preventing the fly’s turning. For a cylinder rotation speed of 10 /s, Camhi’s “closed loop gain” is w0.9, i.e. the turning response is about 9 /s, and the slip speed or actual input would be w1 /s. If the reference is 0 /s, then the error is w1 /s. This suggests that the actual loop gain, at least the portion due to the output function, is close to 9. The output gain is usually the major contributor to the loop gain, capable of amplifying a very small amount of error into the appropriate corrective output. This can only be measured when the loop is opened, since in a closed loop the error is self-reducing. Contrary to intuition, a high loop gain does not mean that a large response will be produced, so long as there is negative feedback. A high gain allows the system to oppose the effects of disturbance very quickly. The higher the gain, the more closely the output will mirror the disturbance (rather than the actual input or slip speed in Camhi’s example). If the gain is 90, then roughly 99% of the disturbance effects are canceled. With sufficient loop

40 SECTION | A Why do we need perceptual control theory?

gain, the tracking will be virtually perfect. With a loop gain of 9, roughly 90% of the effect of the disturbance on the controlled input variable is canceled by the turning behavior. To be sure, this is still a low gain, but Camhi’s estimates are based on Mittelstaedt’s measurements from input/output experiments. Neither Camhi nor Mittelstaedt had any understanding of control systems when they performed the measurements. Assuming the controlled variable is perceived motion or cylinder velocity, now the error is 10 /s instead of 1 /s due to behavioral feedback that cancels 90% of the effect of disturbance. Camhi’s calculated open loop gain is 6e8; so with a disturbance 10 /s the output might be 60e80 /s. That the output is much higher than the disturbance does not suggest a higher gain in the system. In the open loop condition, the effect of disturbance cannot be rejected through the feedback function. There appears to be an integrator in the output function, so that the error signal simply accumulates. Whether or not the feedback function is eliminated, the same system properties are being measured. A similar problem is found in the study of oculomotor behavior in monkeys (Robinson, 1981, 1990; Lisberger, Morris, & Tychsen, 1987). Unlike flies and other lower organisms, monkeys are capable of relatively isolated eye movements, especially when their heads are fixed, as in most experiments on eye movements. It is then possible to elicit movements by manipulating visual target stimuli. Traditionally this is considered an example of the sensorimotor transformation that converts visual input (e.g. target motion) into the appropriate eye movement. Given a highly reliable relationship between visual target motion and eye movement, a massive literature has been produced on the neural substrates of the oculomotor system in monkeys. Yet surprisingly there is no working model of the oculomotor circuit for eye movement. Despite thousands of studies on the superior colliculus, a structure critical for eye movements, there is still no consensus on what the neural signals in the superior colliculus represent (Wurtz & Goldberg, 1971; Sparks, 2002). Many different terms have been used to describe the neural signals, such as spatial attention, motor memory, response selection, response preparation, motor set, and target selection (Schall, 2004). Oculomotor behavior in monkeys is normally studied using pursuit or saccade tasks. Pursuit movement generally tracks slow motions of some visual target stimulus whereas saccade movement is a large and rapid eye movement in response to a large movement of the target (Robinson, 1968). The target velocity is usually treated as the input, and the eye movement as the output (Lisberger, Evinger, Johanson, & Fuchs, 1981). The gain is calculated in the conventional way: eye movement velocity divided by target velocity. Again, the disturbance is treated as the input to the system. The oculomotor field is thus misled by the incorrect application of control theory by some of its pioneers (Robinson, 1981; Steinman, 1986). They start with the wrong premise that the comparison function is outside the organism, and that the role of the oculomotor system is to transform the error signal into

The crisis in neuroscience Chapter | 3

41

the appropriate output. The input signal is called the retinal error, the result of a comparison between output and disturbance. But as we have seen, there is no such thing as retina error. The visual input at the retina is simply the perceptual signal. There are two major sources of influence on this perceptual signal: the feedback effect produced by behavior and the disturbance effect produced by the environment. For the sake of simplicity, we can assume that the behavior is just the eye movement, and the disturbance comes from the target motion. These two influences act in opposite directions if there is an effective control system present. This control system contains an internal reference signal, adjustable by top down signals. It acts to make the actual retinal input match the reference. Yet the mistake of treating the oculomotor system as an input/output device leads to extraordinarily tortuous reasoning about the internal organization of the system. It may appear that the system is well calibrated, with a highly linear causal link from target motion to eye movement. But where is the input that leads to the reliable eye movement? According to the conventional answer, the input is the motion of the visual target. As already mentioned earlier, that is only true from the perspective of the observer. It is not true from the perspective of the subject, whose eye movement manages to reduce the actual motion of the visual target on the retina. As a result, the retina does not ever get to receive the input that is received by the observer, because the purpose of the system is to reduce the visual motion in the first place. The eye movement stabilizes the visual image, but the input-output link cannot function when the input is stabilized by the behavior itself. For ‘linear causation’ system to work, it must obtain the “actual velocity” from the observer’s perspective, not the actual image velocity on the retina. But instead of concluding that this is an example of reductio ad absurdum, and that the oculomotor system cannot be some input/output device, Robinson and other pioneers in the field decided to accept the absurd. They reasoned that, indeed, what was needed was an internal reconstruction of the visual target motion. Since this information is not directly available to the eye, the brain must somehow reconstruct the target velocity signal, by adding a copy of the pursuit command (eye velocity with respect to the head) to the retinal velocity error (target velocity with respect to the eye). In other words, self-movement plus the perceived motion are combined to generate a representation of the disturbance. To recreate the right stimulus, internal positive feedback of the motor command for eye velocity was used (Young, Forster, & Van Houtte, 1968; Lisberger et al., 1987). To obtain the “right” input, some copy of the system output is actually needed, but how does the system’s own output reflect the movement velocity in the first place? Starting from the wrong premise, this train of reasoning results in infinite regress. One consequence of this line of reasoning is that it attributes all variability in movement to the “information” sent to the motor system. Thus, exactly what type of behavior is produced is determined by what the sensory system

42 SECTION | A Why do we need perceptual control theory?

tells the motor system. Variability in output must be attributed to variability in input, or noise in the sensory system. If only the visual system can provide the same signal every time! If only our perceptual systems can always provide the motor system with the right instructions! Then performance would not vary so much. But the fact that all performance is always variable is brushed aside, and no attempt is made to explain why the system output is necessarily variable. In negative feedback control systems, it is not necessary to reconstruct target velocity at all. The thermostat, for example, does not reconstruct the actual outside temperature. Its main function is to keep the reading of its sensors at a particular level, a level that is determined by the internal reference, and independent of the external environment. The controlled variable in visual tracking systems can be displacement of the target from the fovea or the movement of the visual field. The retina cannot report the error between eye and target. That error is generated at higher levels inside the brain, using internal comparison functions that generate the discrepancy between the reference and input. The difference between the target representation and the fovea is the error that must be converted into the output. The function of the orientation controller reduces the error between fovea and target. What Robinson calls positive feedback appears to be integration of error signals in the output function of a negative feedback controller. Again, what is overlooked is the internal reference that specifies how much perceived motion deviation of the target is tolerated. Object motion constitutes a disturbance because it forces the perceptual inputs to deviate from the values specified by the internal reference signal. Consequently, any detected motion can generate an error signal that results in movement of the eyes and body. But a change in reference will change how the system behaves given the same target motion. Just like Mittelstaedt and Camhi, the students of oculomotor behavior were puzzled by the behavior of a simple control system, because they used the wrong methods to analyze it. The incorrect analysis can ultimately be traced back to Wiener’s diagram and conventional engineering control theory. Tragically, it was precisely their training in engineering control theory that led to their confident rejection of the simple control model as the explanation for oculomotor behavior.

The proper study of behavior It is beyond the scope here to present a model of a viable control hierarchy, and to explain the neural implementation of such a model. That is a task for the future, though initial steps in that direction have been taken (Yin, 2014). The aim here is only to explain the origin of the current crisis in neuroscience and the need for a paradigm shift. But having described at length the problems with the linear causation paradigm, it would be fitting to suggest some solutions for resolving the current crisis. Below I shall briefly outline how

The crisis in neuroscience Chapter | 3

43

control theory, properly understood, can inform studies of neural function and establish a new paradigm in neuroscience. First, it is necessary to avoid nature’s trap for the students of behavior, namely the behavioral illusion. Although neuroscientists today do not usually use decerebrate preparations, they still force their subjects to behave like input-output systems, by eliciting behavioral output with a disturbance (stimulus). Traditional measures of behavior show a series of pulse-like events, and neglect everything in between, as if the animal is waiting for the next trigger to behave again. Such an animal is considered the ideal subjectdan assumption implicit in the common practice to heavily restrain animals so that they will show only one discrete behavior that the experimenter wishes to study. When behavior is thus conveniently truncated and labeled, it is all too easy to make up arbitrary task variables from the perspective of the experimenter, and examine the neural correlates of only these variables while neglecting everything else. Imposing the observer’s perspective and linking arbitrary imagined variables to the neural activity are perhaps the most common errors in neuroscience. The truncated input/output designs in traditional neuroscience experiments make it easy to avoid any inconvenient and continuous data. As we have seen with the oculomotor field, as soon as one attempts to be quantitative, insurmountable problems arise. For input/output devices have very stringent requirements that are never met in biological organisms, unless they are dead. If the retinal image is stabilized by the action of a control system, then that image itself can no longer “cause” the output appropriately. The experimenter is then forced to invent all kinds of ad hoc mechanisms to restore the sacred input, as we have seen in the discussion of oculomotor behavior. We must therefore abandon the assumption of the input/output device altogether, and measure behavior continuously. The discrete categories of events (no movement, movement, cue, reward) commonly used in behavioral analysis are largely imposed by the researcher. While they may be convenient labels, these labels make it easy to neglect how behavior is changing over time, which should be the subject under study in the first place. In fact, categorical perceptions that approximate digital or logical processing only emerge at the highest levels of the control hierarchy. They cannot be understood without understanding the lower level continuous processes that give rise to such all-or-none perceptions. To understand the neural substrates underlying behavior, then, it would be necessary to view it, and to measure it, as a continuous process. It would be necessary to allow the organism under study to behave freely, and to analyze the relationship between neural activity and continuous and timevarying behavioral variables. When behavior is thus measured, the failure of traditional models of behavior will become abundantly clear. The remarkable complexity and variability in the measured behavior is anticipated by control theory, as the output of a negative feedback control system will always vary

44 SECTION | A Why do we need perceptual control theory?

according to fluctuations in disturbances to the controlled variable. For this variability to be comprehensible, we must first understand the controlled variable. Research based on control theory, then, involves three steps: (1) formulating a hypothesis about the controlled variable; (2) testing for the controlled variable; and (3) modeling of the control system. First, when observing any behavior, we begin with the question of which perceptual inputs to the organism are altered by the behavioral output. From possible answers to this question we can formulate a hypothesis about what the controlled variable is. Second, we must test this hypothesis, by generating a disturbance to the hypothesized controlled variable (Marken, 1997). It is possible to estimate how much effect this disturbance will have on the hypothetical controlled variable in the absence of control, i.e. under open loop conditions. Disturbances that alter that variable along will be opposed. If the predicted change fails to occur, or is smaller than predicted, then the hypothetical variable is not being controlled, and one must go back to the drawing board and propose another hypothesis. As analyzed above, in traditional research on reflexes and oculomotor behavior, disturbances are also used to produce behavioral output. What then is the difference between the traditional approach and the approach suggested here? The main difference is the identification of the controlled variable. In the pursuit task, we assume that the internal reference specifies no change in position of the target, because any target motion will generate corresponding pursuit movement. Instead, it is possible to manipulate the feedback functiondthe effect of the output on the visual target, and assess the compensatory increase in the output in order to control input, or to use such a method to clamp the error signal entering the integrator in the controller and measure the properties of the integrator. Given the wide range of possible variables that can be controlled, it may seem nearly impossible to select the correct one. But we are greatly aided in this by the fact that we ourselves are a collection of control systems. Introspectionddisciplined, informed by theory, and always subject to empirical testingdcan be a powerful tool in understanding living control systems. Because we perceive and have control over what we perceive, we can hypothesize what others perceive and control. This approach would be hopelessly circular but for the empirical testing that must always follow the hypothesis formulation. Thus in the analysis of the control system, the formulation of a hypothesis is based on first person perception and imagination, and the verification or falsification process is based on third person observations. Having identified the controlled variable, we can now proceed to measure the properties of the system. For neuroscientists, this would involve identifying the neural implementation of the control loop components at each level of the hierarchy (input, output, comparator). The challenge is to distinguish between

The crisis in neuroscience Chapter | 3

45

neural signals representing perceptual, reference, or error signals. The anatomical organization can provide clues. Just as important is the pattern of disturbance and the temporal relationship between the measured signals given a disturbance to the hypothetical closed loop. For example, perceptual signals can be stabilized as output and error vary according to the disturbance. Here the key advantage of neuroscience is that, unlike behavioral studies, the experimentalist can directly measure the signals inside the organism. Indeed, by studying the relationship between different signals in the control hierarchy, and testing specific hypotheses generated by working models, a neurobiological analysis is the only method that can give us detailed information on the functional organization of the hierarchy.

Conclusions Negative feedback control systems have properties that are counterintuitive, especially to those influenced by the linear causation paradigm. However shocking these properties may be, they can be easily demonstrated. Folk psychology is right to insist on the existence of goals and purposes, though it lacks any explanation of what they are and how they operate. The victory of Galileo and modern physics over the school of Aristotle does not invalidate teleology. The control loop explains what Aristotle calls the final cause, “that for the sake of which.” For example, a man can run in order to improve his health. If he is not healthy at the moment, how can he become healthy by running? Becoming healthy is a goal, and running is the means to that goal (telos, or end). The telos is similar to the internal reference signal of a control system. It is this signal that makes control possible. No behavior can make any sense without invoking some internal reference signal representing the “should be value” of the controlled variable. The emergent property of a negative feedback organization is teleology, in a literal sense, rather than linear causation. All control systems are teleological systems, even though there is no ‘ghost in the machine,’ and no violation of physical law. The attempt to remove teleology from the study of behavior, therefore, will be remembered as one of the greatest scientific blunders in history. When properly understood, control theory contradicts the prevailing opinions of the learned and questions the most ingrained paradigm in neuroscience, the paradigm of linear causation. Not only does it refute the idea that external inputs or environmental influences cause behavior, but also the idea that thoughts and plans and internal models cause behavior (James, 1890; Seligman, Railton, Baumeister, & Sripada, 2013). It suggests that the input/ output approach to experimental design will necessarily fail, and that the General Linear Model and statistics have been inappropriately applied to the study of living organisms (Friston, Harrison, & Penny, 2003). Grasping these remarkable implications is the first step toward a resolution of the current crisis in neuroscience.

46 SECTION | A Why do we need perceptual control theory?

References Adrian, E. D. (1957). The analysis of the nervous system. Proceedings of the Royal Society of Medicine, 50(12), 991. Alivisatos, A. P., Chun, M., Church, G. M., Greenspan, R. J., Roukes, M. L., & Yuste, R. (2012). The brain activity map project and the challenge of functional connectomics. Neuron, 74(6), 970e974. Ashby, W. R. (1956). An introduction to cybernetics. John Wiley and Sons. Ashby, W. R. (1958). Requisite variety and its implications for the control of complex systems. Cybernetica, 1(2), 83e99. Bernstein, N. (1967). The coordination and regulation of movements. Oxford: Pergamon Press. Britten, K. H., Shadlen, M. N., Newsome, W. T., & Movshon, J. A. (1992). The analysis of visual motion: a comparison of neuronal and psychophysical performance. Journal of Neuroscience, 12(12), 4745e4765. Brunswik, E. (1943). Organismic achievement and environmental probability. Psychological Review, 50(3), 255. Camhi, J. M. (1984). Neuroethology. New York: Sinauer. Craik, K. J. (1947). Theory of the human operator in control systems1. British Journal of Psychology. General Section, 38(2), 56e61. Friston, K. J., Harrison, L., & Penny, W. (2003). Dynamic causal modelling. NeuroImage, 19(4), 1273e1302. Gallistel, C. R. (1980). The organization of action: A new synthesis. Hillsdale, N. J.: Lawrence Erlbaum Associates. Glimcher, P. W. (2005). Indeterminacy in brain and behavior. Annual Review of Psychology, 56, 25e56. Gomez-Marin, A., Paton, J. J., Kampff, A. R., Costa, R. M., & Mainen, Z. F. (2014). Big behavioral data: psychology, ethology and the foundations of neuroscience. Nature Neuroscience, 17(11), 1455e1462. Green, A. M., & Angelaki, D. E. (2010). Internal models and neural computation in the vestibular system. Experimental Brain Research, 200(3e4), 197e222. Hodgkin, A. L., & Huxley, A. F. (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of Physiology, 117(4), 500. Hull, C. L. (1934). The concept of the habit-family hierarchy, and maze learning. Part I. Psychological Review, 41(1), 33. James, W. (1890). The principles of psychology (vol. 1). New York: Henry Holt. Kuhn, T. S. (1962). The structure of scientific revolutions. Chicago: University of Chicago press. Lisberger, S., Evinger, C., Johanson, G., & Fuchs, A. (1981). Relationship between eye acceleration and retinal image velocity during foveal smooth pursuit in man and monkey. Journal of Neurophysiology, 46(2), 229e249. Lisberger, S. G., Morris, E., & Tychsen, L. (1987). Visual motion processing and sensory-motor integration for smooth pursuit eye movements. Annual Review of Neuroscience, 10(1), 97e129. Marken, R. S. (1997). The dancer and the dance: methods in the study of living control systems. Psychological Methods, 2(4), 436. Markram, H. (2006). The blue brain project. Nature Reviews Neuroscience, 7(2), 153e160. McIntyre, J., & Bizzi, E. (1993). Servo hypotheses for the biological control of movement. Journal of Motor Behavior, 25(3), 193e202.

The crisis in neuroscience Chapter | 3

47

McIntyre, J., Zago, M., Berthoz, A., & Lacquaniti, F. (2001). Does the brain model Newton’s laws? Nature Neuroscience, 4(7), 693e694. Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167e202. Mittelstaedt, H. (1964). Basic control patterns of orientational homeostasis. Paper presented at the symposia of the society for experimental biology, 18, 365e385. Neuringer, A. (2004). Reinforced variability in animals and people: implications for adaptive action. American Psychologist, 59(9), 891e906. Osborne, L. C., Lisberger, S. G., & Bialek, W. (2005). A sensory source for motor variation. Nature, 437(7057), 412e416. Penrose, R. (1999). The emperor’s new mind: concerning computers, minds, and the laws of physics. Oxford University Press. Pouget, A., & Snyder, L. H. (2000). Computational approaches to sensorimotor transformations. Nature Neuroscience, 3, 1192e1198. Powers, W. T. (1973a). Behavior: control of perception. New Canaan: Benchmark Publications. Powers, W. T. (1973b). Feedback: beyond behaviorism. Science, 179(71), 351e356. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960). A general feedback theory of human behavior. Perceptual & Motor Skills, 11, 71e88. Robinson, D. (1981). The use of control systems analysis in the neurophysiology of eye movements. Annual Review of Neuroscience, 4(1), 463e503. Robinson, D. (1990). Why visuomotor systems don’t like negative feedback and how they avoid it. In Paper presented at the vision, brain, and cooperative computation. Robinson, D. A. (1968). Eye movement control in primates. Science, 161(3847), 1219e1224. Rosenblueth, A., Wiener, N., & Bigelow, J. (1943). Behavior, purpose, and teleology. Philosophy of Science, 10(1), 18e24. Schall, J. D. (2004). On building a bridge between brain and behavior. Annual Review of Psychology, 55, 23e50. Seligman, M. E., Railton, P., Baumeister, R. F., & Sripada, C. (2013). Navigating into the future or driven by the past. Perspectives on Psychological Science, 8(2), 119e141. Shadmehr, R., & Wise, S. P. (2005). The computational neurobiology of reaching and pointing: A foundation for motor learning. MIT press. Sherrington, C. S. (1906). The integrative action of the nervous system. New Haven: Yale University Press. Skinner, B. F. (1953). Science and human behavior. Simon and Schuster. Sparks, D. L. (2002). The brainstem control of saccadic eye movements. Nature Reviews Neuroscience, 3(12), 952e964. Steinman, R. M. (1986). The need for an eclectic, rather than systems, approach to the study of the primate oculomotor system. Vision Research, 26(1), 101e112. Stevens, S. S. (1975). Psychophysics. Transaction Publishers. Todorov, E., & Jordan, M. I. (2002). Optimal feedback control as a theory of motor coordination. Nature Neuroscience, 5(11), 1226e1235. von Holst, E., & Mittelstaedt, H. (1950). The reafference principle. In The collected papers of erich von Holst. Coral Gables: University of Miami Press. Wiener, N. (1948). Cybernetics. Paris: Hermann & Cie Editeurs. Wolpert, D. M., & Kawato, M. (1998). Multiple paired forward and inverse models for motor control. Neural Networks, 11(7), 1317e1329.

48 SECTION | A Why do we need perceptual control theory? Wurtz, R. H., & Goldberg, M. E. (1971). Superior colliculus cell responses related to eye movements in awake monkeys. Science, 171(3966), 82e84. Yin, H. H. (2013). Restoring purpose in behavior. In Computational and robotic models of the hierarchical organization of behavior (pp. 319e347). Berlin: Springer. Yin, H. H. (2014). How basal ganglia outputs generate behavior (p. 768313). Advances in neuroscience, 2014. Yin, H. H. (2016). The basal ganglia and hierarchical control in voluntary behavior. In J.J. Soghomonian (Ed.), The basal ganglia-novel perspectives on motor and cognitive functions. Berlin: Springer (Vol. (in press). Young, L., Forster, J., & Van Houtte, N. (1968). A revised stochastic sampled data model for eye tracking movements. In Paper presented at the fourth annual NASA-university conference on manual control.

Chapter 4

When causation does not imply correlation: robust violations of the faithfulness axiom Richard Kennaway University of East Anglia, School of Computing Sciences, Norwich, United Kingdom

Introduction The problem of deducing causal information from correlations in observational data is a substantial research area, the simple maxim that “correlation does not imply causation” having been superseded by methods such as those set out in earlier work (Pearl, 2000; Spirtes, Glymour, & Scheines, 2001), and in shorter form (Pearl, 2009). I shall exhibit a substantial class of systems to which these methods and recently developed extensions of them fail to apply. This class is a type of system that occurs throughout the life and social sciences, and in engineering: control systems. The works just cited limit attention to systems whose causal connections form a directed acyclic graph, together with certain further restrictions, and also do not consider dynamical systems or time series data. As such, dynamical systems with feedback lie outside their scope. Attempts have been made to extend these methods toward dynamical systems and systems with cyclic causation, by relaxing one or more of the basic assumptions. However, among the class of dynamical systems with feedback is a subclass of ubiquitous occurrence in the real world, for which, I argue, no such extension of these methods can succeed. These are control systems: systems which have been designed, whether by a human designer or by evolution, to destroy the correlations that these causal inference methods work from. In addition, they tend to create high correlations between variables that are only indirectly causally linked. In a later section I discuss in detail some of the papers that have attempted such extensions. It is not an accident that in every case the assumptions made, while allowing some The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00004-6 Copyright © 2020 Elsevier Inc. All rights reserved.

49

50 SECTION | A Why do we need perceptual control theory?

dynamical systems with feedback, exclude control systems. I show where control systems violate their assumptions, and analyze the result of applying their methods anyway, exhibiting where they break down.

Preliminaries Causal inference I briefly summarise the concepts of causal inference set out in the works cited above. For full technical definitions the reader should consult the original sources. A hypothesis about the causal relationships existing among a set of variables V is assumed to be expressible as a directed acyclic graph G (the DAG assumption). An arrow from x to y means that there is a direct causal influence of x on y, and its absence, that there is none. Given such a graph, and a joint probability distribution P over V, we can ask whether this distribution is consistent with the causal relationships being exactly as stated by G: could this distribution arise from these causal relationships? Besides the DAG assumption, there are two further axioms that are generally required for P to be considered consistent with G: the Markov condition and the Faithfulness property. P satisfies the Markov condition if it factorises as the product of the conditional distributions PðVi jpredðVi ÞÞ, where predðVi Þ is the set of immediate predecessors of Vi in G. This amounts to the assumption that all of the other, unknown influences on each Vi are independent of each other; otherwise put, it is the assumption that G contains all the variables responsible for all of the causal connections that exist among the variables. It can be summed up as the slogan “no correlation without causation”. The Faithfulness assumption is that no conditional correlation among the variables is zero unless it is necessarily so given the Markov property. For example, if G is a graph of just two nodes x and y with an arrow from x to y, then every probability distribution over x and y has the Markov property, but only those yielding a non-zero correlation between x and y are faithful. It is not obvious in general which of the many conditional correlations for a given graph G must be zero, but a syntactically checkable condition was given by (Pearl, 1998), called d-separation. Its definition will not concern us here. Faithfulness can be summed up as the slogan “no causation without correlation”. The idea behind Faithfulness is that if there are multiple causal connections between x and y, then while it is possible that the causal effects might happen to exactly cancel out, leaving no correlation between x and y, this is very unlikely to happen. Technically, if the distributions P are drawn from some reasonable measure space of possible distributions, then the subset of nonfaithful distributions has measure zero.

When causation does not imply correlation Chapter | 4

51

When these assumptions are satisfied, the correlations present in observational data can be used to narrow the set of causal graphs that are consistent with the data. The assumptions have all been the subject of debate, but I am primarily concerned here with the Faithfulness assumption. Attacks on it have been based on the argument that very low correlations may be experimentally indistinguishable from zero, and therefore that one may conclude from a set of data that no causal connection can exist even when there is one. But, it can be countered, that merely reflects on the inadequate statistical power of one’s data set, the response to which should be to collect more data rather than question this axiom. However, I shall not be concerned with this argument. Instead, our purpose is to exhibit a large class of robust counterexamples to Faithfulness: systems which contain zero correlations that do not become nonzero by any small variation of their parameters, nor by the collection of more data, yet are not implied by the Markov property. Some of these systems even exhibit large correlations (absolute value above 0.95) between variables that have no direct causal connection, but are only connected by a series of direct links, each of which is associated with correlations indistinguishable from zero. These systems are neither exotic, nor artificially contrived for the sole purpose of being counterexamples. On the contrary, systems of this form are common in both living organisms and man-made systems. It follows that for these systems, this general method of causal analysis of nonexperimental data cannot be applied, however the basic assumptions are weakened. Interventional experiments are capable of obtaining information about the true causal relationships, but for some of these systems it is paradoxically the lack of correlation between an experimentally imposed value for x and the observed value of y that will suggest the presence of a causal connection between them.

Zero correlation between a variable and its derivative As a preliminary to the main results, I consider the statistical relation between a function, stochastic or deterministic, and its first derivative. In the Appendix I demonstrate that under certain mild boundedness conditions, the correlation between a differentiable real function and its first derivative is zero. (The obvious counterexample of ex, identical to its first derivative, violates these conditions.) An example of a physical system with two variables, both bounded, one being the derivative of the other is that of a voltage source connected across a capacitor. The current I is related to the voltage V by I ¼ C dV=dt, C being the capacitance. If V is the output of a laboratory power supply, its magnitude continuously variable by turning a dial, then whatever the word “causation” means, it would be perverse to say that the voltage across the capacitor does not cause the current through it. Within the limits of what the power supply

52 SECTION | A Why do we need perceptual control theory?

can generate and the capacitor can withstand, I can be caused to follow any smooth trajectory by suitably and smoothly varying V. The voltage is differentiable, so by Theorem 4.1 of the Appendix, on any finite interval in which the final voltage is the same as the initial, cV;I is zero. By Theorem 4.2, the boundedness of the voltage implies that the same is true in the limit of infinite time. This is not a merely fortuitous canceling out of multiple causal connections. There is a single causal connection, the physical mechanism of a capacitor. The mechanism deterministically relates the current and the voltage. (The voltage itself may be generated by a stochastic process.) Despite this strong physical connection, the correlation between the variables is zero. Some laboratory power supplies can be set to generate a constant current instead of a constant voltage. When a constant current is applied to a capacitor, the mathematical relation between voltage and current is the same as before, but the causal connection is reversed: the current now causes the voltage. Within the limits of the apparatus, any smooth trajectory of voltage can be produced by suitably varying the current. It can be argued that the reason for this paradox is that the product-moment correlation is too insensitive a tool to detect the causal connection. For example, if the voltage is drawn from a signal generator set to produce a sine wave, a plot of voltage against current will trace a circle or an axis-aligned ellipse. One can immediately see from such a plot that there is a tight connection between the variables, but one which is invisible to the productmoment correlation. A more general measure, such as mutual information, would reveal the connection. However, let us suppose that V is not generated by any periodic source, but varies randomly and smoothly, with a waveform such as that of Fig. 4.1A. This waveform has been designed to have an autocorrelation time of 1 unit: the correlation between V(t) and V(ted) is zero whenever jdj  1. (It is generated as the convolution of white noise with an infinitely differentiable function which is zero outside a unit interval.1) Choosing the capacitance C, which is merely a scaling factor, such that V and I have the same standard deviation, the resulting current is shown in Fig. 4.1B. A plot of voltage against current is shown in Fig. 4.1C. One can clearly see trajectories, but it is not immediately obvious from the plot that there is a simple relation between voltage and current and that no other unobserved variables are involved. If we then sample the system with a time interval longer than the autocorrelation time of the voltage source, then the result is the scatterplot of Fig. 4.1D. The points are connected in sequence, but each step is a random jump whose destination is independent of its source. Over a longer time, this sampling produces the

1. Matlab was used to perform all the simulations and plot the graphs. The source code is available as supplementary material.

When causation does not imply correlation Chapter | 4

53

FIG. 4.1 Voltage and current related by I ¼ dV=dt. (A) Voltage versus time. (B) Current versus time. (C) Voltage versus current. (D) Voltage versus current, sampled. (E) Voltage versus current, sampled for a longer time.

scatterplot of Fig. 4.1E. All mutual information between V and I has now been lost: all of the variables Vi and Ii are close to being independently identically distributed. Knowing the exact values of all but one of these variables gives an amount of information about the remaining one that tends to zero as the sampling time step increases. All measures of any sort of mutual information or causality between them tend to zero, not merely the correlation coefficient. The only way to discover the relationship between V and I is to measure them on timescales short enough to reveal the short-term trajectories instead of the Gaussian cloud.

Control systems A control system, most generally described, is any device which is able to maintain some measurable property of its environment at or close to some set value, regardless of other influences on that variable that would otherwise tend to change its value. That is a little too general: a nail may serve very well to prevent something from moving, despite the forces applied to it, but it would not be considered a control system. Control systems, more usefully demarcated, draw on some energy source to actively maintain the controlled variable at its reference value. Some everyday examples are a room thermostat that

54 SECTION | A Why do we need perceptual control theory?

FIG. 4.2 The basic block diagram of any feedback control system. The controller is above the shaded line; its environment (the plant that it controls) is below the line.

turns heating or cooling mechanisms up and down to maintain the interior at a constant temperature despite variations in external weather, a cruise control maintaining a car at a constant speed despite winds and gradients, and the physiological processes that maintain near-constant deep body temperature in mammals. The general form of a feedback controller is shown in Fig. 4.2. The variables have the following meanings: P: The controller’s perceptual input. This is a property of the environment, which it is the controller’s purpose to hold equal to the reference signal. R: The reference signal. This is the value that the control system tries to keep P equal to. It is shown as a part of the controller. In an industrial setting it might be a dial set by an operator, or it could be the output of another control system. In a living organism, R will be somewhere inside the organism and may be difficult to discover. O: The output signal of the controller. This is some function of the perception and the reference (and possibly their past history). This is the action the control system takes to maintain P equal to R. Often, and in all the examples of the present paper, O depends only on the difference ReP, also called the error signal. D: The disturbance: all of the influences on P besides O. P is some function G of the output and the disturbance (and possibly their past history). I shall now give some very simple didactic examples of control systems, and exhibit the patterns of correlations they yield among P, R, O, and D under various circumstances.

Example 1 Fig. 4.3 illustrates a simple control system acting within a simple environment, defined by the following equations, all of the variables being time-dependent. O ¼ kðR  PÞ

(4.1)

When causation does not imply correlation Chapter | 4

55

FIG. 4.3 Block diagram of a simple feedback control system.

P¼O þ D

(4.2)

Eq. (4.1) describes an integrating controller, i.e. one whose output signal O is proportional to the integral of the error signal ReP. Eq. (4.2) describes the environment of the controller, which determines the effect that its output action and the disturbing variable D have upon the controlled variable P. In this case O and D add together to produce P. Fig. 4.4 illustrates the response to step and

FIG. 4.4 Responses of the controller. (A) Step change in D, R ¼ 0. (B) Step change in R, D ¼ 0. (C) R ¼ 0, randomly varying (d). (D) R and D both randomly varying.

56 SECTION | A Why do we need perceptual control theory?

random changes in the reference and disturbance. The random changes are smoothly varying with an autocorrelation time of 1 s. The gain k is 100. Observe that when R and D are constant, P converges to R and O to R e D. The settling time for step changes in R or D is of the order of 1/k ¼ 0.01. The physical connections between O, R, P, and D are as shown in Fig. 4.3: O is caused by R and P, P is caused by O and D, and there are no other causal connections. We now demonstrate that the correlations between the variables of this system bear no resemblance to the causal connections. We generate a smoothly randomly varying disturbance D, which varies on a timescale much longer than 1/k, with a standard deviation of 1 (in arbitrary units). The reference R is held constant at 0. Table 4.1 shows the correlations in a simulation run of 1000 s with a time step of 0.001 s. The performance of the controller can be measured by its disturbance rejection ratio, s(D)/s(ReP) ¼ 23.2. The numbers vary between runs only in the third decimal place. For this system, correlations are very high (close to 1) exactly where direct causal connection is absent, and close to zero where direct causal connection is present. There is a causal connection between D and O, but it proceeds only via P, with which neither variable is correlated. Here is a visual contrast between the causal links and the non-zero pairwise correlations.

It is difficult to calculate significance levels for these correlations, since the random waveforms that we generate for D (and in some of our simulations, also for R) are deliberately designed to be heavily bandwidth-limited. They have essentially no energy at wavelengths below about 0.2 s. Successive samples are therefore not independent, and formulas for the standard error of the correlation do not apply. Empirically, if the simulation is repeated many times, we find that the variance of the correlation between two independently generated waveforms like D is proportional to the support interval of the filter TABLE 4.1 Correlations for an integrating controller (Example 1). Cf. Fig. 4.4C. P D O 0.002 –0.999 P 0.043

When causation does not imply correlation Chapter | 4

57

that generates them from white noise. This implies that as one reduces the sampling time step below the point at which there is any new structure to discover, the variance of the correlation will not decrease. When the samples already contain almost all the information there is to find in the signal, increasing the sampling rate cannot yield new information. All of the simulations presented here use runs of 1 million time steps of 0.001 s, and a coherence time for each of the exogenous random time series of 1 s. For two such series, independently generated, the standard deviation of the correlation (estimated from 1000 runs) is 0.023, compared with 0.001 for the same quantity of independent samples of white noise. Therefore, for the examples reported here, any correlation whose magnitude is below about 0.05 must be judged statistically not significant.2 The correlations observed here between D and both O and P are thus indistinguishable from zero. These correlations are approximately summarised in Table 4.2. This behavior is quite different from that of a passive equilibrium system, such as a ball in a bowl (or something nailed down, which is a similar situation with a much higher spring constant). In the latter system, if we identify D with an external force applied to the ball, P with its position, and O with the component of gravitational force parallel to the surface of the bowl, we will find (assuming some amount of viscous friction, and a measurement timescale long enough for the system to always be in equilibrium), that O and P are both proportional to D. There will also be a steady-state error. This is not the case for the control system above, which has zero steady-state error. Given any constant value of D, O will approach a value proportional to D while P tends to zero with time.

TABLE 4.2 Rounded correlations (Example 1). O P

P 0

D –1 0

2. Significance in the everyday sense of practical usefulness deserves a mention. For the practical task of estimating the value of one variable from another it is correlated with, far higher correlations are required. For a bivariate Gaussian, even with a correlation of 0.5, the probability of guessing from the value of one variable just whether the other is above or below its mean is only 67%, compared with 50% for blind guessing. To be right 9 times out of 10 requires a correlation of 0.95. To estimate its value within half a decile 9 times out of 10 takes a correlation of 0.995. And to be sure from a finite sample that the correlation really is that high would require its true value to be even higher.

58 SECTION | A Why do we need perceptual control theory?

Real control systems often have to deal with some amount of transport lag in the environment, which we can model by changing Eq. (4.2) to P(t) ¼ O(tel) þ D(t), where l is the time delay. Transport lags are common in control systems in which the environment literally transports a substance O from where the controller produces it to where it affects P. Examples abound in chemical process engineering and in biological systems. This particular control system will only be stable in the presence of lag if its gain is below about 1/l. When this is so, the correlations and rejection ratio are little affected by the presence of lag. This remains true if correlations are calculated between lagged and unlagged quantities.

Example 2 If we modify Example 1 by letting R vary in the same manner as D, but independently from it, the correlations are now as shown in Tables 4.3 and 4.4. Here is a graphical comparison of the causal relationships and the correlations:

TABLE 4.3 Correlations for varying R and D (Example 2). P R E D O 0.718 0.717 –0.002 –0.718 P 0.998 –0.027 –0.031 R 0.039 –0.032 E –0.024

TABLE 4.4 Rounded correlations (Example 2). P R E D O 0.7 0.7 0 –0.7 0 1 0 P 0 R 0 0 E

When causation does not imply correlation Chapter | 4

59

O now has a substantial correlation with every variable except the only one that it directly depends on, E, with which its correlation is zero. In the limit of increasing gain, P and O are almost identical to R and R e D respectively. Since R and D are independently identically distributed, the correlation of O with R or D tends to 1/O2. All but one of the causal links (that from O to P) has a correlation of zero; all but one of the non-zero correlations corresponds to a causal link.

Example 3 In all of the systems discussed so far, there has been no noisedthat is, signals of which the experimenter knows nothing except possibly their general statistical characteristics. Some of the variables are randomly generated waveforms, but they are all measured precisely, with no exogenous noise variables. In the next example, I show that the introduction of modest amounts of noise can destroy some of the correlations among variables whose amplitudes are far larger than the noise. In Example 1, if we measure the correlation between O þ D and P, then it will of course be identically 1, and we might consider this correlation to be important. However, in practice, while the variables P, R, and O may be accurately measureable, D in general is not: it represents all the other influences on P of any sort whatever, known or unknown (The control system itselfdEq. 4.1ddoes not use the value of D at all. It senses only P and R, and controls P without knowing any of the influences on P.). To model our partial ignorance concerning D, I shall split it into D0, the disturbances that can be practically measured, and D1, the remainder. Let us assume that D0 and D1 are independently randomly distributed, and that the variance of D ¼ D0 þ D1 is ten times that of D1. That is, 90% of the variation of the disturbance is accounted for by the observed disturbances. The correlations that result in a typical simulation run, with randomly varying D and constant R are listed in Tables 4.5 and 4.6. When D1 has amplitude zero, the system is identical to the earlier one, for which O þ D0 ¼ P. But when the additional disturbance D1 is added,

TABLE 4.5 Correlations with noisy measurement (Example 3). P O + D0 D0 D1 D O 0.002 0.308 –0.947 –0.311 –0.999 0.132 0.042 0.011 0.043 P 0.012 –0.990 –0.302 O + D0 D0 –0.006 0.948 0.311 D1

60 SECTION | A Why do we need perceptual control theory?

TABLE 4.6 Rounded correlations (Example 3). P O + D0 D 0 D1 D 1 O weak 0 –1 –weak very weak 0 0 0 P –weak 0 –1 O + D0 0 1 D0 D1 weak

accounting for only one-10th the variation of D, the correlation between O þ D0 and P sinks to a low level. The reason is that the variations in P are much smaller than the noise. For this run, the standard deviations were s(O) ¼ 0.999, s(D0) ¼ 0.953, s(D1) ¼ 0.318, and s(P) ¼ 0.046. So although the unmeasurable O þ D0 þ D1 is identical to P, the measurable O þ D0 correlates only weakly with P, and the better the controller controls, the smaller the correlation.

A digression on disturbances Example 3 also demonstrates the error of a common naive idea about control systems. In every functioning control system, O is just the output required to oppose the effects of the disturbances D on P. It is sometimes assumed that a control system works by sensing D and calculating the value of O required to offset it. This does not work. It would result in every unsensed disturbance affecting P unopposed, and as we have seen, the controller in Example 3 performs far better than this. It would also require the control system to contain a detailed model of how the output and the disturbances affect Pdthat is, a model of its environment. Any inaccuracies in this model will also produce unopposed influences on P. For Example 4 below, control by calculating O from measurements of DO and DP would allow any error in measuring due to produce an error in P growing linearly with time, a complete failure of control. While controllers can be designed that do make some use of an environmental model and sensed disturbances to improve their performance, it remains the case that no control is possible without also sensing the actual variable to be controlled. Control cannot be any better than the accuracy of that measurement. Sensing or modeling anything else is not a necessity for control.

Example 4 A slightly different control system is illustrated in Fig. 4.5. This is very similar to the previous one, but now the output is proportional to the error and the integrator is part of the environment. Two disturbances are present, DO adding

When causation does not imply correlation Chapter | 4

61

FIG. 4.5 Another simple feedback control system.

to the output and DP adding to the perception. DP has standard deviation 1, and DO has a standard deviation chosen to produce the same magnitude of error signal as DP. These are the equations describing the system: O ¼ kðR  PÞ controller Z P ¼ DP þ ðO þ DO Þdt environment With a constant reference signal, the correlations we obtain for a sample run of this system are shown in Table 4.7, and in rounded form in Table 4.8. The general pattern of correlations is this: Again we see patterns of correlation that do not resemble the causal relationships:

O is proportional to the difference between P and R, and is not directly influenced by anything else, but is correlated with neither. DP has no correlation with any other variable. O correlates perfectly with E, because O ¼ kE

TABLE 4.7 Correlations for a proportional controller (Example 4). E DO DP P R O –0.041 0.034 1.000 –0.527 –0.031 P 0.997 –0.041 0.008 0.001 R 0.034 –0.032 –0.002 –0.527 –0.031 E DO –0.006

62 SECTION | A Why do we need perceptual control theory?

TABLE 4.8 Rounded correlations (Example 4).

P R E

O 0 0 P 1 R E DO

1 0 0

DO DP –0.5 0 0 0 0 0 –0.5 0 0

by definition, but the only other variable it has a correlation with is DO, whose causal influence on O proceeds via an almost complete circuit of the control loop. As with the integral control system, the addition of transport lag of no more than 1/k does not change this behavior. The apparently paradoxical relationships between causation and correlation in all these examples can in fact be used to discover the presence of control systems. Suppose that D1 is an observable and experimentally manipulable disturbance which one expects to affect a variable P (because one can see a physical mechanism whereby it should do so), but when D1 is experimentally varied, it is found that P remains constant, or varies far less than expected. This should immediately suggest the presence of a control system which is maintaining P at a reference level. In Powers’ earlier work (Powers, 1974, 1998) this is called the Test for the Controlled Variable. Something else must be happening to counteract the effect of D1, and if one finds such a variable O that varies in such a way as to do this, then one may hypothesise that O is the output variable of the control system. Further investigation would be required to discover the whole control system: the actual variable being perceived (which might not be what the experimenter is measuring as P, but something closely connected with it), the reference R (which is internal to the control system), the actual output (which might not be exactly what the experimenter is observing as O), and the mechanism that produces that output from the perception and the reference. For this test to be effective, the disturbance must be of a magnitude and speed that the control system can handle. Only by letting the control system control can one observe that it is doing so. A further paradox of control systems also appears here. Unlike the control systems of the earlier examples, this one has a certain amount of steady-state error. When R, DO, and DP are constant, the steady-state values of E and O are eDO/k and eDO respectively. When the gain k is large, E may be experimentally indistinguishable from zero, while O may be clearly observed. E is, however, the only immediate causal ancestor of O. For the systems of Examples 1e3, the paradox is even stronger: the steady-state error and output

When causation does not imply correlation Chapter | 4

63

signals are both zero, for any constant values of the disturbances and reference. Superficial examination of the system could lead one to wrongly conclude that they play no role in its functioning.

Example 5 In all of the above examples, the disturbances and the reference signal vary only on timescales much longer than the settling time of the control system: 1 s versus 0.01 s. If we repeat these simulations with signals varying on a much shorter timescale than the settling time, then we find completely different patterns of correlation. Table 4.9 shows the general form of the results for examples 1e4 with a gain of 1 (giving a settling time of 1 s) and a coherence time for the random signals of 0.1 s. Notice that Faithfulness is still violated. For example, the new version of Example 1 has a zero correlation between O and P, although the only causal arcs involving O are with P. For this system, O is just the integral of P, and so Theorem 4.2 applies: O and P necessarily have zero correlation regardless of the signal injected as D. For the new versions of Examples 2 and 3, O does not correlate with any other variable, and for Example 4, DO correlates with no other variable. When a variable correlates with nothing else, no system of causal inference from non-interventional data can reveal a causal connection that includes it.

TABLE 4.9 Correlations in the presence of fast disturbances (Example 5).

O

P

P 0

O P O + D0 D0 D1

D 0 1

Example 5(1)

O P R E

P 0

R 0 0

E 0 –0.7 0.7

Example 5(2)

D 0 1 0 –0.7

O P R E DO

P O + D0 0 0 1

P –0.7

D0 0 1 1

D1 0 weak 0 0

Example 5(3) R 0.7 0

E DO 1 0 –0.7 0 0.7 0 0

Example 5(4)

D 0 1 1 1 0

DP –0.7 0.7 0 –0.7 0

64 SECTION | A Why do we need perceptual control theory?

Summary Many more examples could be discussed of control systems of different architectures and different types of disturbances, but these are enough to illustrate the general pattern. The presence of control systems typically results in patterns of correlation among the observable variables that bear no resemblance to their causal relationships. These patterns are robust: neither varying the parameters of the control systems nor collecting longer runs of data would reveal stronger correlations or diminish the extreme correlations between variables with no direct connection. I shall later discuss why this is so, but first I examine some current work on causal inference and demonstrate in each case that the authors’ hypotheses exclude systems such as the above, and, if applied despite that, that their methods indeed fail to correctly discover the true causal relations.

Current causal discovery methods applied to these examples When a theorem shows that certain causal information can be obtained from nonexperimental observations on some class of systems, and yet no such information is obtainable from the systems we have exhibited, it follows that these systems must lie outside the scope of these theorems. Here I shall survey some such methods and exhibit where the systems considered here violate their hypotheses. Dynamical systems are excluded from all of the causal analyses of references (Pearl, 2000; Spirtes et al., 2001), which do not consider time dependency. In addition, control systems inherently include cyclic dependencies: the output affects the perception and the perception affects the output. Control systems therefore fall outside the scope of any method of causal analysis that excludes cycles, and both of these works restrict attention to directed acyclic causal graphs. Lacerda et al. (Lacerda, Spirtes, Ramsey, & Hoyer, 2008) consider dynamical systems sampled at intervals, and allows for cyclic dependencies, but a condition is imposed that in any equation giving xnþ1 (the value of x at time n þ 1) as a weighted sum of variables at time n, the coefficient of xn in that sum must be less than 1. This excludes any relation of the form x ¼ dy/dt, which in discrete time is approximated by the difference equation xnþ1 ¼ xn þ yn dt. In addition, they recommend sampling such systems on timescales longer than any transient effects. As can be seen from Fig. 4.1C and D,E, the organised trajectories visible when the system is sampled on a short time scale vanish at longer sampling intervals: only the transient effects reveal anything about the true relation between the variables. This recommendation thus rules out any possibility of discerning causal influences from nonexperimental data in the presence of control systems.

When causation does not imply correlation Chapter | 4

65

The problems that dynamical systems pose for causal analysis have been considered in terms of the concept of “equilibration”. An earlier work (Iwasaki & Simon, 1994) demonstrates how the apparent causal relationships in a dynamical system may depend on the timescale at which one views it, the timescale determining which of the feedback loops within the system have equilibrated. (Although he describes a variable which depends on its own derivative as “self-regulating”, the didactic example that he discusses, of a leaky bathtub, is not a control system.) Later works (Dash, 2003; Dash, 2005) consider the interaction of the Equilibration operator (Iwasaki & Simon, 1994) and the Do operator (Pearl, 2000), and considers the question of when they commute. They show that they often do not, and recommends that when this is so, manipulation (Do) be performed before equilibration. This amounts to recommending that the system be acted on and sampled on a timescale shorter than its equilibration time, so that transient behavior may be observed. Even when this is done, however, the true causal relationships may fail to manifest themselves in the correlations, as shown by the voltage/current example, and in Example 5, where the disturbance is ten times as fast as the controller’s response time. When intermediate timescales of disturbance are applied to the example control systems, some mixture of the equilibrium and non-equilibrium patterns of correlation will be seen. With more complicated systems of several control systems acting together at different timescales (as in the case of cascade control, where the output of one control system sets the reference of another, usually faster one), the patterns of correlations in the face of rapid disturbances will be merely confusing. A further moral to be drawn from this is that “equilibration” is not necessarily a passive process, like the ball-in-abowl described after Example 1, or the bathtub example (Dash, 2003, 2005; Iwasaki & Simon, 1994). Unlike those systems, control systems typically exhibit very small or zero steady-state error (which is almost their defining characteristic). Duvenaud and colleagues (Duvenaud, Eaton, Murphy, & Schmidt, 2008) are concerned with techniques for making successful predictions, rather than learning the correct causal structure. However, for some of our examples this is impossible even in principle. If a variable has zero correlation with any other variable, conditional on any set of variables, then no information about its value can be obtained from the other variables. This is the case, for example, with Example 5(1) and the variable O, and for the voltage/current example. A concurrent article (Itani, Ohannesian, Sachs, Nolan, & Dahleh, 2008) proposed a method of causal analysis capable of deriving cyclic models. It first generalises causal graphs to the cyclic case. Whereas in the acyclic case one can demonstrate that the conditional probability distribution on the whole graph factorises into conditional distributions of each variable given its immediate causes, this is not always so for cyclic graphs. It therefore considers only graphs for which a generalisation of this holds. Given the

66 SECTION | A Why do we need perceptual control theory?

conditional distribution of each variable given its immediate causal predecessors, a joint distribution of all the variables is said to be induced by them if (i) a local version of factorisation holds, and (ii) nodes are independent under d-separation. (We refer to that paper for the complete technical definition.) The requirement that the authors impose is that the conditional distributions must induce a unique global distribution. This fails for the control systems considered here, because the presence of high correlations between variables connected by paths of low correlation results in nonuniqueness of such a global distribution. For Example 1, the causal graph is D/P4O. The required condition involves distributions f (D), f (O; P), and f (P; O, D). The actual global distribution has O and D normally distributed with means of 0, standard deviations of 1, and correlation e0.997. P has distribution dP(O þ D), by which we denote the distribution in which all of the probability is concentrated at P ¼ O þ D. This implies what the conditional distributions must be: f (D) ¼ N(0,1), f (O; P) ¼ N(0,1), and f (P; O, D) ¼ dP(O þ D). But these are induced by any global distribution of the form P(O, D) dP(O þ D), where P(O, D) is a bivariate Gaussian with any correlation, and unconditional standard deviations for O and D of 1. This happens because the conditional distributions omit any information about the joint distribution of O and D, variables which are connected only via a third variable with which they have no correlation. Voortman and colleagues (Voortman, Dash, & Druzdzel, 2010) consider dynamical systems with cyclic dependencies, but its results on the learnability of such systems depend on the Faithfulness assumption, which, the authors note, is violated when there is an equilibrium. Our Examples 1e4 all maintain equilibria, and indeed the authors’ Theorem 4.2 fails to apply to them. However, the voltage/current example, and Example 5 are not systems in equilibrium. Even those systems violate Faithfulness, and the conclusions of the authors’ theorems do not hold for them. For the voltage/current example, the property is trivially satisfied, but the same distribution between the two variablesda bivariate Gaussian with zero correlationdwould satisfy the assumptions for the causal graph on V and I with no edges. So long as data are collected at intervals long compared with the coherence time of the waveforms, none of the four possible graphs on two nodes can be excluded. A study attempting an alternative approach (Zhang & Spirtes, 2008) considers the possibility of making a weaker Faithfulness assumption which is still sufficient to conduct causal analysis. It demonstrates that Faithfulness implies two properties which they call Adjacency Faithfulness and Orientation Faithfulness and that, while these do not together imply Faithfulness, they are (given the Markov assumption) all that is necessary for standard methods of causal inference. They then prove that if Adjacency Faithfulness is satisfied, then Orientation Faithfulness can be tested from the data, obviating the need to assume it. If Orientation Faithfulness fails the test, then the

When causation does not imply correlation Chapter | 4

67

data are not faithful to any causal graph. They also find a condition weaker than Adjacency Faithfulness, called Triangle Faithfulness, having a similar property: if Triangle Faithfulness is satisfied, Adjacency Faithfulness can be tested from the data. Each of these Faithfulness conditions is a requirement that some class of conditional correlations be non-zero. The work is restricted to acyclic graphs, so control systems are ruled out on that ground. However, if we investigate Triangle Faithfulness anyway for our examples, we find that the only triangle present in any of our causal graphs is the cyclic triangle connecting P, E, and O in Examples 2 and 4. The three vertexes of this triangle are all non-colliders (i.e. the causal arrows do not both go toward the vertex). Triangle Faithfulness requires all of the correlations between any two of these vertexes, conditional on any set of variables not including the third, to be non-zero. However, in Example 2, six of these twelve correlations are zero. The following table shows the correlations obtained from simulation (round values) between each pair of variables named at the left, conditional on the set of variables shown in the top row.

The values that are here 1 all result from mathematical identities. For example, c(OPjD) ¼ 1 because P ¼ O þ D. They are nonetheless valid correlations. In practice, measurement noise would make these correlations slightly smaller, but they will still be extreme. Correlations more typical of real experiments can be obtained only by assuming gross amounts of measurement error. In view of the cyclicity of the triangle, we might instead test the Triangle Faithfulness condition that applies in the case of a collider, but we fare no better. Each of the following correlations would be required to be non-zero.

As before, the 1 entries are mathematical identities. The undefined entries are due to the fact that fixing some of these variables may also fix one or both of the variables whose correlation is being measured. For example, c(OPjER) is undefined because fixing E and R fixes P ¼ R e E. Adding measurement noise can make these correlations well-defined (indicated by the parenthetical values), but this creates several more zeroes. Example 4 gives similar results. Triangle Faithfulness is therefore not satisfied on any interpretation of how one might apply it to these graphs, and so the failure of Faithfulness for these

68 SECTION | A Why do we need perceptual control theory?

examples is not detectable by this method. No approach along these lines can avail for these systems, because the data generated by each of them are in fact faithful to some causal graphdbut in no case are the data faithful to the real causal graph.

The fundamental problem We have seen that control systems display a systematic tendency to violate Faithfulness, whether they are at equilibrium or not. Low correlations can be found where there are direct causal effects, and high correlations between variables that are only indirectly causally connected, by paths along which every step shows low correlation. This follows from the basic nature of what a control system does: vary its output to keep its perception equal to its reference. The output automatically takes whatever value it needs to, to prevent the disturbances from affecting the perception. The very function of a control system is to actively destroy the data that current techniques of causal analysis work from. What every controller does is hold P close to R, creating a very strong statistical connection via an indirect causal connection. For constant R, variations in P measure the imperfection of the controller’s performancedthe degree to which it is not doing what it is supposed to be doing. This may be useful information if one already knows what it is supposed to be doing, as will typically be the case when studying an artificial control system, of a known design and made for a known purpose. However, when studying a biological, psychological, or social system which might contain control systems that one is not yet aware of, correlations between perception and actiondin other terminology, input and output, or stimulus and responsedmust fail to yield any knowledge about how it works. Current causal analysis methods can only proceed by making assumptions to rule out the problematic cases. However, these problematic cases are not few, extreme, or unusual, but are ubiquitous in the life and social sciences. For psychology, this point has been made experimentally (Marken & Horth, 2011; Powers, 1978). Control systems also create problems for interventions. To intervene to set a variable to a given value, regardless of all other causal influences on it, one must act so as to block or overwhelm all those influences. This is problematic when the variable to be intervened on happens to be the controlled perception of a control system. One must either act so strongly as to defeat the control system’s own efforts to control, or fail to successfully set the perception. In the former case, the control system is driven into a state atypical of its normal operation, and the resulting observations may not be relevant to finding out how it works in normal circumstances. In the latter case, all one has really done is to introduce another disturbing variable. One has not so much done surgery on the causal graph as attach a prosthesis. Arguably, introducing a new variable is the only way of intervening in a system, in the absence of hypotheses about the causal relationships among its existing variables.

When causation does not imply correlation Chapter | 4

69

When interventions are performed, a lack of appreciation of the phenomena peculiar to control systems can lead to erroneous conclusions. A candle placed near to a room thermostat will not warm it up, but apparently cool the rest of the room. If one does not know the thermostat is there, this will be puzzling. If one has discovered the furnace, and noticed that the presence of the candle correlates with reduced output from the furnace, one might be led to seek some mechanism whereby the furnace is sensing the presence of the candle. And yet there is no such sensor: the thermostat senses nothing but its own temperature and the reference setting. It cannot distinguish between a nearby candle, a crowd of people in the room, or warm weather outside. To test for the presence of control systems, one must take a different approach, by applying disturbances and looking for variables that remain roughly constant despite there being a clear causal path from the disturbance to that variable (the Test for Controlled Variables). When both a causal path and an absence of causal effect are observed, it is evidence that a control system may be present. If, at the same time, something else changes in such a way as to oppose the disturbance, that is a candidate for the control system’s output action. Discovering exactly what the control system is perceiving, what reference it is comparing it with, and how it generates its output action, may be more difficult to discover. For example, it is easy to demonstrate that mammals control their deep body temperature, less easy to find the mechanism that they sense that temperature with. The task is made more difficult by the fact that in a well-functioning control system, the error may be so small as to be practically unmeasurable, even though the error is what drives the output action.

Conclusion Dynamical systems exhibiting equilibrium behavior are already known to be problematic for causal inference, although methods have been developed to extend methods of causal inference to include some parts of this class. But the subclass of control systems poses fundamental difficulties which cannot be addressed by any extension along those lines. They specifically destroy the connections between correlation and causation which these methods depend on.

Appendix: Sufficient conditions for zero correlation between a function and its derivative Here I demonstrate the absence of correlation between a function satisfying certain weak boundedness conditions and its first derivative. Before attending to the technicalities, I note that the proofs for both theorems are almost im b Rb _ ¼ 12x2 . mediate from the observation that a xxdt a

70 SECTION | A Why do we need perceptual control theory? Theorem 4.1. Let

x be a differentiable real function, defined in the interval [a, b], such that x(a) ¼ x(b). If x is not constant then the correlation of x and x_ over [a, b] is defined and equal to zero. Proof. Write xa,b and x_a,b for the means of x and x_ over [a, b]. By replacing x by x e xa,b we may assume without loss of generality that xa,b is zero. x_a,b must exist and equal zero, since Z b 1 xðbÞ xðaÞ _ ¼ ¼0 x_a;b ¼ xdt ba a ba The correlation between x and x_ over [a, b] is defined by: Z b 1 _ xxdt ba a cx;x_ ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 10 1ffi u0 Z b Z b u 1 u@ 1 t x2 dtA@ x_2 dtA ba a ba a   xðbÞ2  xðaÞ2 =2 ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 10 1ffi u Z Z b u b u t x2 dtA@ x_2 dtA a

a

The numerator is zero and the denominator is positive (since neither x nor x_ is identically zero). Therefore cx,x_ ¼ 0. Theorem 4.2. Let x be a differentiable real function. Let x and x_ be the averages of x and x_over the whole real line. If these averages exist, and if the correlation of x and x_ over the whole real line exists, then the correlation is zero. Proof. Note that the existence of the correlation implies that x is not constant. As before, we can take x to be zero and prove that x_ is also zero. The correlation between x and x_ is then given by the limit: Z b 1 _ xxdt ba a vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi cx;x_ ¼ lim 10 1ffi a/N;b/N u0 Z b Z b u 1 u@ 1 t x2 dtA@ x_2 dtA ba a ba a

  xðbÞ2  xðaÞ2 =2 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ lim 10 1ffi a/N;b/N u Z Z b u b u t x2 dtA@ x_2 dtA a

a

When causation does not imply correlation Chapter | 4

71

Since this limit is assumed to exist, to prove that it is zero it is sufficient to construct some particular sequence of values of a and b tending to N, along which the limit is zero. Either x(b) tends to zero as b/N, or (since x ¼ 0 and x is continuous) there are arbitrarily large values of b for which x(b) ¼ 0. In either case, for any ε > 0 there exist arbitrarily large values of b such that jx(b)j < ε. Similarly, there exist arbitrarily large negative values a such that jx(a)j < ε. For such a and b, the numerator of the last expression for cx,x_ is less than ε2/2. However, the denominator is positive and non-decreasing as a/eN and b/N. The denominator is therefore bounded below for all large enough a and b by some positive value d. If we take a sequence εn tending to zero, and for each εn take values an and bn as described above, and such that an/eN and bn/N, then along this route to the limit, the corresponding approximant to the correlation is less than εn/d. This sequence tends to zero, therefore the correlation is zero. The conditions that x(a) ¼ x(b) in the first theorem and the existence of x in the second are essential. If we take x ¼ et, which violates both conditions, then x_ ¼ x and the correlation is 1 over every finite time interval. That x_ and cx,x_ exist is a technicality that rules out certain pathological cases such as x ¼ sin(log(1þjtj)), which are unlikely to arise in any practical application. We remark that although we do not require them here, corresponding results hold for discrete time series, for the same reason in its finite difference form: that ðxi þxiþ1 Þðxiþ1 xi Þ ¼ xiþ1 2  x2i .

References Dash, D. (2003). Caveats for causal reasoning with equilibrium models. Doctoral dissertation. University of Pittsburgh. Retrieved from http://d-scholarship.pitt.edu/7811/. Dash, D. (2005). Restructuring dynamic causal systems in equilibrium. In R. G. Cowell, & Z. Ghahramani (Eds.), Proc. 10th international workshop on artificial intelligence and statistics (aistats 2005) (pp. 81e88). Society for Artificial Intelligence and Statistics. Duvenaud, D., Eaton, D., Murphy, K., & Schmidt, M. (2008). Causal learning without DAGs. In Nips 2008 workshop on causality (pp. 177e190). Retrieved from http://jmlr.csail.mit.edu/ proceedings/papers/v6/. Itani, S., Ohannesian, M., Sachs, K., Nolan, G. P., & Dahleh, M. A. (2008). Structure learning in causal cyclic networks. In Nips 2008 workshop on causality (pp. 165e176). Retrieved from http://jmlr.csail.mit.edu/proceedings/papers/v6/. Iwasaki, Y., & Simon, H. A. (1994). Causality and model abstraction. Artificial Intelligence, 67(1), 143e194. https://doi.org/10.1016/0004-3702(94)90014-0. Lacerda, G., Spirtes, P., Ramsey, J., & Hoyer, P. O. (2008). Discovering cyclic causal models by independent components analysis. In D. A. McAllester, & P. Myllyma¨ki (Eds.), Proc. 24th conference on uncertainty in artificial intelligence (pp. 366e374). AUAI Press. Marken, R. S., & Horth, B. (2011). When causality does not imply correlation: More spadework at the foundations of scientific psychology. Psychological Reports, 108, 1e12. Pearl, J. (1998). Probabilistic reasoning in intelligent systems. Morgan and Kaufman.

72 SECTION | A Why do we need perceptual control theory? Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge University Press. Pearl, J. (2009). Causal inference in statistics: An overview. Statistics Surveys, 3, 96e146. https:// doi.org/10.1214/09-SS057. Powers, W. T. (1974). Behavior: The control of perception. Aldine. Powers, W. T. (1978). Quantitative analysis of purposive systems: Some spadework at the foundations of scientific psychology. Psychological Review, 85(5), 417e435. Powers, W. T. (1998). Making sense of behavior: The meaning of control. Benchmark. Spirtes, P., Glymour, C., & Scheines, R. (2001). Causation, prediction, and search. MIT Press. Voortman, M., Dash, D., & Druzdzel, M. J. (2010). Learning why things change: The differencebased causality learner. In Proc. 26th conference on uncertainty in artificial intelligence (pp. 641e650). AUAI Press. Retrieved from http://event.cwi.nl/uai2010/papers/UAI2010 0100.pdf. Zhang, J., & Spirtes, P. (2008). Detection of unfaithfulness and robust causal inference. Minds and Machines, 18(2), 239e271. https://doi.org/10.1007/s11023-008-9096-4.

Chapter 5

Unraveling the dynamics of dyadic interactions: perceptual control in animal contests Sergio M. Pellis1, Heather C. Bell2 1 Department of Neuroscience, University of Lethbridge, Lethbridge, AB, Canada; 2Behavior and Evolution, University of California at San Diego, La Jolla, CA, United States

Introduction A rock and a falcon hurtling through the sky toward the ground are subject to the same laws of physics, but the same laws that can account for the trajectory and speed of the rock’s descent are not sufficient to account for the behavior of the falcon (Cziko, 2000). The trajectory and speed of the falcon’s descent varies, seemingly capriciously, and so it plummets toward the ground in a manner very different from the rock. Hurtling toward the ground is a falcon’s way of attacking a prey, such as a pigeon, that is flying closer to the ground. Not only is the pigeon flying, so creating a moving target to which the falcon has to adjust its trajectory of descent, but also, once the pigeon detects the oncoming falcon, it takes evasive action, creating greater disturbance to the falcon’s ability to keep its descent targeted on the prey. In this context, all of the small changes in wing and tail position that lead to changes in the speed and direction of descent by the falcon make sense, as the falcon is adjusting its behavior to maintain its orientation to the pigeon so that, when the distance is closed, it can grasp the prey with its talons. That is, the falcon alters its behavior to maintain a constant perception of the world - in this case, the location of the pigeon in its visual field. This was the great insight made by Bill Powers (Powers, Clark, & McFarland, 1960a, 1960b; Powers, 1973, 2005) that led him to develop perceptual control theory (PCT), a concept that can explain why behavioral output can be so variable and yet organisms are able to achieve functional outcomes, such as the falcon catching the pigeon (e.g., Bell, 2014; Bell, Johnson, Judge, Cade, & Pellis, 2012; Carey, Mansell, & Tai, 2014; Marken & Mansell, 2013; Pellis & Bell, 2011). In PCT, an organism is envisioned as The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00005-8 Copyright © 2020 Elsevier Inc. All rights reserved.

75

76 SECTION | A Why do we need perceptual control theory?

behaving so as to maintain a perception at some specified reference level constant, which means that the organism has to be able to detect a deviation from that reference level e by a comparator subtracting the actual value of the signal from the preferred value (Ashby, 1956) e and then, act to correct that error. In this way, the overt behavior observed, such as the changing trajectory and the speed of descent of the falcon, can be accounted for as a mechanism that corrects the error signal. Moreover, this error-correcting device is envisioned by PCT to be embedded within a hierarchy, whereby higher levels of control can change lower levels by altering their reference signals. Such an arrangement allows for learning through a process described in PCT as reorganization. This reorganization allows for past experiences, such as having previously hunted in the same terrain, and multiple disturbances - such as the falcon flying upwards to avoid striking a clump of trees before resuming its descent toward the pigeon - to be taken into account by the control system (Powers, 1973, 2005). In this chapter, interactions between pairs of animals are considered from the perspective of PCT. Whether friendly interactions, such as play and sex, or agonistic ones, such as conspecific fighting and predation, the issue is what may be the relevant controlled variables (i.e., the perceptions that are maintained constant at some specified reference level). How does identifying a controlled variable (CV) in such an encounter change the way in which the overt behavior is measured compared to more traditional, non-PCT approaches? What are the complications that emerge when animals have to juggle multiple, potentially competing CVs, in an interaction? What are the methodological challenges to making sense of actions during encounters when it is not evident what and how the organisms are able to perceive the CV? These are the issues confronted in this chapter, with the view that it is the iterative interaction between observations of real-life phenomena and theory that provides a way to both improve our observations of nature and refine the theory.

Lessons from fighting The classic image of deer fighting is that they size up one another, and then, if they judge that they can beat their opponent, they face one another head-tohead, rush forward and then clash with their antlers, which may then be followed by some wrestling as the antlers remain intertwined (Clutton-Brock, Guinness, & Albon, 1982). Described in this manner, the fight clearly gives the impression of a ritualized affair, in which the animals test their strength and avoid doing serious harm to one another (Eibl-Eibesfeldt, 1961). However, this view underestimates the risk of retaliation by the opponent as a restraining factor (Maynard Smith, 1974). But restraint, whatever its causes, may be illusory (Geist, 1974). For example, when one of the deer stumbles, its

Unraveling the dynamics of dyadic interactions Chapter | 5

77

opponent may swing around and slam its antlers into an exposed flank (Geist, 1978). That such ‘taking’ advantage of an opponent’s misfortune is not an aberration is indicated by an analysis of museum hides of male deer over eighteen months old showing that 84% have puncture marks on the flanks (Geist, 1986). Members of a fighting dyad are doing their best to inflict harm, but usually fail to do so because of the actions of the opponent block attacks (Geist, 1971). When viewed from this perspective, many of the differences in fighting styles across species make sense (Geist, 1966). For example, bighorn sheep, like deer, orient head-on to clash horns, but also, like the deer, if one of the animals stumble, the other is likely to maneuver to its opponent’s side and strike its exposed flank. The large horns and reinforced skulls attest to the importance of the head-to-head strike, but as the flank strike shows, these massive weapons can be lethal when used against a non-reinforced target (Geist, 1971). Blocking access to weaker flanks requires the animals to do some behavioral maneuvering, forcing the attack to be directed at the body area best suited to receive the blow: the top of the skull. In contrast, mountain goats, another species from the North American Rockies, have short, pointed horns and they fight by adopting an anti-parallel position, with their heads opposing their opponent’s mid-flank position, and then, from this position, they lower their heads and strike upwards, plunging their horns into their opponent’s underbelly (Geist, 1965). Protection of the underbelly is provided in two ways. First, the abdominal skin is thicker than the skin in non-target areas (Geist, 1967). Second, and more critical, defensive maneuvers are used that lessen the likelihood of a successful strike or reduce the severity of the impact if the strike is delivered (Geist, 1965, 1966). The above examples of combat in several different species of ungulates graphically highlight an important principle about fighting in particular, and social interactions in general (Pellis & Pellis, 2015). Fighting involves contact by one animal’s weapons with particular targets on the other animal’s body, with the opponent then acting to block successful contact (Blanchard & Blanchard, 1994; Geist, 1978; Pellis, 1997). As a consequence, accessing the target and avoiding such contact creates a dynamic system with particular actions by either partner only comprehensible within the context of the actions by the other animal. Specifically, from a PCT perspective, the distance between the attacker’s weapons (e.g., horns, teeth) and the species-specific body targets is the perception that is controlled during interactions (Pellis & Bell, 2011). In this conceptual context, social interactions are dynamic, involving continuously changing behavior by both animals, with the behavioral output of each animal being both a disturbance to the CV of the other animal and a compensation for the disturbance created by the other animal. The insights derived from this “target as CV” perspective have implications for how to study and interpret interactions.

78 SECTION | A Why do we need perceptual control theory?

Some thoughts on methodology The most commonly used approach to study social interactions, such as conspecific fighting, is to count the occurrence of readily identifiable behavior patterns. In the combat of cockroaches, for example, such behavior patterns include lashing with their antennae, kicking and head butting (e.g., Breed, Meaney, Deuth, & Bell, 1981; Guerra & Mason, 2005). An element of the dynamic nature of the encounters can sometimes be captured that evaluate the statistical association between the actions performed by the animals (e.g., Clark & Moore, 1994; Nelson & Fraser, 1980). These approaches, however, imply that encounters are about producing particular actions, but as shown above, such abstracted actions may not be the property of the performer, but arise from the combined moves and countermoves of both animals. Given the continuous maneuvers and counter-maneuvers by the opponents, a tool that can simultaneously track the movements of both animals facilitates such an analysis, especially in the early stages when it is not clear what the target (as CV) may be. One such method is the Eshkol Wachmann Movement Notation (EWMN) (Eshkol & Wachmann, 1958). EWMN is a globographic system, designed to express relations and changes of relation between parts of the body, with the body treated as a system of articulated axes (i.e., body and limb segments). A limb is any part of a body that either lies between two joints or has a joint and an extremity. These are imagined as straight lines (axes), of constant length, which move with one end fixed to the center of a sphere. The body is represented on a horizontally ruled page into columns that denote units of time (e.g., frames of a video). The signs for movement are read from left to right and from bottom to top. Movements by any limb segment, or the body as a whole, can be described as the distal end moves across the surface of the sphere, with the proximal end being anchored in the center of the sphere. Typically, the locations on the sphere (horizontal and vertical) are at 45
angles, but the unit of angular measurement can be reduced (e.g., 22.5
) if finer grain comparisons are needed. An important feature of EWMN is that the same movements can be notated from several different perspectives: the coordinates for the position of the body segments can be scored with reference to the environment, to the body segment to which it is connected, and the movement by one animal can be described relative to the body of the other animal. By transforming the description of the same behavior from one coordinate system to the next, invariance in the behavior may emerge in some coordinates but not others, with invariance providing a clue as to the existence of a potential CV (Golani, 1976). In interactions between two animals, three measures have proven to be particularly useful to track the inter-animal relationship (e.g., Moran, Fentress, & Golani, 1981; Pellis, 1982). For simplicity, these measures are shown in an

Unraveling the dynamics of dyadic interactions Chapter | 5

79

example in which the animals’ movements were only tracked in the horizontal plane (Pellis et al., 2013). The three measurements are: (1) Partnerwise orientation: This refers to the relationship of the longitudinal axis of one animal relative to the other. One animal is selected as the focal animal and the 45
units are situated in a circle around the longitudinal axis (0e7), with 0 being situated in the direction in which the animal is facing. Wherever the other animal is located in space, its longitudinal axis is envisaged as transecting that of the focal animal carrying the EWMN coordinates, with the number pointed at by the anterior of the opponent being given that numerical value for the partnerwise orientation. For example, in Fig. 5.1Aa, the focal animal (with the numerals surrounding its body) is facing upward on the page and the other animal is standing facing the bottom of the page, thus pointing in the direction of 4 on the focal animal, giving the pair a partnerwise angle of 4. Then, as the focal animal changes its position in space, so does the other animal, leading them to maintain the same partnerwise angle (Fig. 5.1Ab).

FIG. 5.1 Three of the EWMN coordinates used to track a pair of interacting animals are illustrated in the horizontal domain (from a dorsal view) for greater sage grouse engaged in an aggressive encounter. These include partnerwise orientation (A), opposition (B) and relative distance (C). See text for further explanation. Reprinted from Pellis, S. M., Blundell, M. A., Bell, H. C., Pellis, V. C., Krakauer, A. H., & Patricelli, G. L. (2013). Drawn into the vortex: The facing-past encounter and combat in lekking male greater sage-grouse (Centrocercus urophasianus). Behaviour, 150, 1567e1599 with permission of Brill.

80 SECTION | A Why do we need perceptual control theory?

(2) Opposition: With this measure, the part of the body of one animal closest to the body part on another is scored. To score this, imagine the EWMN sphere being deflated, so that it is wrapped around each animal’s body. The front of the sphere (taking the horizontal value only) would be 0 and this value would be attached to the tip of the beak or snout, with the rearmost point as 4. Similarly, each side of the body (head, shoulder, torso) would be labeled 2 for the right side and 6 for the left side. The body parts opposed by the two animals can then be tracked during the course of the encounter. For example, in Fig. 5.1Ba, the two animals are standing in such a way that the right sides of their heads are opposing one another (2H/2H). Then, as the animals move, the points on their bodies of closest opposition changes (Fig. 5.1Bb) to the right side of their shoulders (2S/2S). (3) Relative distance: Given that videotapes are often not taken with a measurable frame of reference, the absolute distance in a metric, such as centimeters, is not possible, but the distance in terms of animal lengths (i.e., from the tip of snout or beak to the base of the tail when the animal is standing in a relaxed posture) can be used to track the relative distance, during encounters, between the animals. For example, in Fig. 5.1Ca, the two animals are standing side-by-side, facing opposite directions and are two animal lengths apart. Then, following some movement by one or both of the animals, they maintain the same orientation, but move closer together (Fig. 5.1Cb), ending up only half an animal distance apart. Combining these inter-animal variables with the actual movements performed by each animal on a notated score sheet permits the researcher to identify what each animal is contributing to the encounter. If a particular opposition remains constant despite multiple movements by the animals, the score sheet can be used to identify which partner is initiating movement away from that opposition (i.e., the disturbance) and which animal is performing the counteracting movements (i.e., the compensatory actions) to maintain that opposition, thus alerting one to the possibility that that opposition may be a CV (Fig. 5.2). A concrete example will illustrate this point. In the combat of male giant Madagascar hissing cockroaches, animals head butt their opponent’s flanks and head, with most researchers lumping these two locations of butting as one common action, “head butting” (e.g., Breed et al., 1981; Nelson & Fraser, 1980). From the perspective of the organization of fighting sequences, is it accurate to lump all cases of head butting together as if they represented a unitary behavior pattern? Videotaped staged encounters between pairs of cockroaches were analyzed using EWMN, revealing that the animals maneuver to access their opponent’s flank. However, as one cockroach circles to gain access to the other’s flank (Fig. 5.3A), its opponent rotates to face the approaching animal and so block that access (Fig. 5.3B). As a result, the two cockroaches end up facing one another (Fig. 5.3C). Frame-by-frame

Unraveling the dynamics of dyadic interactions Chapter | 5

81

FIG. 5.2 An example from a sequence of combat in greater sage grouse showing an EWMN score (upper panel), illustrating the way in which the movements of two birds are simultaneously tracked over time (video frame numbers are shown at the top of the score). In this case, the movement of the bodies of the two birds is recorded in terms of shift of body weight (e.g., leaning to the right) and changes in front, which tracks the orientation of the body in space (e.g., the bird rotates to the left). In between these two measurements of the birds’ movements are the three interanimal measurements. An aerial view of the movements of the two birds (lower panel) shows the relative positions of the birds at three instances depicted in the notation (see frames labeled (i), (ii) and (iii)). The arrows indicate the trajectories of movement by each bird. Note that, while the distance between the birds is reduced by two-thirds, the orientation and opposition remain the same due to the compensatory movements by each bird relative to the movement of the other bird. Reprinted from Pellis, S. M., Blundell, M. A., Bell, H. C., Pellis, V. C., Krakauer, A. H., & Patricelli, G. L. (2013). Drawn into the vortex: The facing-past encounter and combat in lekking male greater sage-grouse (Centrocercus urophasianus). Behaviour, 150, 1567e1599 with permission of Brill.

comparison of the movements of the two animals shows that, as one, and then the other, move to gain access to the flank, they each countermove and so continue to face their opponent. After several of these attempts to access the other’s flank, the cockroaches will butt heads, using the knobby outgrowths on their pronatum, a forward extending shield from the anterior thorax, to lock their heads together and so enable them to wrestle. If one cockroach is thrown off balance, the winner quickly circles to its opponent’s exposed flank and attacks. Similarly, if one cockroach approaches before its opponent reacts, it may manage to access the other’s flank, and then butt its body. The manner in which the opponent’s flank is attacked is also revealing. The anterior of the Madagascan hissing cockroach’s head shield protrudes forward with a slight upward curve. The cockroach dips its shield downward, wedging it beneath the lateral edge of its

82 SECTION | A Why do we need perceptual control theory?

FIG. 5.3 An aerial view of two Madagascar hissing cockroaches is shown. The two cockroaches, labeled (A) and (B), are moving and counter-moving, as one attempts to gain access to the other’s flank and its opponent turns to block it by facing its attacker. The arrows indicate the trajectories of movement of each cockroach. The three panels show that in A, cockroach (B) approaches cockroach (A) orienting toward its flank, but as (B) approaches, (A) turns to face, which leads to a change in their orientation in space but not in relation to each other, as (B) also compensates by moving laterally and around (B). Continued maneuvering by (B) to access (A)’s flank is countered more vigorously by (A), ending in a head-to-head orientation (C) (see text).

opponent’s body, thus between it and the ground. The defender counters this maneuver by rotating around its longitudinal axis, pressing the lateral surface of its body against the ground. If the attacker manages to press its head shield between its opponent’s body and the ground, it then flicks its head upward and thrusts its body forward, which flips its opponent over. Once flipped over, the attacker can bite the helpless cockroach on the exposed, softer undersides of its body. Quantitative evidence supports the conclusions drawn from the qualitative EWMN analyses (Bell, 2013). Selecting attacks in which the recipient failed to move before contact was achieved, the location of head butts on the opponent’s body was scored (Fig. 5.4A). When the flank is exposed in this manner, it is the lower thorax and abdomen that is most likely to be struck, not the head or the upper thorax (Fig. 5.4B). Once struck, the probability of the defender being flipped was recorded. The data show that the probability of being flipped over is much higher if the flank is struck rather than the head (Fig. 5.4C). Even though the data are limited, as they depend on cases of relatively rare uncontested head butts, these differences are significant (Bell, 2013). Moreover, frame-by-frame analysis suggests that, when the defender actively moves to protect its flank, the attacker maneuvers toward the defender’s lower flank, reducing the risk that the defender can turn to face its attacker and retaliate. However, given the vigorous defensive maneuvers by the defender, successful access to the lower flanks is usually blocked before contact can be made. That is, the animals compete for access to their opponent’s flank and use it as a vantage point from which to unbalance their opponent. The flank is thus a target that functions as the CV, with the actions of the two animals becoming understandable within that context. If the opponent does not countermove, or does so too slowly, the attacker immediately orients

FIG. 5.4 The orientation and consequences of head butts on different parts of the opponent’s body are shown for cockroaches. In A, an aerial view of a cockroach shows how the body was divided to score the body areas contacted by the head butts (1 ¼ side of the head and pronatum, 2 ¼ side of lower thorax, 3 ¼ anterior half of the abdomen and 4 ¼ posterior half of the abdomen). In B, the frequency with which different parts of the flank are butted is shown. In C, the probability that such butting leads to the defender being flipped over onto its back is shown. For panels B and C, the data in the bar graphs are given as means and standard deviations. Reprinted from Bell, H. C. (2013). Control in living systems: An exploration of the cybernetic properties of interactive behaviour. Unpublished doctoral dissertation, Lethbridge, AB: University of Lethbridge.

84 SECTION | A Why do we need perceptual control theory?

toward it and strikes it on its flank. In contrast, the opponent’s head is butted only because of its opponent’s countermoves, which blocks the access to its flank by facing the attacker. In this dynamic context, the head butting of the opponent’s flank arises from the attacker reducing the distance to zero on the targeted body area (the CV), whereas head butting the opponent’s head is a byproduct of the joint offensive and defensive actions of both animals. Labeling them as a single behavior pattern, as is commonly done, misses the differences in how they arise. Since the two cockroaches both avoid the situation in which the opponent gains the advantageous configuration to flip its opponent, the animals mostly maintain the position that favors defense. In a sense, the head-to-head configuration is a default CV. Such default CVs, that are actively created and defended, can be thought of as a joint articulating the two animals’ bodies together (Golani, 1976). However, as shown by the example of the cockroaches, they may arise as a compromise between CVs in which both opponents compete for the same body target (for another example involving a different default CV, see Pellis et al., 2013). Indeed, modeling multiple agents interacting in the context of competing CVs has shown that the system can settle into a stable configuration that maintains a ‘compromise CV’ constant (McClelland, 2004) e supporting the view that, in such cases, characterizing the conflicting CVs and the conditions that lead to defense of stable ‘virtual’ CVs is important for understanding the mechanisms producing the overt behavior.

Some thoughts on interpretation During play fighting, Syrian hamsters compete for access to the cheeks, which if contacted, are gently nibbled (Pellis & Pellis, 1988a). In contrast, during serious fighting, like many other murid (i.e., mouse-like) rodents (Pellis, 1997), the lower flanks and lower dorsum are bitten (Pellis & Pellis, 1988b). The cheeks are also nibbled by the male in adult sexual interactions (Pellis & Pellis, 1988b), a pattern consistent with the play fighting seen in many murid rodents (Pellis, 1993). While the actual targets contacted may differ - the nape in rats and voles, the cheeks in Syrian hamsters, and the mouth in Djungarian hamsters - what unites the pattern of play fighting in murid rodents is that they all compete for access to species-typical precopulatory targets. This consistency across murid species led to the hypothesis that in murid rodents play fighting is derived from immature versions of adult sexual behavior (Pellis & Pellis, 2009). As they approach sexual maturity, the play fighting of some of these species becomes increasingly difficult to distinguish from sex (Pellis, 1993), but in some, the play has been modified so that, while involving the same targets, at all ages, play remains distinct from sex (Pellis & Pellis, 2009). For Syrian hamsters, an alternative explanation has been posited for the origin and development of play fighting.

Unraveling the dynamics of dyadic interactions Chapter | 5

85

Based on the involvement of some common neural circuitry in both serious and playful fighting, Yvon Deville and his colleagues argued that play fighting in Syrian hamsters is an immature version of serious fighting and that, over development, play fighting grades into serious fighting (Cheng, TaravoshLahn, & Delville, 2008). An important part of the evidence for this hypothesis is an apparent change in the targeting over the course of development. Like other murid rodents, in these hamsters, the rump is the main target for serious aggression (Pellis & Pellis, 1988b). In the test paradigm used by Delville, when play fighting first starts to occur shortly after weaning, the cheeks are the predominant target, but then, as the animals approach sexual maturity, the rump becomes the predominant target. Over the course of this transition, there are intermediary bites that are directed at the mid-flank. Measured in terms of the shifting location of biting, from cheeks to mid-flank to lower flank, it seems reasonable to assume that, with age, playful biting grades seamlessly into aggressive biting, and so draw the conclusion that play fighting is an immature form of serious fighting (Delville, David, TaravoshHahn, & Wommack, 2003; Taravosh-Lahn & Delville, 2004; Wommack, Taravosh-Lahn, David, & Delville, 2003). However, this interpretation is contingent on how the bites directed to the mid-flank arise. When defending against a playful attack to the cheek, the recipient of the attack rolls over onto its back, and then, from this position in which it faces the attacker, it blocks access to its cheeks (Pellis & Pellis, 1988a). The attacker attempts to maneuver to gain access to its opponent’s cheeks by either moving slightly laterally or by rotating its head and neck around its longitudinal axis so as to bring its mouth to a position opposite one of its opponent’s cheeks. However, as shown in Fig. 5.5, the defender does not remain passive, but moves its head to keep its mouth oriented toward its partner’s mouth and uses its forepaws to help in blocking its partner’s moves (see panels A and B). Then, following continued defensive maneuvers by the supine hamster that block access to its cheeks, the attacker delivers bites lower down its opponent’s flank (see panels C, D and E). That is, the location of the area on the flank ‘targeted’ changes, but not because the attacker is switching from playful to aggressive biting, but rather, as in the cockroach example above, because the defender denies the attacker access to its cheeks e the preferred target. In Syrian hamsters, biting of an opponent’s mid-flank mostly occurs in the context of this on-back/on-top wrestling in which the animals compete for access to the cheeks. In contrast, bites to the rump, are typically delivered when the defender is on all fours (Pellis & Pellis, 1988a, 1988b). That is, bites to the mid-flanks do not arise as a consequence of defense of the rump (Pellis & Pellis, 1988b). All that is necessary for mid-flank bites to increase in frequency with age is for the hamsters to become more vigorous in the defense of their cheeks during play. With the onset of sexual maturity, hamsters housed together form dominance relationships and subordinate hamsters tend to lie passively when the dominant initiates nibbling of the cheeks (Pellis & Pellis,

86 SECTION | A Why do we need perceptual control theory?

FIG. 5.5 Drawings taken from videotaped material show a sequence of play fighting in golden hamsters in which they are attacking and defending the cheek. The panels (AeF) show different phases of the interaction (see text). Note that, in the last frame (F), as the attacker has shifted its attack to its opponent’s lower flank, the defender launches a counterattack to the side of the attacker’s face. Reprinted from Pellis, S. M., & Pellis, V. C. (1988a). Play-fighting in the Syrian golden hamster Mesocricetus auratus Waterhouse, and its relationship to serious fighting during post-weaning development. Developmental Psychobiology, 21, 323e337 with permission of Wiley.

1993). In Delville’s paradigm, unfamiliar hamsters of different ages are tested in a neutral cage (Delville et al., 2003). In the absence of an established dominance relationship, the play of these unfamiliar hamsters can become very rough, with the play leading to more vigorous competition for the cheeks and an increased likelihood that the encounter escalates into serious aggression. Thus, in this testing paradigm, with increased age, the increased vigor of defense during play increases the likelihood of upper and mid-flank bites as the defense is more likely to block access to the cheeks. Also, as the play is more likely to escalate to serious fighting, bites directed to the rump become increasingly frequent. These considerations strongly suggest that mid-flank bites are not transitional ones that reflect a gradual shift from play to aggression, but rather, are an outcome of the dynamics of attack and defense of the cheek. If so, then aggression and play remain distinct at all ages, one involving bites to the rump and the other the nibbling of the cheeks. Measuring mid-flank bites as a distinct category of biting may thus be an inappropriate abstraction from the interactions.

Unraveling the dynamics of dyadic interactions Chapter | 5

87

Problems needing novel solutions The concept of the target as a CV makes sense of much of the behavior that occurs during competitive interactions. At a simple level of organization, the target can be thought of as a moving location in the environment, much like the example of the pigeon evading the falcon that began this chapter. When viewed in this perspective, there are useful precedents that have established how such a system may work. For example, chasing a ball or a Frisbee involves the subject maintaining constant some perception of the object relative to the eye and the horizon, and then moving faster or slower to maintain that perception until the distance is closed and the object is caught (Shaffer & McBeath, 2005; Shaffer, Krauchunas, Eddy, & McBeath, 2004). Indeed, the organization of such a tracking system has been usefully conceptualized in an explicitly PCT framework (Marken, 2005). However, a ball or a Frisbee does not have a control system that guides its movements so as to avoid capture, in the way a pigeon does to avoid a falcon or a cockroach does to avoid being head butted on its flank by another cockroach. To be sure, there are unpredictable disturbances that arise when a Frisbee moves through the air, such as a sudden gust of wind, but the difficulty faced by the catcher is of a lower order of magnitude compared to when the object actively makes compensatory movements to prevent capture. Nonetheless, from a PCT perspective, all that matters is the error signal, not how that error is created - so in that sense, the situation with the competitor is only more complex because of the number of compensatory actions required. However, in competitions between two animals, the defender not only moves to avoid being caught, but may also counterattack. For example, in Fig. 5.5, once the hamster in the ontop position bites the flank of the supine hamster, its opponent, in turn, bites the side of its face (see panels E and F). Such retaliatory, defensive bites to the side of the face are common in both playful and serious fighting of murid rodents (Pellis, 1997; Pellis & Pellis, 2009). How animals adjust to competing CVs is an important avenue for understanding dyadic interactions. Moreover, analyses of various competitive interactions also highlight some puzzles that require extending PCT models. A central tenet of PCT is that only CVs that can be perceived can be controlled (Cziko, 2000), but sometimes, how the subject perceives the target is unclear. The predatory behavior of some small mammals illustrates this problem. First, many predators actively hunt at night, making vision less useful in guiding an attack. For nocturnal predators locating and tracking prey, vision may be used when light is available but audition and touch (usually via the vibrissae) may be substituted when it is not (Huang, 1986; Pellis, Vanderlely, & Nelson, 1992). Such sensory substitution would imply a higher-level reference signal that can be compared to many, different, lower-level sensory inputs. In other situations, the solution to an obstructed perceptual input

88 SECTION | A Why do we need perceptual control theory?

cannot be so readily explained. For example, consider a cat attacking and killing a mouse. The target for delivering a lethal bite is the nape of the neck (Leyhausen, 1979). If a cat approaches a mouse from its rear, its nape is in sight, but if the mouse turns to face the cat, then its face blocks the cat’s view of the nape. Yet, the cat maneuvers to gain a side or rear orientation to the mouse (Pellis et al., 1988). Similarly, smaller predators, such as marbled polecats (Ben-David, Pellis, & Pellis, 1991) and various species of a family of marsupial carnivores (Dasyuridae), actively maneuver to reach around for the nape of their prey when their access and view of its nape is blocked by the prey’s orientation toward its attacker (Pellis & Nelson, 1984; Pellis & Officer, 1987). That is, the target continues to attract attacks even though it is not directly perceivable. Indeed, one small marsupial carnivore, the kowari, overcomes the mouse’s frontal defense to reach over and bite its nape (Fig. 5.6). In this case, the kowari launches an attack on its prey to a body target that is not directly visible. It is possible that there are a series of CVs that are sequentially engaged at different points in the attack. However, if every conceivable counter maneuver by the prey requires a matching CV, then this could lead to an infinite number of CVs. A more parsimonious possibility for how an animal can maneuver to gain access to an unseen target is that there is an ‘image’ of where in relation to the partner’s body the target resides, so that views of the partner from any perspective, even if the target cannot be sensed, can enable the performer to move so as to get closer to the target. If this is so, where in the control hierarchy is this ‘image’ constructed from the available lower-level perceptions and what are the reference signals that are changed by the higher-level control circuits that enable the animal to continue to track the target? As argued by Gibson (1979), the subject may simply extract out a higher order invariant from the sensory array, without the need for any reference signals. Pattern extraction may be achieved in multiple ways: by built-in biases as to what features of the world are attended (Buchanan, 2009), by the ability to fill in the gaps in an incomplete perceptual array (Wertheimer, 1961), or by some more complex computational algorithm (Marr, 1982). How any or all of these theoretical approaches may be usefully integrated with PCT to deal with the tracking of ‘unseen’ targets remains to be determined. Another example will illustrate an extension to the problem and lay the foundations for how a PCT approach may be deployed. As noted above, most small carnivores kill mice with a bite to the nape of its neck (Pellis & Officer, 1987). If the mouse runs away, they will run directly after it, using their mouth and forepaws in an attempt to grasp the hindquarters of the fleeing mouse, which, once caught, is held and a bite is delivered to the nape from the rear or the side (Pellis & Nelson, 1984). The kowari, however, does something very different. If the mouse runs away, the kowari runs in an arc, around the mouse, so that it ends up in front of the mouse by intercepting

Unraveling the dynamics of dyadic interactions Chapter | 5

89

FIG. 5.6 Drawings taken from videotaped material show a sequence of predatory attack on a mouse by a kowari. The panels are numbered from 1 to 12, representing frame numbers. Note that the kowari has to raise its snout up and over the mouse’s head to access the nape. Reprinted from Pellis, S. M., & Officer, R. C. E. (1987). An analysis of some predatory behaviour patterns in four species of carnivorous marsupials (Dasyuridae), with comparative notes on the eutherian cat Felis catus. Ethology, 75, 177e196 with permission of Paul Parey.

its path (Fig. 5.7). Once the mouse is halted by the kowari, attacking its nape from the front raises the same issue as above: the attack is launched from an orientation from which the target may not be directly perceptible. But the kowari also illustrates a second problem: rather than directly following the prey as it attempts to reduce the error signal (i.e., reducing the distance between the target and the mouth) as do other small carnivores (Pellis & Nelson, 1984; Pellis & Officer, 1987), it runs away from the mouse, initially increasing the distance between them. Once the kowari is in place in front of the oncoming mouse, that distance is reduced (as in Fig. 5.6). The question arises as to how the control system can be organized so that an increasing distance from the target can be part of the same system that results in the eventual decreasing of the distance with the target to zero.

90 SECTION | A Why do we need perceptual control theory?

FIG. 5.7 An aerial view that shows a kowari running in an arc around a mouse that is fleeing from it, so as to intercept and attack the mouse from a frontal orientation. The mouse is shown as a black arrow, with the gray arrow indicating its direction of movement. Note that for the kowari to get in front of the mouse, it initially moves away from the mouse, before moving back toward the mouse (thin arrow). For more details, see text and Pellis & Officer (1987).

Both problems - the ‘image’ of a target with or without sensory substitution and moving away from the target in order to gain access to the target at the terminal part of the strategy - require careful consideration with regard to how control systems can be organized to solve them. A thought experiment may be useful to identify the items that need to be accounted for in a control system. Consider hunting a deer with a rifle. You come across the deer as it is grazing in an open meadow. As one of the rules that you follow is to cause the deer the minimum pain possible, you need a clean shot through its heart. However, as the deer’s heart is not visible, you use cues about the deer’s body in order to guess its likely location and that is where you aim. With experience, you learn to modify that location slightly, based on the size and age of the deer. So, even though it is not directly visible as a target, the heart can still act as the ‘perception’ that is controlled by altering your behavior (i.e., the aiming of the rifle), with experience providing a means of reorganizing higher-level

Unraveling the dynamics of dyadic interactions Chapter | 5

91

reference signals and the estimation of the size of error signals. Something like this may apply to the ‘not directly visible target’ discussed for animal contests above. Successful contact with the target in the case of hunting becomes more complex if the deer is not standing still, but is moving. If the deer is running, the rifle is aimed not at the presumed heart, but a little ahead of the desired target e this is known as ‘leading the target’. The further away the deer is to the hunter and the faster the deer is running, the gap between the heart and the location in space at which the rifle is actually being aimed becomes greater, as the gap reflects the time differential between the bullet traveling to the target and the deer moving forward. Essentially, the rifle is aimed in such a way that it ‘predicts’ where the target (the heart) to be intercepted by the bullet will be. This is much like what the kowari does, with you, the hunter, and the kowari, both failing to make contact with the target if the prey zigzags to one side, rather than continuing to travel in the same direction. However, maintaining your aim is a simpler control problem than is the case for the kowari. In the case of the deer hunter, all you need to do is to maintain a constant lead between the location aimed at and the location of the deer’s heart, with the only complication being to estimate the size of the gap based on the distance and speed of the deer. In contrast, the kowari, by traveling in an arc, first away from the mouse and then back toward it, is continually changing the distance between itself and its target. Several possible control systems could be envisioned to explain the arcing run of the kowari. For example, there could be a hierarchy of control circuits that could alter the acceptable value of the distance, to some maximum, before reducing it to zero. Alternatively, all that may be needed is a higher-level control circuit that modifies the reference signals of two lower-level circuits: the first directs the kowari to reach a virtual target ahead of its prey, and then, once that position is achieved, the second involves moving directly toward the prey, the ultimate target. That the outward journey involves the control of different perceptions is suggested by the hunting of jumping spiders, which take such long detours that the prey are removed from their visual field before they re-engage visually with them (Wilcox & Jackson, 2002). However, given that the direction and speed of the mouse may change, using either of these control systems would require the kowari to update continually its relative distance with the mouse or the location of the virtual target. Another alternative may simplify the demands on such updating. What may be controlled is the changing pattern of the relationship between the predator and prey as a whole (Golani, 1981; Pellis, Pellis, & Iwaniuk, 2014). Such homeokinetic constancy would be more like the ‘guess’ of where the prey may be, and so what would be defended would be an image of what that relationship should look like. A malleable image could arise from a CV with a changing loading factor (i.e., CV þ N) that is updated by perceptions at different points in the sequence. In this way, one could imagine how an error signal could be generated, but as in the case of the imagined location of the

92 SECTION | A Why do we need perceptual control theory?

hidden target, the issue is how and where in an hierarchical control system such an ‘image’ is constructed and how lower-level reference signals are altered to maintain the image in the face of disturbances. Even if any of these hypothesized control systems can adequately account for the arcing, the complexity of the mechanisms involved in producing the arcing maneuver also needs to be determined. For example, the evidence suggests that the sensorimotor arrangement for the hunting of jumping spiders may not require a physical, independent component acting as a comparator (Wilcox & Jackson, 2002). This would be akin to the Watt governor, a mechanism developed by James Watt in the 18th century to control the speed of steam locomotives. A flywheel device’s arms extend as it rotates when more steam enters and subsides when there is less steam. When the arms of the flywheel rise, a lever is pulled up that shuts off the fuel valve, reducing motor output, whereas, when the arms lower, more fuel enters and the motor’s output increases. Thus, the speed of the locomotive is maintained constant, irrespective of changes in terrain or fuel levels. Here also, there is no separate comparator and reference signal, so no computation of an error signal is needed (Clark, 1998). However, the particular speed maintained by a governor depends on the weight of the flywheels, the length of the flywheel’s arms, the aperture of the valve, and so on. That is, the device is constructed so as to bias it to maintain a particular speed, and it achieves this by a negative feedback loop. In this sense, the whole device may be thought of as the comparator. Similarly, the reference signal may be thought of as the point at which the movement of the arms begins to influence the flow of fuel to the engine. Whether control systems require a physical comparator or not, higher levels in the hierarchy of the control system may be engaged to modify the CV or how effectively the CV is regulated. For example, rats protect small pieces of food held in their mouths by swerving away when another rat approaches its mouth (Whishaw, 1988). The magnitude of the dodging is based on the food holder’s gaining and maintaining a particular distance with the robber’s mouth (Bell & Pellis, 2011): The greater the movement toward the owner by the robber, the greater the size of the dodge. On average, the inter-animal distance at which a rat initiates a dodge is around 4 cm. The robber’s approach is detected visually by the defender or by the robber contacting the tips of the defender’s vibrissae. Visual occlusion and trimming the vibrissae diminishes this distance to under 2 cm, as the animal now uses contact with the guard hairs on their fur to detect the approaching robber (Pellis et al., 1996). That is, the distance regulated depends on the sensory input available. Regulatory modification of the CV at a higher level in the hierarchy is possible. Visually intact, adult rats that as juveniles had been reared in social isolation begin dodging at a greater inter-animal distance (Bell, 2014). As the sensory input is unchanged, this change in the CV seems to arise from a brain-level alteration on reference level for the CV. Finally, in normally reared rats, damage to their medial prefrontal cortex leads to animals

Unraveling the dynamics of dyadic interactions Chapter | 5

93

that maintain an inter-animal distance like that of intact rats, but they are less effective in doing so, with the variation in the distance maintained being greater (Himmler et al., 2014). Given that these rats are not deficient in motor output, the problem is one related to a change in a control mechanism needed for coordinating movements with the robber.

Conclusion: where to go from here? Given that many different control system models could be generated to explain the overt behavior, one potentially valuable research tool with which to identify the possibilities that are most likely to be fruitful would be to develop computer simulations that take the known elements into account and then see what is necessary to create a model that is sufficient to recreate the actual behavior (Powers, 2009). The iterative process of modeling and comparing those models to the real behavior (e.g., Bell, Bell, Schank, & Pellis, 2015; May et al., 2006; Schank & Alberts, 1997) may guide us in identifying the kinds of mechanisms that are needed. This would then generate hypotheses that could be tested on real animals. A simple model could start with the tracking of an object in space, as already noted for catching a ball or a Frisbee (Marken, 2013; see Powers, 2009, for examples of such simulations). In such a situation, there is one CV with most of the behavior being explained by how it maintains the reference value for that CV. A somewhat more complex scenario can arise when there is more than one CV involved. For example, packs of wolves hunt in a coordinated fashion, suggesting complex, cognitive evaluations of what both the prey and pack mates are doing (MacNulty, Mech, & Smith, 2007). A computer simulation of such hunting has shown that each wolf in the pack needs only to follow two rules e track and close the distance with the prey and maintain a constant spatial relationship (distance and relative position) with the closest wolf e to produce the seemingly complex coordination between pack members during the hunt (Muro, Escobedo, Spector, & Coppinger, 2011). In terms of PCT, each wolf is simultaneously fulfilling two CVs. As already discussed above, such simultaneously maintained CVs may lead to a compromise configuration that is itself maintained and so acts as a virtual CV (e.g., Golani, 1976; McClelland, 2004; Pellis et al., 2013). In other cases, CVs may compete, leading to a higher-level control circuit that needs to select which CV to follow at particular moments in an interaction. For example, newly hatched precocial birds, such as geese, learn to follow their mother by a process referred to as ‘imprinting’ (Lorenz, 1935). The focus of research has been with regard to how goslings imprint on the mother and so preferentially follow her (e.g., Bolhuis & Honey, 1998). PCTbased computer simulations (Powers, 2009), building on a model called ‘crowd demo’ (see www.billpct.org for an animated view of this model and Tucker, Schweingruber, and McPhail (1999), for an application of this model

94 SECTION | A Why do we need perceptual control theory?

to explore the composite behavior of multiple agents), provide insight into this phenomenon. Using this modeling approach (Pellis & Bell, 2011), we found that when a single CV was used e the distance between each individual gosling and the mother e the end result would be a swarm of goslings around the mother (Fig. 5.8A). Yet, inspection of the many pictures of imprinted

FIG. 5.8 Drawings of simulated goslings (g) as they follow their mother (M). In the sequence (1e3) under A, the goslings orient their movement exclusively to the mother, with the end result being that the gosling form a semicircular swarm around the mother. In the sequence (1e3) under B, the goslings orient their movement to both the mother and each other, with the end result being that the goslings form a line in front of the mother. Reprinted from Pellis, S. M., & Bell, H. C. (2011). Closing the circle between perceptions and behavior: A cybernetic view of behavior and its consequences for studying motivation and development. Developmental Cognitive Neuroscience, 1, 404e413 with permission from Elsevier.

Unraveling the dynamics of dyadic interactions Chapter | 5

95

goslings following their mother or, even more famously, Lorenz himself, reveals that more often than not, the goslings follow in a line, not a swarm. The only way we could produce virtual goslings to end in a linear relationship with one another was to have them follow two rules, each with a distinctive CV e follow the largest object and follow the nearest object. This led to the gosling nearest the mother to follow her and the other goslings to follow the nearest sibling, thus producing the linear arrangement among the goslings in following the mother (Fig. 5.8B). Critically, the two CVs need to have a hierarchical arrangement relative to one another, in which one has priority in one case and the other has priority in the other. It is beyond the scope of the present chapter to develop simulation models that can use basic PCT principles to produce virtual animals that can replicate the behavior of real animals that continue to attack targets that are not directly perceived or move to head off the escape path of prey. Nonetheless, the examples discussed above illustrate how a simulation-based approach can start to identify how many CVs need to be incorporated and what their hierarchical relationship may be in recreating behavior and outcomes that look like that of the real animals (e.g., Bell et al., 2015; Muro et al., 2011; Schank & Alberts, 1997). Using this approach, we suggest, will allow researchers to move between theory and observation of real phenomena in productive ways in trying to understand how complex inter-animal interactions are organized.

Acknowledgments We thank Bill Powers for teaching us so much about how to make sense of behavior from a control systems perspective and Louise Barrett, Warren Mansell, Kent McClelland and Vivien Pellis for their thoughtful comments on the present work.

References Ashby, W. R. (1956). An introduction to cybernetics. London, UK: Chapman & Hall Ltd. Bell, H. C. (2013). Control in living systems: An exploration of the cybernetic properties of interactive behaviour. Unpublished doctoral dissertation. Lethbridge, AB: University of Lethbridge. Bell, H. C. (2014). Behavioral variability in the service of constancy. International Journal of Comparative Psychology, 27, 196e217. Bell, H. C., Bell, G. D., Schank, J. A., & Pellis, S. M. (2015). Attack and defense of body targets in play fighting: Insights from simulating the ‘keep away game’ in rats. Adaptive Behavior, 23, 371e380. Bell, H. C., Johnson, E., Judge, K. A., Cade, W. H., & Pellis, S. M. (2012). How is a cricket like a rat? Insights from the application of cybernetics to evasive food protective behaviour. Animal Behaviour, 84, 843e851. Bell, H. C., & Pellis, S. M. (2011). A cybernetic perspective on food protection in rats: Simple rules can generate complex and adaptable behaviour. Animal Behaviour, 82, 659e666.

96 SECTION | A Why do we need perceptual control theory? Ben-David, M., Pellis, S. M., & Pellis, V. C. (1991). Feeding habits and predatory behaviour of marbled polecats (Vormella peregusna syriaca): I. Killing methods with relation to prey size and prey behaviour. Behaviour, 118, 127e143. Blanchard, R. J., & Blanchard, D. C. (1994). Environmental targets and sensorimotor systems in aggression and defence. In S. J. Cooper, & C. A. Hendrie (Eds.), Ethology and psychopharmacology (pp. 133e157). New York: John Wiley & Sons. Bolhuis, J. J., & Honey, R. C. (1998). Imprinting, learning and development: From behavior to brain and back again. Trends in Neurosciences, 21, 306e311. Breed, M. D., Meaney, C., Deuth, D., & Bell, W. J. (1981). Agonistic interactions of two cockroach species, Gromphadorhina portentosa and Supella longipalpa (orthoptera (dictyoptera): Blaberidae, blattellidae). Journal of the Kansas Entomological Society, 54, 197e208. Buchanan, B. (2009). Onto-ethologies: The animal environments of uexku¨ll, heidiger, merleauponty, and deleuze. Albany, NY: SUNY Press. Carey, T. A., Mansell, W., & Tai, S. J. (2014). A biopsychological model based on negative feedback and control. Frontiers in Human Neuroscience, 8. https://doi.org/10.3389/ fnhum.2014.00094. Cheng, S.eY., Taravosh-Lahn, K., & Delville, Y. (2008). Neural circuitry of play fighting in golden hamsters. Neuroscience, 156, 247e256. Clark, A. (1998). Time and mind. Journal of Philosophy, 95, 354e376. Clark, D. C., & Moore, A. J. (1994). Social interactions and aggression among male Madagascar hissing cockroaches (Gromphadorhina portentosa) in groups (Dictyoptera: Blaberidae). Journal of Insect Behavior, 7, 199e215. Clutton-Brock, T. H., Guinness, F. E., & Albon, S. D. (1982). Red deer: Behavior and ecology of two sexes. Chicago, IL: The University of Chicago Press. Cziko, G. (2000). The things we do. Using the lessons of Bernard and Darwin to understand the what, how, and why of our behavior. Cambridge, MA: MIT Press. Delville, Y., David, J. T., Taravosh-Hahn, K., & Wommack, J. C. (2003). Stress and the development of agonistic behavior in golden hamsters. Hormones and Behavior, 44, 263e270. Eibl-Eibesfeldt, I. (1961). The fighting behavior of animals. Scientific American, 205, 112e122. Eshkol, N., & Wachmann, A. (1958). Movement notation. London, UK: Weidenfeld & Nicholson. Geist, V. (1965). On the rutting behavior of the mountain goat. Journal of Mammalogy, 45, 551e568. Geist, V. (1966). The evolution of hornlike organs. Behaviour, 27, 175e214. Geist, V. (1967). On fighting injuries and dermal shields of mountain goats. Journal of Wildlife Management, 31, 192e194. Geist, V. (1971). Mountain sheep. Chicago, IL: The University of Chicago Press. Geist, V. (1974). On the fighting strategies in animal combat. Nature, 250, 354. Geist, V. (1978). On weapons, combat and ecology. In L. Krames, P. Pliner, & T. Alloway (Eds.), Advances in the study of communication and affect. Vol. 4, Aggression, dominance and individual spacing (pp. 1e30). New York, NY: Plenum Press. Geist, V. (1986). New evidence for high frequency of antler wounding in cervids. Canadian Journal of Zoology, 64, 380e384. Gibson, J. J. (1979). The ecological approach to visual perception. Boston, MA: Houghton Mifflin. Golani, I. (1976). Homeostatic motor processes in mammalian interactions: A choreography of display. In P. P. G. Bateson, & P. H. Klopfer (Eds.), Perspectives in ethology (Vol. 2, pp. 69e134). New York, NY: Plenum Press.

Unraveling the dynamics of dyadic interactions Chapter | 5

97

Golani, I. (1981). The search for invariants in behavior. In K. Immelmann, G. W. Barlow, L. Petrinovich, & M. Cain (Eds.), Behavioral development: The Bielefeld interdisciplinary project (pp. 372e390). Cambridge, UK: Cambridge University Press. Guerra, P. A., & Mason, A. C. (2005). Information on resource quality mediates aggression between male Madagascar hissing cockroaches, Gromphadorhina portentosa (Dictyoptera: Blaberidae). Ethology, 111, 626e637. Himmler, B. T., Bell, H. C., Horwood, L., Harker, A., Kolb, B., & Pellis, S. M. (2014). The role of the medial prefrontal cortex in regulating inter-animal coordination of movements. Behavioral Neuroscience, 128, 603e613. Huang, C. (1986). Detection of prey by captive Ningaui ridei and Sminthopsis griseoventer (marsupialia: Dasyuridae). Australian Mammalogy, 10, 23e26. Leyhausen, P. (1979). Cat behavior: The predatory and social behavior of domestic and wild cats. New York: Garland STPM Press. Lorenz, K. (1935). Der Kumpan in der Umwelt des Vogels. Journal of Ornithology, 83, 137e213. MacNulty, D. R., Mech, L. D., & Smith, D. W. (2007). A proposed ethogram of large-carnivore predatory behavior, exemplified by the wolf. Journal of Mammalogy, 88, 595e605. Marken, R. S. (2005). Optical trajectories and the informational basis of fly ball catching. Journal of Experimental Psychology, 31, 630e634. Marken, R. S. (2013). Taking purpose into account in experimental psychology: Testing for controlled variables. Psychological Reports, 112, 184e201. Marken, R. S., & Mansell, W. (2013). Perceptual control as a unifying concept in psychology. Review of General Psychology, 17, 190e195. Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. New York, NY: Freeman. Maynard Smith, J. (1974). The theory of games and the evolution of animal conflicts. Journal of Theoretical Biology, 47, 209e221. May, C. J., Schank, J. G., Joshi, S., Tran, J., Taylor, R. J., & Scott, I.-E. (2006). Rat pups and random robots generate similar self-organized and intentional behavior. Complexity, 12, 53e66. McClelland, K. (2004). The collective control of perception: Constructing order from conflict. International Journal of Human-Computer Studies, 60, 65e99. Moran, G., Fentress, J. C., & Golani, I. (1981). A description of relational patterns of movement during ‘ritualized fighting’ in wolves. Animal Behaviour, 29, 1146e1165. Muro, C., Escobedo, R., Spector, L., & Coppinger, R. (2011). Wolf-pack (Canis lupus) hunting strategies emerge from simple rules in computational simulations. Behavioural Processes, 88, 192e197. Nelson, M. C., & Fraser, J. (1980). Sound production in the cockroach, Gromphadorhina portentosa: Evidence for communication by hissing. Behavioral Ecology and Sociobiology, 6, 305e314. Pellis, S. M. (1982). An analysis of courtship and mating in the cape barren goose Cereopsis novaehollandiae latham based on Eshkol-Wachman movement notation. Bird Behavior, 4, 30e41. Pellis, S. M., O’Brien, D. P., Pellis, V. C., Teitelbaum, P., Wolgin, D. L., & Kennedy, S. (1988). Escalation of feline predation along a gradient from avoidance through “play” to killing. Behavioral Neuroscience, 102, 760e777.

98 SECTION | A Why do we need perceptual control theory? Pellis, S. M. (1993). Sex and the evolution of play fighting: A review and a model based on the behavior of muroid rodents. The Journal of Play Theory & Research, 1, 56e77. Pellis, S. M. (1997). Targets and tactics: The analysis of moment-to-moment decision making in animal combat. Aggressive Behavior, 23, 107e129. Pellis, S. M., & Bell, H. C. (2011). Closing the circle between perceptions and behavior: A cybernetic view of behavior and its consequences for studying motivation and development. Developmental Cognitive Neuroscience, 1, 404e413. Pellis, S. M., Blundell, M. A., Bell, H. C., Pellis, V. C., Krakauer, A. H., & Patricelli, G. L. (2013). Drawn into the vortex: The facing-past encounter and combat in lekking male greater sagegrouse (Centrocercus urophasianus). Behaviour, 150, 1567e1599. Pellis, S. M., McKenna, M. M., Field, E. F., Pellis, V. C., Prusky, G. T., & Whishaw, I. Q. (1996). Uses of vision by rats in play fighting and other close quarter social interactions. Physiology & Behavior, 59, 905e913. Pellis, S. M., & Nelson, J. E. (1984). Some aspects of predatory behaviour of the quoll Dasyurus viverrinus. Australian Mammalogy, 7, 5e15. Pellis, S. M., & Officer, R. C. E. (1987). An analysis of some predatory behaviour patterns in four species of carnivorous marsupials (Dasyuridae), with comparative notes on the eutherian cat Felis catus. Ethology, 75, 177e196. Pellis, S. M., & Pellis, V. C. (1988a). Play-fighting in the Syrian golden hamster Mesocricetus auratus Waterhouse, and its relationship to serious fighting during post-weaning development. Developmental Psychobiology, 21, 323e337. Pellis, S. M., & Pellis, V. C. (1988b). Identification of the possible origin of the body target which differentiates play-fighting from serious fighting in Syrian golden hamsters Mesocricetus auratus. Aggressive Behavior, 14, 437e449. Pellis, S. M., & Pellis, V. C. (1993). The influence of dominance on the development of play fighting in pairs of male Syrian golden hamsters (Mesocricetus auratus). Aggressive Behavior, 19, 293e302. Pellis, S. M., & Pellis, V. C. (2009). The playful brain. Venturing to the limits of neuroscience. Oxford, UK: Oneworld Press. Pellis, S. M., & Pellis, V. C. (2015). Are agonistic behavior patterns signals or combat tactics e or does it matter? Targets as organizing principles of fighting. Physiology & Behavior, 146, 73e78. Pellis, S. M., Pellis, V. C., & Iwaniuk, A. N. (2014). Pattern in behavior: The characterization, origins and evolution of behavior patterns. Advances in the Study of Behavior, 46, 127e189. Pellis, S. M., Vanderlely, R., & Nelson, J. E. (1992). The roles of vision and vibrissae in the predatory behaviour of northern quolls Dasyurus hallucatus (Marsupialia: Dasyuridae). Australian Mammalogy, 15, 55e60. Powers, W. T. (1973). Behavior: The control of perception. London, UK: Wildwood House. Powers, W. T. (2005). Behavior: The control of perception (2nd ed.). New Canaan, CN: Benchmark Publications, Inc (The second edition explores novel topics in new chapters). Powers, W. T. (2009). Living control systems: The fact of control. New Canaan, CN: Benchmark Publications, Inc. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960a). A general feedback theory of human behavior I. Perceptual & Motor Skills, 11, 71e88. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960b). A general feedback theory of human behavior II. Perceptual & Motor Skills, 11, 309e323. Schank, J., & Alberts, J. (1997). Self-organized huddles of rat pups modeled by simple rules of individual behavior. Journal of Theoretical Biology, 189, 11e25.

Unraveling the dynamics of dyadic interactions Chapter | 5

99

Shaffer, D. M., Krauchunas, S. M., Eddy, M., & McBeath, M. K. (2004). How dogs navigate to catch Frisbees. Psychological Science, 15, 437e441. Shaffer, D. M., & McBeath, M. K. (2005). Naı¨ve beliefs in baseball: Systematic distortion in perceived time of apex for fly balls. Journal of Experimental Psychology, 31, 1492e1501. Taravosh-Lahn, K., & Delville, Y. (2004). Aggressive behavior in female golden hamsters: Development and the effect of repeated social stress. Hormones and Behavior, 46, 428e435. Tucker, C. W., Schweingruber, D., & McPhail, C. (1999). Simulating arcs and rings in gatherings. International Journal of Human-Computer Studies, 50, 581e588. Wertheimer, M. (1961). Experimental studies on the seeing of motion. In T. Shipley (Ed.), Classics in psychology (pp. 1032e1089). New York, NY: Philosophical Library (originally published in 1912). Whishaw, I. Q. (1988). Food wrenching and dodging: Use of action patterns for the analysis of sensorimotor and social behavior in the rat. Journal of Neuroscience Methods, 24, 169e178. Wilcox, S., & Jackson, R. (2002). Jumping spider tricksters. In M. Bekoff, C. Allen, & G. M. Burghardt (Eds.), The cognitive animal: Empirical and theoretical perspectives on animal cognition (pp. 27e34). Cambridge, MA: MIT Press. Wommack, J. C., Taravosh-Lahn, K., David, J. T., & Delville, Y. (2003). Repeated exposure to social stress alters the development of agonistic behavior in male golden hamsters. Hormones and Behavior, 43, 229e236.

Chapter 6

How the brain gets a roaring campfire: Structuring for perceptual results Erling O. Jorgensen Riverbend Community Mental Health, Concord, NH, United States

Introduction Picture a summer evening, on a warm vacation night, in a quiet campground. The campfire is nearly dying, down to a few embers now, still with a bit of glow for the couple seated on camp chairs nearby. It is too early to snuggle off into the tent, but some decaf coffee (with a touch of liqueur) would be, oh, so nice, before settling down for the night. Problem: ‘This is the wrong sort of campfire! We’re not going to get any boiling water out of this!’ There is a mismatch, between the “roaring” campfire of our imagination and the “mildly glowing” thing in front of us. Whenever a mismatch arises, action is generally taken, to make one’s perceptions more closely match what is wanted. If only we could alter a set of relationships e “more” wood, of the right “size,” placed “upon” the existing embers e we might get the fire (and the coffee) we desire. Wood does not move by itself, so it will be necessary to enact a series of movements, in the right sequence, to bring about the behavioral event of “stoking.” Movements do not just float in the air. They involve manipulations of tangible objects, to bring about the right sensations of “crackling yellow” flames, that will signal us the boiling water is not far off. Every intention that is not currently being met can be deconstructed down into a series of sub-perceptions that must be brought to their own preferred values, for the higher level perception to be enacted. What seems to happen so simply is really quite a complicated process for the brain to undertake. A series of questions may give some idea of the scope of problem. How does the brain know what is happening out there in the world? How can it change what it sees happening? If the brain wants something different, what form should that

The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00006-X Copyright © 2020 Elsevier Inc. All rights reserved.

101

102 SECTION | B Models of brain and behavior

desired goal take? How much needs to be known about the means of implementation? Must every potential consequence of action be foreseen, so the right ones can be chosen? Or is it enough to monitor the direction of change for any given action, like some extensive Hot-or-Cold game, where “Getting warmer” or “Getting cooler” is sufficient to home in on the desired target? To ask these questions about the brain is to ask about the neurophysiology. There is a frightfully wondrous nervous system carrying out these processes. However, sampling even a small portion of research into the brain can seem like wandering into a dense overgrown thicket. The neural twists and tangles go in every which direction, and it can be hard to make even some initial sense of all the complexity. That is where maps can serve an important function. This chapter will look first at some neocortical mapping of brain function, before broadening in the following chapter (with additional supplementary materials - see online version) to include some of the subcortical structures and how they function. To be more specific, Chapter 6 on “Structuring for Perceptual Results” draws upon two maps in particular. One is a broad functional map with roots in engineering and cybernetics, known as Perceptual Control Theory (PCT). This map has rigorous and compelling proof-of-principle models as to the application of negative feedback for bringing perceptual input to its desired states, and the implications that follow from that key insight. Where it is weak is in neurophysiological implementations of a range of perceptual input functions. It specifies the types of perceptions the brain may be operating with, but not many of the potential physiological mechanisms. This is where the other map may have something to offer. The second map, known as Hierarchical Temporal Memory (HTM), provides a more detailed rendering of how neurons, specialized in layers and columns in the neocortex, may be enacting various functions. Specifications of this theory are available both in the popular literature (Hawkins & Blakeslee, 2009) and with all the detail of a doctoral dissertation (George, 2008), along with other derivative papers in the literature (e.g., George & Hawkins, 2005, 2009; Hawkins, George, & Niemasik, 2009). There are also commercial and scientific ventures to further develop the technology of HTM (Numenta, 2011). As an application of Bateson’s (1979) method of double description, this chapter will attempt to overlay these two maps e PCT and HTM e upon one another, to see what increments of understanding emerge. Both sets of maps are broadly integrative, in that they suggest what functions to be looking for amidst the neural tangles. In this respect, the PCT map may have a more extensive scope, because it seeks to outline a very broad range of human functioning by means of a relatively small set of core concepts. To the extent that instantiations of those concepts can be discerned in the physiology of the nervous system, such an approach may yield fruitful avenues and insights as to

How the brain gets a roaring campfire Chapter | 6

103

the underlying terrain. The HTM map is more confined to the neocortex, deriving common processes and algorithms that may be utilized regardless of sensory modality (Numenta, 2011). The following Chapter 7, focusing on “Input and Output Functions”, draws in additional research or ‘maps’ of both subcortical and neocortical structures of the brain. One set of maps is derived from the work of Vadim Glezer (1995), a Russian investigator of the brain’s visual system. He offers a detailed exploration of visual transfer functions at numerous spots in the nervous system, from the retina through the lateral geniculate nucleus of the thalamus, through the striate and peristriate cortices, and on up to the temporal and parietal cortices. His work suggests a more extensive list of how perceptual input functions get constructed than that proposed for PCT. Additional maps supplement that presentation, in terms of more circumscribed models of visual input functions. In addition, the matter of PCT output functions will be broached, by way of focused output through the basal ganglia (see online version of Chapter 7) en route to motor areas, using the analogy of various gating functions operative in that region (Chevalier & Deniau, 1990; Groenewegen, 2003; McFarland & Haber, 2002; Mink, 2003; Yin, 2014). Further online materials supplementary to Chapter 7 will present an extensive discussion of modulated input through the thalamus (Guillery, 1995; Sherman, 1996, 2007), including a novel understanding of how PCT comparator functions may get implemented there, thus becoming output structures as well. Such areas of the brain appear to be structured in key ways, which may allow for important features of PCT-style negative feedback control to be enacted.

Features of perceptual control How to structure a goal These chapters take as their key starting point the work of William T. Powers (1973, 1978, 2008), and his proposals regarding Perceptual Control Theory. The heart of the theory is that organisms stabilize multiple variables of importance to them, by measuring the variables at the point of perceptual input. In other words, organisms can only know about their environment by means of what they sense. If they can then control for favorable states of those perceptions, they can achieve a wide variety of important goals e everything from physiological requirements for optimal functioning, to useful spatial arrangements in the objects around them, to high level systems of affiliation or social checks and balances. A key concept is that organisms do not have goals for action but goals for results. And the best way to gauge results is to measure them perceptually. Corrective negative feedback thus becomes a way to progressively move toward the attainment of important goals. A non-technical description of some basic PCT concepts is presented in Fig. 6.1. All the

104 SECTION | B Models of brain and behavior

FIG. 6.1 Perceptual control loop. Modular structure of a basic PCT control loop, regardless of level or scale at which it operates. Non-technical descriptors show its basic functioning. It can be operationalized in terms of a few simultaneous differential equations (see Powers, 2008, p. 82).

functions and signals of a control loop have precise definitions in the literature of PCT (Powers, 2008). What is useful about Fig. 6.1 is to realize that those same precise concepts have common-sense correlates in non-technical language. The basic components of a negative feedback control loop, as conceived within PCT, comprise a fairly simple modular structure. Externally-derived input called a Perception is compared to an internally-generated Reference signal (called a “Goal” in Fig. 6.1) specifying the preferred state of that perception. The difference, calculated by a Comparator, comprises an Error signal (or “Discrepancy”), which drives the Output (or “Behavior”) of the control loop. Typically, an integrating function is used to generate the output, amplified by a gain factor and graded by a slowing factor. The import of such a function is to start to move action in the needed direction. At the lowest level, the output has behavioral effects in the external environment, the most important of which is to influence a Controlled Variable (called “What you care about” in Fig. 6.1), which may also be affected by various Disturbances in the environment. When control is good, effects of the output compensate for and typically cancel the effects of the disturbance on the controlled variable. The state of the controlled variable is continually checked out and measured in terms of appropriate sensors by the organism, to generate various types of perceptions,

How the brain gets a roaring campfire Chapter | 6

105

thereby closing the loop. Because such an arrangement stabilizes and controls1 the values of those perceptions, PCT emphasizes and indeed takes as its name, Perceptual Control Theory. The elemental control loop described above utilizes a proportional-integral (PI) form of what is called a “PID controller” (Wikipedia, 2014, PID controller) in its output function.2 This promotes stable control from a limited set of computations. Dynamic stability can be further increased by stacking the elemental control modules depicted in Fig. 6.1 and utilizing cascaded PI-control, where the output of a higher controller provides the set point reference for a lower controller. This principle is laid out clearly in the earliest published works that later developed into PCT. Powers, Clark, and McFarland (1960) express it this way: “To put it graphically, the output of a second-order system is not a muscular force, but a goal toward which first-order systems automatically adjust their input-signals” (p. 79). This hierarchical arrangement is fundamental to PCT conceptualizations. What is essential here is for the lower level to operate on a faster time scale than the higher level.3 The time scale can be built into the equations by means of the slowing factor mentioned above. The result of this structural arrangement is for the lower level to handle and filter out fast-changing dynamics, while the higher level operates at a slower pace with the stabilized perceptual results the lower level provides.4 The control task is thereby partitioned into different perceptual components, improving the dynamic range of overall control. Such a modular array of control loops within a hierarchy is in effect a functional template that can be applied to a wide variety of biological architectures, at various degrees of scale. This allows it to have a broad level of generality, utilizing a small toolbox of concepts. As alluded to above, the aim

1. Elaboration on this cybernetic concept of control is included in the online version of this chapter. 2. In response to the question, what becomes of the derivative portion, PCT offers three answers. (1) Cascade control allows lower level controllers to filter out high frequency changes in the process variable, so that higher level controllers can operate with a more limited bandwidth for error. (2) The rate of change in the process variable is itself subject to direct control as its own hierarchical level, known as “transitions.” Moreover, using smoothly graded set point ramping reduces excessive derivative movement. (3) The human emotion system may itself be based on the error derivative component. That is to say, the slope of change in the error term may activate systemic parameter changes, which are experienced as transitory emotional states. A PCT prediction or hypothesis in this regard might be that the amygdala or portions of the limbic system would be structured as sensors for derivatives of error, with input coming as collaterals from areas of the brain identified as PCT comparators. 3. Stability of the resulting cascaded control is discussed further in the online version of this chapter. 4. There is developmental evidence for the relative speeds of different orders of control systems, based on ethological studies of free-ranging neonatal chimpanzees (Plooij, 1990; Plooij & van de Rijt-Plooij, 1990).

106 SECTION | B Models of brain and behavior

of these chapters is to apply this conceptual toolbox to various neurophysiological findings, and to explore the nature of perceptual input functions that the brain may be constructing.

Organizational rationale A few words should be said about the organization of these chapters, and the rationale for choices as to material reviewed here. PCT in general is an exercise in modeling,5 specifically trying to model how living control systems operate. The central fact that PCT points to is that living systems control the state of their perceptions (Powers, 2008). This is the environment within which they operate. What is encouraging about Hawkins and George and their HTM model6 is that they seem attuned to this matter that the really hard part is to build the internal perceptual models. Likewise, Glezer, reviewed in the next chapter, is a careful modeler. He, too, demonstrates an amazing sensitivity to the richness of how visual perceptions seem to be constructed in the nervous system. Within the PCT frame, almost by definition, perceptions are what matter to a control system. Certainly this introduces a computational problem7 for those who would model all the components of control systems. But that problem is preeminently on the perceptual side of the loop, in the series of framing decisions that set up cascaded perceptions in the first place.8 Within a fuller

5. Construction of a Proportional-Integral output function with a slowing factor is discussed in more detail in the online version of this chapter, including references to timing studies, with their implications for relative hierarchical placement of levels of control. 6. There is an extended discussion of the HTM model’s reliance on Bayesian inference and prediction in the online version of this chapter. In brief, Bayesian inference is a computational way of improving the accuracy of interpretations of subsequent data, based on what has come before. It accomplishes this by characterizing data in terms of probabilistic functions, and then adjusting their conditional probabilities given the receipt of updating evidence. That evidence can be temporally prior, in a bottom-up feedforward manner, or contextually relevant in a top-down hierarchical manner. This approach has a natural affinity for computational models based on the notion of prediction. It is contrasted with the PCT notion of a goal, which serves not as a prediction, but a specification. By measuring desired goals perceptually, PCT allows the environment to model itself, without the necessity for prediction. 7. An argument for PCT’s initial set of equations is presented in the online version of this chapter, with placeholders for the neurophysiological transfer functions which will vary level by level. 8. An important concept in PCT (beyond the scope of this chapter) is what has become known as E. coli reorganization (Marken & Powers, 1989), for generating or altering control functions or parameters. Based on Koshland’s (1980) work on behavioral chemotaxis, Powers (2008) describes this process as “a mechanism that would cause organization to keep changing as long as the result was unfavorable. A favorable result then would let the process of change slow down or stop” (p. 109, emphasis in original). Selective retention following random trial and error may thus serve as a baseline default method for creating control system properties, even if it is sometimes supplemented by guided development.

How the brain gets a roaring campfire Chapter | 6

107

neural model envisioned by PCT, constructed perceptions cascade upwards e what neurophysiological research tends to call feedforward inputs e and references9 for needed perceptions cascade downwards, specifying the implementing perceptual goals at each lower level of the neural hierarchy. The loop is then closed via the forces and actions that are performed in the environment, with their resulting perceptual effects. As an overview of the presentation to come, the following is a succinct summary of the case. Essentially two questions e one descriptive, the other predictive e are being raised in these companion chapters: (1) Are the principles of PCT being enacted at a neurophysiological level in the brain? (2) Are there neurophysiological implementations of PCT components waiting to be found? This chapter is organized around the recurring campfire illustration, which is woven throughout numerous spots in the text. An early form of structure here is presenting PCT in terms of its predictions. These may be laid out as follows: (a) There is a clear modular structure for how the brain functions. (b) It can be captured in terms of a hierarchical arrangement of cascaded control. (c) This offers a robust integrative framework for understanding brain research. (d) PCT has a number of hypotheses about the types of perception to be found, as well as their relative placement. (e) There are specific predictions about what is needed for reference signals to function in a commensurate way with their respective perceptions. In order to explore these predictions, an approach that is both wide-ranging and computational is needed. The model known as Hierarchical Temporal Memory is well suited in this regard. A number of specific arguments can be summoned, in favor of bringing in the HTM model: It is an integrative map of sufficient scope to be compared to PCT. It provides potential physiological mechanisms. It proposes neural algorithms for cortical micro-circuits. It tackles the difficult problem of perceptual framing. It matches a prediction of PCT, to implement commensurate framing of reference signals. (f) It utilizes a hierarchical arrangement.

(a) (b) (c) (d) (e)

9. Powers’ notion of reference signals as associative address signals is introduced in the online version of this chapter, in the context of potential overlap with a feature of the HTM model, allowing for commensurate communication between levels in the cortex.

108 SECTION | B Models of brain and behavior

(g) (h) (i) (j) (k) (l)

It It It It It It

has some recognition of feedback relationships. emphasizes the importance of time. incorporates memory and context into the process of sensing. offers an implementation of the PCT imagination connection. offers one possible mechanism for PCT gain adjustment. makes sequence data central for many types of perceptions.

Despite the broad compatibility between the PCT and HTM frameworks, several challenges can be posed for the HTM model. These would include the following: (a) Does PCT control provide a further de facto “supervisor” (in addition to time), for constructing and stabilizing perceptual niches? (b) Can HTM equations be used to model a fuller range of perceptions? (c) Can the form of PCT control equations indeed be integrated with HTM algorithms? (d) Can HTM implement all the mechanisms that may be needed for PCT references? These are the issues to be addressed, in the remainder of this chapter.

Modeling with falsifiability in mind It is worth noting that the implementation of a PCT control loop is a ‘toy model’, in the sense used within physics. It lays out the essential relationships in terms of a working model, with a minimum number of variables and the functional equations that connect them (Powers, 2008). A model in this sense is more than just a rearranged description of the emergent behavior it is trying to emulate. Rather, it specifies inner properties among the subparts of a system, which when allowed to run mathematically on its own generates its own approximation of the behavior under examination.10 That makes its predictions specific enough to be open to falsifiability. This is a key reason why the current chapter utilizes the HTM model, especially as specified via George’s (2008) algorithms, as a base of comparison with the PCT model. While both are admittedly described verbally in the body of the text here, leaving the actual equations to the original source documents, both models have actual working implementations underneath, going beyond the level of verbal abstraction and generalization. Indeed, if the comparisons raised in this chapter are at all justified, that raises the possibility that either model might be further refined in terms of parameters articulated in the other model.

10. Implications of the mathematical specification of PCT working models, even in simplified form, are presented in the online version of this chapter.

How the brain gets a roaring campfire Chapter | 6

109

By way of initial proof-of-principle demonstrations, the core PCT model is exceedingly robust.11 Because of the central role of negative feedback in a PCT control loop, its paradigmatic expression is that of a tracking task. When examples of tracking tasks have been operationalized with PCT equations, tuned with reasonable parameters and quantified via computer sampling of the data at 60 times per second, the point-by-point match between a subject’s performance and that predicted by the model has been on the order of 0.98 (Marken, 1992; Powers, 2008), even with different disturbance tables or on multiple runs widely separated in time. This is a strong challenge for a behavioral model to meet, especially by comparison with the flow-diagram verbal ‘models’ often reported within the psychological literature. In addition to the mathematical rigor, there is a heuristic value to viewing a wide range of biological phenomena as instances of tracking. Whenever a perception is made to match a fixed reference signal, overriding the effects of disturbances pulling the perception in other directions, that is an example of compensatory tracking. And whenever a perception is made to follow a changing reference signal, with or without the difficulty from disturbances, that is an example of pursuit tracking. In either instance, a given state of affairs is brought to a preferred state and maintained there, as measured by perceptual monitoring. The actions are somewhat incidental, as long as they serve to make the perception match the reference. It could be argued that any situation where a “purpose” can be discerned might well fall under this conceptual paradigm (Marken, 1992, 2002, 2014). That suggests there could be a wide applicability to this notion of tracking and the PCT model that explains it so well (Powers, 1978, 1992). For instance, in the campfire illustration at the beginning of this chapter, the campers seemed set on getting some liqueur-flavored decaf coffee before heading off into their tent. That projected state of affairs became a fixed reference signal against which the result of various implementing actions could be measured. They proceeded to enact that goal with a sequence of semi-flexible behaviors, each one geared to its own interim perceptual results. So then, the position of additional wood needed to change, via smoothly changing reference signals, to bring first kindling and then larger pieces of firewood close enough to existing embers for them to match a reference for catching fire. That step needed to happen first, before the mess-kit pot of cold water would be positioned, so there was some kind of reference signal for serial ordering of events. Once the water started to boil e another match of perceptual result to desired reference state e it became necessary to retrieve the pot while monitoring heat sensations, relative to a reference of not-getting-burned-in-the-process. 11. Examples of such demos can be found at http://www.livingcontrolsystems.com/demos/tutor_ pct.html and at http://www.mindreadings.com/demos.htm, with links to further PCT literature at http://www.pctweb.org/index.html.

110 SECTION | B Models of brain and behavior

These various actions were taking place via a variety of unspecified-in-thisexample yet controlled joint angles, and eye-hand coordination, and perceptual assessment of what type of fire would result in what sensation of coffee to cap off a just-right evening together. Every step in the process is essentially a tracking task for making a perceptual result come out the right way. Fig. 6.2 presents a schematic rendering of the kind of layering of perceptions and perceptual results involved in this campfire example. Now it must be said, Fig. 6.2 is just a visual and verbal portrayal of the range of possible perceptions, far from a “testable model” in the sense used above. Each link would require its own test, with disturbance-resistance being the hallmark for whether a given perception was being controlled.12 To move down the

FIG. 6.2 Hierarchical layering of perceptions in PCT. Schematic rendering of how basic control loops could be linked hierarchically into cascaded control loops. Higher levels set references for perceptions to be controlled at lower levels, which then help to comprise the perceptions being controlled at higher levels. To move up the hierarchy is to ask why a given perception is important, while moving down the hierarchy is to ask how the perception comes about. Eleven distinctive categories or classes of perception are represented by the layered sheets, according to ideas proposed and refined by Powers (1973, 1979, 1990) over the years. Representative examples are suggested for the different levels, based on the recurring campfire illustration within this chapter.

12. Among computer simulations of control, this concept of disturbance-resistance has been operationalized in terms of a “stability factor,” where the actual variance (in the denominator of the term) due to environmental disturbance is much less than otherwise expected, when the quantity is being controlled (Powers, 1978; Marken, 1983, 1989).

How the brain gets a roaring campfire Chapter | 6

111

hierarchy of control is to ask how each step is being implemented. To move up the hierarchy of control is to ask why it is important, that is, from what frame of reference and to what purpose.

Predictive proposals of PCT Hierarchical classes of perception With regard to the notion of a hierarchical super-structure of inter-nested control loops, Powers (1973, 1979, 1990) provides specific ideas about various classes of perception, including the types that might need to be constructed in order for a higher level of perception to be constructed. This aspect of Powers’ theory has been termed Hierarchical Perceptual Control Theory, and is admittedly more speculative than the quantitative rigor of the core model. Nonetheless, even though the specific names and arrangement of perceptual layers (as listed in Fig. 6.2) are not essential elements of PCT per se, they do offer a fruitful way to categorize and partition the various perceptions of one’s experience. A brief description of those proposed levels follows, working up from the bottom, together with some indication of the basic organizational principle that might be involved with some of them.13 The first and lowest order of perception could be called Intensity, basically representing “how much” stimulation of sensory nerve endings is occurring. At the campfire scene, the degree of “glowing” among the embers constituted an intensity perception. The second order of perception could be called Sensation, made up of a “weighted sum” combination of various intensity signals. The “yellow” flames of the campfire, and the “crackling” sound, would both be examples of sensation perceptions. The third order of perception could be Configuration, occurring at the next higher level, with an organizing principle of “feature co-occurrence.” Distinguishing objects that can be moved for “firewood” is a campground example of a configuration perception, along with the limb and joint configurations for moving them. A fourth order of perception is that of Transition, involving the “rate of change” of a lower level perception. The desired “flickering” of the campfire flames would be an example of a transition perception. A fifth order of perception could be called Event, which is not just a sense of movement but a repetitive movement functioning as a unitary package. Steadily “stoking” the fire and “placing” each piece of firewood into position represent two instances, of slightly different durations, that comprise behavioral event perceptions. A sixth order of perception could be called Relationship, consisting of “co-variation” among lower level perceptions that otherwise might vary

13. The online version of this chapter presents a fuller explication and expansion of these ideas, as does the introductory chapter to this volume.

112 SECTION | B Models of brain and behavior

independently of one another. Gathering “lots” of kindling, and positioning it “near” the embers of a campfire, would be examples of these kinds of relationship perceptions. A seventh order of perception, called Category, is a form of constructed perception consisting of “class membership.” Within the campground, the image of a “roaring” campfire, occurring at several neighboring sites, as compared to the “sputtering” one at this site, were essentially category perceptions, because any number of different styles of fire might be classed as one or the other type. An eighth order of perception is a level called Sequence, referring to “serial ordering.” At the campsite, an extended series of results was being prepared, that of deriving “a-bigger-fire-for-getting-boiling-waterfor-coffee,” the attainment of which required getting the perceptions to occur in the right sequence. Any other position for the coffee in that series would mean drinking it cold. A ninth order of perception could be called Program control, enacted as a set of decision nodes, each comprised of an if-then test. A basic decision node in the campfire scene is, “if no bubbling water, more heat.” This would form part of a program perception, likely affecting the initial planning for building up the fire, as well as gauging when to remove the mess-kit pot from near the flames. Movement through the network is conditional on the perceptual state existing in the world as each node is reached. A 10th order of perception could be called Principle, in the sense of “general guiding heuristics” for selecting among programmatic algorithms. The campers within the illustration were likely operating with a somewhat fuzzy notion of having a “nice” evening, as one of their principle perceptions. An 11th order of perception could be called System Concept, in the sense of a systematic unity among principles, procedures, and relationships. At the campground, it seems the reason for all the other implementing perceptions was for the campers to “enjoy this time together around a campfire,” as an episodic system concept perception subserving the broader narrative system concept of “enriching our marriage.” These diverse classes of perception imply quite different perceptual input functions occurring in the nervous system of the brain. This captures some of PCT’s version of a frame problem (Jorgensen, 2005) e not that of inferring behavioral consequences, but of constructing perceptual inputs. And the myriad of ways of constructing those perceptions is where the more daunting challenge lies for research; namely, making sense of the brain’s complexities surrounding perceptual framing. To reiterate, Chapters 6 and 7 survey two key approaches to the specific matter of perceptual framing. The Hierarchical Temporal Memory model of Hawkins and Blakeslee (2004) and George (2008), examined in this chapter, proposes canonical cortical circuitry for how the brain functions. In Chapter 7, Glezer’s (1995) work on visual input

How the brain gets a roaring campfire Chapter | 6

113

functions, especially the role of spatial frequencies and what is called the ventral stream from the striate cortex to the inferotemporal cortex, is explored at length.

How to maintain commensurate matching There is a companion framing difficulty e similar to perceptual framing e that comes along for the ride: How does the error signal deriving from a mismatch for one form of perception e a higher one according to a PCT hierarchy e become a specified reference signal for quite a different form of perception at the next lower level? This is no trivial issue. It seems that Powers (1973) was well aware of the potential dilemma, because he mentions at one point, almost in passing: “This, incidentally, solves the problem of translating an error in a higher-order variable into a specific value of a lowerorder perception, a problem that has quietly been lurking in the background” (p. 217). What problem is he referring to, and what solution is he proposing? To give it a name, it is the issue of incommensurability, the question of comparing or measuring two things according to a common standard. It is the problem of interface. How will two things communicate in a way that is understandable to both sides? What language or units or frames do they hold in common? This issue is glossed over in the toy-model simulations generated so far to test the principles of PCT. Basically commensurability between the different levels is presupposed. It is handled by the computer. The equations assign numbers to the different variables, and the computations proceed straightaway. This is a defensible simplification, based as it is on an assumption that variables within control loops perform as scalar quantities, even if they actually represent more complex functions. Powers (1973) lays out that scalar assumption early in his book when talking about the concept of “neural current, defined as the number of impulses passing through a cross section of all parallel redundant fibers in a given bundle per unit time” (p. 22, italics omitted). However, by returning much later to this ‘quietly lurking problem in the background,’ Powers (1973) demonstrates awareness that just dealing in numbers does not guarantee that they are meaningful numbers, for any given equation. So he introduces a key postulate, as a first approximation: “We will assume from now on that all reference signals are retrieved recordings of past perceptual signals” (p. 217, italics omitted). This is a first step in handling incommensurability. If the pattern of a current perceptual signal is being compared to the pattern of a past perceptual signal at the same point in the brain where that signal first arose, then at least apples are being compared to apples. He goes on to draw the implication of that postulate: “This requires giving the outputs from higher-order systems the function of address signals,

114 SECTION | B Models of brain and behavior

whereas formerly they were reference signals. The address signals select from lower-order memory those past values of perceptual signals that are to be recreated in present time” (p. 217).

Getting a workable address This is a very insightful move. Reference signals are address signals. It is like saying, “Give me more of that.” The reference doesn’t even need to know what “that” signifies. The “address” here is a way of ensuring the reference signal is getting to the right locale, and by taking the form of a retrieved perceptual signal, both the reference and the perception are guaranteed to be speaking a comparable “language.” There is a further step that Powers (1973) takes, when speaking of associative addressing: “In associative addressing, the information sent to the computing device’s address input is not a location number, but a fragment of what is recorded in one or more locations in memory” (p. 212). He continues, “Any part of the stored information can be used as an address, a match resulting in replay of the rest of the information in that memory unit” (p. 213). This results in a pointer, or an address, that is already contextually relevant. In other words, “Give me the rest of that.” A side benefit of using associative addressing is that the signal carrying that fragment can be sent to many different locations. If no match is found in many of the locales e in other words, if that sequence of inputs does not generate an expanded replay of the larger pattern stored in a given area e that top-down hierarchical input just becomes random noise for those non-relevant areas. Either that or it remains sub-threshold, because of not being amplified by the synaptic strengthening established by the original perceptual copy. There is no need for a higher level to know ahead of time the exact locale required. This is a decided advantage, allowing unsupervised addressing of the requisite reference signals. So then, a reference signal has the following features: (a) It only has to be an address signal, essentially signaling more of “that.” (b) It is a recorded copy of what it is looking for, leading to a common currency for comparison between perception and reference. (c) It only needs to carry a fragment of what it is looking for, because it will result in the replay of “the rest” of what it is looking for. This is quite an amazing solution set for keeping signals commensurate between different levels of control and perception in the brain. By giving reference signals these features back in 1973, Powers was making implicit predictions about the brain’s wiring and its functional characteristics.

Cortical implementation Perceptual Control Theory makes a series of broad predictions about how an internal model of reality may be built up via a hierarchy of cascaded perceptions, with higher level ones constructed out of the raw material of lower

How the brain gets a roaring campfire Chapter | 6

115

level perceptions. It provides guidance e hypotheses, one might say e about the types of perceptions to be found in the human nervous system, and about their relative placement within a hierarchy. PCT also predicts what may be needed for reference signals to function in a coherent way, to bring about stable negative feedback control. Simulations based upon PCT models of control have worked exceedingly well, to a very high scientific standard (e.g., Powers, 2008). It remains to be seen whether the principles of PCT are being enacted at the neurophysiological level of the human nervous system. As was mentioned at the outset of this chapter, the complexity of neural research raises a need for broadly integrative maps of this territory, which can be compared with that provided by PCT. Among the most promising ones is work by Jeff Hawkins and his colleagues coming out of the Redwood Neuroscience Institute. Hawkins and Blakeslee (2004) list three crucial features for understanding how the brain functions, which are quite prominent for PCT as well: namely, including time in the functions, the significance of feedback, and the neocortex “organized as a repeating hierarchy” (p. 25). In addition, HTM and PCT both share an appreciation for specifying working models that can be tested. While the full modular architecture of PCT would be enacted throughout the nervous system, it can be useful to begin with cortical implementation. Subcortical implementations will play a more substantial role in the following chapter and its supplementary material (see online version). The emphasis of the HTM model on prediction bears a family resemblance to the PCT notion of controlling and thereby repeating prior perceptual results. It appears that Hawkins and his colleagues have articulated arrangements of cells in the neocortex that may be able to produce some of the features laid out under Powers’ PCT proposals. Hawkins and Blakeslee (2004) present a plausible, neurophysiological way to understand intelligence, and how actual brains may construct and carry out that process. The core of that theory is what was initially called “the memoryprediction model” (p. 5), and later called “Hierarchical Temporal Memory (HTM)” (George, 2008, p. 6; see also; Hawkins et al., 2009); this latter designation is used here to refer to this model. Building on Hawkins’ foundation, Dileep George (2008) in his doctoral dissertation derives neural algorithms for a possible mathematical model of the microcircuits at work in the neocortex. A more detailed examination of the import of some of George’s proposals will be presented below.

PCT reference signals and HTM “name” cells As suggested above, there appears to be a good deal of conceptual overlap between how prediction functions in Hawkins’ model and how control functions in Powers’ model. For instance, Hawkins and Blakeslee (2004) give

116 SECTION | B Models of brain and behavior

an analogy for what a higher region is saying to a lower region when it sends a prediction down the hierarchy: “This is basically an instruction to you about what to look for in your own input stream” (p. 136). That sounds almost like providing a reference for a specified perception, even though for PCT it is changed through external action, not just internal interpretation. Hawkins and Blakeslee come closer when they describe the significance of motor commands: “As strange as it sounds, when your own behavior is involved, your predictions not only precede sensations, they determine sensation” (p. 158). Whether Hawkins and his colleagues fully realize it or not, to speak of determining sensation is to start to speak of control. This is especially so from the standpoint of a PCT understanding of motor commands, which essentially are commands for certain perceptual consequences of motor behavior. The part that appears to relate most closely to Powers’ idea of reference signals as memory address signals is Hawkins and Blakeslee’s (2004) notion of cortical patterns of firing which relay the ‘name’ of a given sequence or co-occurrence of inputs. It is a complicated discussion, but one worth looking at in some detail. The physiological heart of the HTM model is the neocortex, in particular its almost modular arrangement into cortical columns. These columns of cells comprise the basic functional units for constructing invariances and sequential patterns out of the neural firings from the flow of sensory experience. This flow is constantly changing, and the job of the HTM units is to build a spatial and temporal model of that experience in a selflearning manner. “The basic idea is that the patterns that frequently occur together in time share a common cause and can be grouped together” (Hawkins et al., 2009, p. 1203). By means of this feature of temporal cooccurrence, time is the only teacher or supervisor in the HTM model. In conjunction with the columnar integration of cells in the neocortex, there is also a regular laminar structure, with six discernible layers of cells (although it seems layers 2 and 3 are very similar). It appears that different layers may perform different functions within the cortical columns. So then, the main input layer seems to be layer 4 for connections from lower regions, whether thalamic or hierarchically lower regions of the cortex. Calvin (1996) calls layer 4 “the IN box of neocortex,” with a corresponding term of “OUT box” used for the deeper layers 5 and 6 (p. 30). Because of synapses from layer 4 to other layers within the column, it seems “the entire column becomes active when driven from below” (Hawkins & Blakeslee, 2004, p. 148). Layer 6 seems to be the main output layer for downward-projecting connections, which synapse (a) onto layer 1 axons in hierarchically lower cortical regions, and (b) onto thalamic areas that contributed the original bottom-up projection to that cortical area (Guillery, 1995; Sherman, 2006). The significance of this latter downward projection is discussed further in online materials supplementary to Chapter 7. Hawkins and Blakeslee (2004) make a fascinating speculation about an incidental detail with regard to connections involving these two input and

How the brain gets a roaring campfire Chapter | 6

117

output layers of the cortex: “(I)n addition to projecting to lower cortical regions, layer 6 cells can send their output back into layer 4 cells of their own column. When they do, our predictions become the input. This is what we do when daydreaming or thinking” (p. 156). This comes remarkably close to the arrangement Powers (1973) predicted and called the “imagination connection” (p. 221-Fig. 15.3), and it suggests a physiological locale for where and how that feature of intelligence may be taking place. According to the HTM model, the ‘name’ is a key way that columns are communicating vertically, both up and down the hierarchy. This is what George and Hawkins (2005) in a later article call “communicating.in terms of the indices of the high probability sequences” (p. 1813). Hawkins and Blakeslee (2004) give a more intuitive view of what is happening: “All objects in your world are composed of subobjects that occur consistently together; that is the very definition of an object. When we assign a name to something, we do so because a set of features consistently travels together” (p. 126). They continue, speaking from the viewpoint of a cortical column, “So whenever I see any of these events, I will refer to them by a common name. It is this group name, not the individual patterns, that I will pass on to higher regions of the cortex” (p. 129). If this name constitutes an address fragment that will subsequently be used as a reference signal from above, it is worth noting that the fragment appears to originate from below. That may assist even more with commensurate recognition, if a higher level sends a diffuse, unsupervised address signal back down through a tangle of synapses.

Neocortical layers in action The import of this truncated form of vertical communication is described as follows, starting with going up the hierarchy. “The layer 4 cells in every column receive input fibers from several regions below it and will fire if they have the right combination of inputs. When a layer 4 cell fires, it is ‘voting’ that the input fits its label” (Hawkins & Blakeslee, 2004, p. 147). Its synapses get the whole column active, and further projections continue up the hierarchy. Furthermore, there is lateral inhibition of nearby columns, to further shape and refine the name, so that higher regions do not get a jumble of possible names from below. Going down the hierarchy, a layer 6 cell does the communicating, saying in effect, “I speak for my region of cortex, .my job is to tell the lower regions of cortex what we think is happening. I represent our interpretation of the world” (Hawkins & Blakeslee, 2004, p. 154f.). This supplies context for the lower level, which from a PCT perspective means instruction on how to subsequently perceive something. So the PCT message from a layer 6 cell may be stronger than that for HTM: not only ‘this is my interpretation,’ but also ‘keep that constituting perception coming.’ It appears that layer 1 is a receiving layer, for inputs from higher regions of the cortex and for collateral inputs (conveyed via the thalamus) from other

118 SECTION | B Models of brain and behavior

active columns within the same region. Hawkins and Blakeslee (2004) suggest the functional significance of these two types of connections through layer 1: “We can think of these inputs to layer 1 as the name of a song (input from above) and where we are in a song (delayed activity from active columns in the same region)” (p. 146). This sounds quite similar to what was posed above about a reference signal being a contextually relevant pointer. That descending output is projected via the axons of layer 1 to cells of layer 2 at the next lower cortical level, which “learn to be driven purely from the hierarchically higher regions of cortex.. The layer 2 cells would therefore represent the constant name pattern from the higher region” (Hawkins & Blakeslee, 2004, p. 153). This picture comes incredibly close to seeing a reference signal in action, as constrained by the features that Powers (1973) laid out. The overlap between Hawkins’ model and that of Powers is quite striking. As Hawkins and Blakeslee (2004) summarize: “Thus the information in layer 1 represents both the name of a sequence and the last item in the sequence. In this way, a particular column can be shared among many different sequences without getting confused. Columns learn to fire in the right context and in the correct order” (p. 149). In other words, there is commensurate communication, which is not getting confused as to where and how it is talking. It uses the right name, in the right context, using a common currency to avoid confusion. This neurophysiological approach of Hawkins’ team to how the cortex works is quite compatible with the broad functional contours that Powers (1973) laid out, some three decades earlier. Although Hawkins’ institute has obviously surveyed a wide swatch of theory and research into the brain, it is not clear from the literature sampled for this chapter that Hawkins is necessarily familiar with Powers’ approach. If there is no direct familiarity, the potential overlap and/or opportunity for ferment of ideas between PCT and HTM would be all the more remarkable.

Modeling the temporal flow of experience How to build contextualized beliefs The above descriptions of HTM draw mostly from a work in the popular literature (Hawkins & Blakeslee, 2004), which necessarily conveys merely a verbal summation of their work. A verbal treatment is not the same as a generative model. The compatibility between PCT and HTM appears to continue even when examining the more detailed algorithmic work of Dileep George (2008), who has provided possible mathematical equations for HTM neuron-like elements in the cortex. Several layers of dynamical equations are spelled out, as a way of modeling the different cellular layers within cortical micro-columns. An outline of those functional layers follows, using the metaphor of various ways to chunk the neural inputs. Fig. 6.3 offers a rudimentary sense of the perceptual functions that might be encoded.

How the brain gets a roaring campfire Chapter | 6

119

FIG. 6.3 Proposed functional connections among cortical layers. Schematic diagram of types of neurons by layer in the neocortex, along with their proposed functional connections for constructing different types of perception. This conceptualization follows HTM proposals put forth by George (2008) and Hawkins and Blakeslee (2004). See text of the chapter for specific explanations of the perceptual chunking occurring layer by layer.

It is important to realize that Fig. 6.3 is a highly schematized rendering of the types of relationships involved between different layers of cortical cells. For purposes of illustration, the disparate examples included in the figure have been artificially conflated as if they occur within the same micro-column. This is a convenient fiction, imposed on much more complex patterns of connection with dynamically changing patterns of firing. Not only do individual cortical neurons have thousands or tens of thousands of synapses, but discernible perceptual features are often the result of networks of cells rather than single ones. However, some form of perceptual and behavioral order nonetheless emerges from such complicated interactions. As a step in discerning that emergent order, a simplified template is here presented, combining insights from both the HTM and PCT models. The discussion begins with a concise description of Fig. 6.3. The job of the first Chunk is to detect co-occurrences. With the first Chunk, stellate cells of cortical layer 4 encode temporal co-occurrence among inputs from lower levels. For example, the sound “b” might co-occur with peripheral sensations of the lips pressed together then pushed apart. The job of the second Chunk is to notice temporal sequences. With the second Chunk, pyramidal cells of layers 2 and 3 (labeled Sequence cells in the figure) receive inputs from the layer 4 stellate cells and have lateral

120 SECTION | B Models of brain and behavior

connections with layer 2 or 3 pyramidal cells in adjoining columns, encoding items happening in temporal sequence. For instance, one sequence cell might encode a “b” perception, while adjoining columns might encode “e” or “g” perceptions. The job of the third Chunk is to compile into a name. With the third Chunk, a different type of pyramidal cell in layer 2 (labeled ‘Name’ cell in the figure) compiles inputs from pyramidal sequence cells from nearby columns. For example, the sequence “b þ e þ g” may be encoded by a particular pattern of firing. The ‘Name’ cell then sends axon projections both to layer 5 cells and ultimately to higher cortical levels. Descending inputs from higher levels, conveyed via layer 1, serve as an address indicator, which can sustain firing in the whole column. The job of the fourth Chunk is to pool local and hierarchical evidence for what is transpiring. With the fourth Chunk, pyramidal cells in layer 5 (labeled Belief cell in the figure) pool contextual inputs from higher cortical levels, conveyed via layer 1, together with the local input within the column from the layer 2 ‘Name’ cell. The result is to encode sequences that are reliable collections across different situations. For instance, the higher level context may indicate that “b-e-g” is a reliable occurrence of use in a variety of perceptual events. A Belief cell projects to a layer 6 pyramidal output cell, as well as sending projections to bursting pyramidal cells in adjoining columns that construct different perceptions out of that reliable sequence. The job of the fifth Chunk is to provide a timing gate, so that a timed perception can be sent to higher cortical levels. This standardizes the input as temporal co-occurrences for that next higher level. With the fifth Chunk, intrinsic-bursting pyramidal cells in layer 5 receive input from a Belief cell about a reliable sequential collection, and send a projection to a non-specific thalamic nucleus, which returns a timing signal via the axonal projections in layer 1. This turns the sequence into a timed event. For instance, the layer 5 bursting cell in one column might represent the word “beg,” while an adjoining column might incorporate that same sequence into the sung ditty “abcdefg”, while still another column might utilize that sequence in the sequentially timed word “begonia.” These timed events are projected to higher order thalamic relay nuclei, to serve as constructed perceptions driving the firing of thalamic relay cells there, and projected from there to the next higher levels of the cortex. Finally, the job of the sixth Chunk is to output the interpretation downwards. With the sixth Chunk, pyramidal cells in layer 6 (labeled Output cell in the figure) receive input from layer 5 Belief cells across several columns about reliable collections of perceptions, sending that output as reference signals back to the lower order thalamic relay area via the thalamic reticular network. The thalamic connections will be explained in the online materials supplementary to Chapter 7. Table 6.1 summarizes these functional chunks, and compares their HTM significance to their PCT significance.

TABLE 6.1 Summary of HTM and PCT functional chunks. Type of cell

HTM significance

PCT significance

First chunk

L4

Stellate cells

Detect co-occurrences: Inbox layer for cortical micro-columns, capturing temporally co-incident inputs

Co-occurrence as initial basis for configuration perceptions

Second chunk

L2 & 3

Pyramidal Sequence cells

Notice lateral sequences: Encoding temporal sequence patterns from below, among layers 2 & 3 in adjoining microcolumns

Constructing perceptions of transitions or sequences

Third chunk

L2

Pyramidal Name cells

Compile into a name: Communicating hierarchically to & from higher cortical areas, providing the “name” of a given sequence of firing

Reference address signals for auto-associative memory

Fourth chunk

L5

Pyramidal Belief cells

Pool the evidence: Pooling top-down & bottom-up evidence, providing a degree-of-certainty appraisal & confirming the reliability of certain sequences of coincidence patterns

Adjusting a gain parameter

Fifth chunk

L5

Intrinsicbursting Pyramidal cells

Provide a timing gate: Providing timing/duration signals for when beliefs from the fourth chunk would be operative; standardizing by temporality for next higher level

Re-indexing ascending perceptions according to time; substrate for event perceptions (i.e., timed sequences)

Sixth chunk

L6

Pyramidal Descending Output cells

Output the interpretation: Pooling evidence of reliable beliefs across several columns & projecting it downward as the best interpretation from this portion of cortex

Reference standards for preferred perceptions sent down to hierarchically lower areas

Presentation and significance of the HTM functional chunks portrayed in Fig. 6.3, including cortical layer and type of cell, as they relate to PCT conceptualizations.

121

Cortical layer

How the brain gets a roaring campfire Chapter | 6

Functional chunks

122 SECTION | B Models of brain and behavior

While the above examples suggest fairly complex perceptual features, it is worth noting that early published tests of the HTM model involve such aspects as recognition of simplified visual patterns (George & Hawkins, 2005) or stick figure representations of motion capture data (Hawkins et al., 2009). These are no small feats for a biologically plausible mathematical model to achieve. Nonetheless, it is obvious that these algorithms are just a start in discerning how cortical columns may be functioning. George (2008) provides a useful formulaic statement for the HTM model: “Intuitively, a learned HTM node can be thought of as having memorized a set of patterns (the coincidence-patterns) and a set of sequences over patterns (the Markov chains)” (p. 107). Those patterns and sequences are analyzed in relation to the patterns and sequences of cortical areas hierarchically above and below a given node. He gives an example for visual area V1 in the striate cortex, where vertical edge patterns may occur in conjunction with certain horizontal edge patterns, with different Markov chains representing the sequence of movements of each of those edges. Here is a simplified way to understand the concept of Markov chains. The brain has to make sense of a temporal stream of input. A given bit of input can either change or stay the same. It can also occur together with other bits, i.e., coinciding with them, or next to other bits, i.e., sequentially either before or after. These are the basic elements that the brain has to work with. A Markov chain is a way to model some of the order-creating activity of the brain. A firstorder Markov chain only deals with immediately adjoining events – what changes into what. Certain regularities in those changes can be captured in terms of “transition probabilities,” i.e., the likelihood of what comes next. In other words, the temporal structure, as interpreted by the brain, changes from random to subsequent events that are more or less likely. Higher-order Markov chains can model the likelihoods of sequences that are several steps in length. When there is a greater likelihood of a subsequent event, that information can depolarize or bias the firing of an adjoining neuron that is helping to recognize a pattern. In this way, networks of pattern recognizers can be built up. To use the campfire illustration, one could think of the vertically oriented flames and their shifting movements, together with the somewhat horizontal lines of firewood underneath, with their more subtle variation of shadows or partially-burned portions, or additional flames dancing out from the underside. While the precise sequencing of the visual representation of flames is not memorized, per se, the ascending and side-to-side regularities of the light images can nonetheless become part of a stabilized higher-level comprehension of “a campfire,” and even contribute to appraisals of what kind of campfire it is, “roaring” or otherwise. Those higher-level interpretations (which were themselves learned by experience) become contextual beliefs for the lower-level modules, as to whether these flickering sensations are proceeding as they should. They also become “predictions,” if you will,

How the brain gets a roaring campfire Chapter | 6

123

(PCT might prefer the term “instructions” or “preferences”), for what will happen with adjustments of firewood, or placement of a pot in the midst of those shifting visual sensations.

First chunk: detect co-occurrences The perceptual complexity of such examples must be faced. While this chapter mostly deals with how the cortex may be constructing its perceptions, it is important to realize that the retina and thalamus have already been constructing perceptual regularities out of the stream of visual input. A rich example such as a campfire underscores that the bottom-up input coming into each HTM node already has a temporal form e that is the ‘T’ component in HTM e and it is the job of the node to make some sense of it. The first place that ascending input arrives is at layer 4 of the cortical column, and George (2008) models the stellate neurons of that layer essentially as coincidence detectors, where “the synapses of the layer-4 neurons represent co-occurrence patterns on its inputs” (p. 127). It is temporal co-occurrence among the message components from below e time is the only supervisor, after all e so it is modeled in terms of multiplication of ascending Markov chain components (p. 112). A certain clarification is needed at this point. George, as well as Hawkins whose work he expands upon, both envision the cortical column module as a template for how the neocortex may be organized in whatever area it occurs. Therefore, ascending and descending projections involve not just subcortical connections, but also cortical regions that are hierarchically above or below the module of interest e this is the ‘H’ component in HTM. A neuro-anatomic fine-point preserved in Fig. 6.3 is that exiting axonal projections seem to be routed past layer 6, even if they are headed to layer 1 of a hierarchically higher level of the cortex. Such projections join a white-matter region of horizontal myelinated axons just below cortical layer 6. It is worth noting that feature co-occurrence may be a key organizing principle for Powers’ (1973) conception of “Configuration” perceptions, although that might involve the overlapping of a wide variety of features of different forms of abstraction. It is of interest that, similar to the low-level position where Powers places configurations within his hierarchy of perceptions, George (2008) begins his mathematical model with equations that calculate the likelihood of coincidence patterns arriving from below.

Second chunk: notice lateral sequences George (2008) makes the evident point that “the same coincidence pattern can be part of different Markov chains” (p. 115), displaying sequential regularities in more than one context. This feature is conveyed by moving from the first Chunk to the second Chunk in Fig. 6.3, with the obvious example that the

124 SECTION | B Models of brain and behavior

perception “b” can be part of multiple sequential regularities. In the HTM model, these are set up by layer 2 and 3 pyramidal sequence cells, seeing as layer 4 projects mainly to that combination of layers in the cortex. A credible way to capture the essential temporality of a stream of input is as sequences that take on probability characteristics in terms of their adjoining components. That is to say, how reliably does a given datum precede or follow another datum? And Markov chains provide a way to model those characteristics. George (2008) assigns this role to some of the pyramidal neurons in layers 2 or 3 of the cortical column, noting, “They become active only in the context of the correct sequence” (p. 130). This seems to be accomplished via excitatory lateral connections with layer 2 or 3 neurons within other cortical columns, where each item in the chain provides meaning and temporal context for the next item in sequence. The equations that capture this dynamic reflect synaptic weighting of the lateral inputs, multiplied by the coincidence input from below (p. 116). This places the coincidence patterns in the context of the sequential patterns, with degree of certainty reflected via a probability function. A case could be made that Powers’ (1973) description of “Transition” perceptions may get constructed via these mechanisms in the cortex, seeing as that type of perception often manifests itself as a sense of movement in whatever sensory modality is operative. If so, it is worth noting that the transition level is the next higher level, after configurations, within Powers’ hierarchical model, similar to its placement within George’s model. Nonetheless, when the focus is upon the exact serial ordering, rather than on just noting the fact of change itself, sequence would seem to be the proper word for that type of perception. This consideration suggests that Powers’ (1979) moving of the “Sequence” level itself to a higher spot in his hierarchy may not fully square with where and how that form of perception may actually get constructed within the neocortex.

Third chunk: compile into a “name” Another role that George (2008) discerns for a different class of layer 2 or 3 pyramidal cells is that of compiling a representation of each reliable Markov chain sequence, for projecting to higher levels in the cortex. This consists of a cell or cells receiving projections across cortical columns, to “pool the outputs of all the coincidence-pattern neurons in the context of a Markov chain” (p. 117). In Fig. 6.3, this process is the third Chunk, represented by what is called a pyramidal ‘name’ cell. Because it takes the form of a probability function, according to George’s equations, this amounts to a feed-forward message about its degree of certainty as to the reliable occurrence of its sequence. As Hawkins and Blakeslee (2004) describe it, somewhat more evocatively, “each cortical region has a name for each sequence it knows. This ‘name’ is a group of cells whose collective firing represents the set of objects in the

How the brain gets a roaring campfire Chapter | 6

125

sequence” (p. 129). That collective firing could be detected as co-occurrence by the L4 stellate cells at the next higher level. Here is another instance of collating or chunking a broader array of neural input into a more compact form for communicating with other hierarchical levels. As was proposed earlier from a PCT standpoint, once this name pattern is formed, it can then be utilized as an associative memory address signal, for conveying a PCT reference signal back from a higher level of perception and control.14 There are pyramidal cells in layers 2 and 3 with apical dendrites in the receiving layer 1, which appear to have the requisite neurophysiological connections to make this a possibility. Moreover, according to Hawkins and Blakeslee (2004), it seems through synaptic strengthening that cells in layers 2, 3, and 5 can learn to keep firing, even when their initiating inputs are no longer active. This conforms with features of associative addressing, raised earlier, where replaying a fragment, i.e., the ‘name,’ leads to replaying the rest of the content to which the name refers.

Fourth chunk: pool the evidence George (2008) models predictions as descending belief vectors derived from conditional probability functions. Within that Bayesian schema, there needs to be a locale for integrating the bottom-up and top-down inputs. This role is assigned to layer 5 pyramidal neurons, which receive local projections from neurons in layers 2 and 3 within the same column, as well as more global or contextual projections from higher levels. This is the fourth Chunk in Fig. 6.3, represented by what is called a pyramidal belief cell. George notes that “the operation of these belief neurons corresponds to the pooling of evidence for a particular coincidence from the different sequences that this coincidence participates in” (p. 131). This type of construction would have to rely on information from higher up, because a given coincidence pattern as it ascends would not ‘know’ to connect with distant sequence sets in other cortical columns. So there is temporal evidence from below multiplied by contextual evidence from above. The effect of that pooling of evidence, including the contextualization provided from above, is to confirm the reliability of certain of those coincidence patterns, precisely because they occur across an array of sequences. In Fig. 6.3, this is conveyed schematically by having a layer 5 pyramidal belief cell represent the reliable sequence of “b” followed by “e” followed by “g”. Within George’s (2008) equations, the top-down evidence appears as a multiplication factor, which thus adjusts the gain and thereby the degree of certainty of what he terms the belief neurons. Gain is an essential parameter 14. This step creates conditions for genuine feedback control. Mere top-down communication by itself is not the same as cybernetic “feedback,” despite that term often being employed in that manner (George, 2008; Numenta, 2011).

126 SECTION | B Models of brain and behavior

within PCT models, comprising the proportional component of the PIcontroller (e.g., Powers, 2008, p. 82).15 However, gain adjustment has yet to be modeled in a physiological manner in PCT, so its appearance within George’s HTM equations is of interest. Degree of certainty amounts to belief in, or reliability of, coincident sequence patterns. That is the nature of the invariant being constructed at this juncture within the neocortical layers, namely, lower level patterns whose usefulness has been tuned up so as to provide predictable elements for other types of perception. To call it an invariant here does not mean the value never changes, but only that it takes on an established structure. The nature of this belief invariant is that of a reliability rating attached to a sequence of coincident occurrences, with the likelihood of the bottom-up evidence being multiplied by a top-down likelihood appraisal (George & Hawkins, 2009). If the reliability component is modeled as a Bayesian multiplication factor, its value could vary between zero and one, signifying weak or strong effects of this coincident sequence wherever it projects. A high degree-of-certainty appraisal would in effect convey two meanings. One is a statement: (a) This pattern of coincidences in sequence reliably occurs. The second is a question, almost a prediction: (b) Can it occur elsewhere; that is, is it useful for other things? When coincidence patterns are first detected in bottom-up fashion, they are simple temporal co-occurrences. An example might be the visual transitions a rolling ball might make, as seen against the occluded transitions of its background. Here, with top-down contextualization provided, coincidences occurring in sequence start to generalize to a broader range of meta-sequences. A top-down example of reliable control might be curling one’s fingers relative to one another, in order to achieve a range of purposes, such as (1) eating with a fork, (2) writing with a pencil, (3) snapping one’s fingers, (4) gesturing to come, or (5) picking up firewood. This starts to implicate cortical regions such as the pre-motor or motor cortex, and it would seem that pyramidal neurons within layer 5 of such areas may provide the locale for the pooling of top-down and bottom-up inputs, to start to bring about such results.

Fifth chunk: provide a timing gate This discussion almost sounds like it is referring to what Powers (1990) calls “Event” perceptions e e.g., the eating, writing, snapping, gesturing, or picking up, listed above. However, before that full parallel can be drawn, an additional component is needed, supplied by what George (2008) calls “an external timing unit” (p. 119). This role he assigns to a different class of layer 5 pyramidal neuron, routed through cortico-thalamic and thalamo-cortical

15. The computer demos used in Powers’ (2008) book are also available at http://www.billpct.org/.

How the brain gets a roaring campfire Chapter | 6

127

connections, “representing the precise time at which the beliefs are going to be active” (p. 132). It appears that layer 5 in the cortex has regular-spiking neurons and intrinsically-bursting neurons, and it is the latter that provide a timing signal, according to the HTM model, via projections to non-specific areas of the thalamus which loop back to layer 1 of the cortex (Llinas, Lesnik, & Urbano, 2002). The reason this re-routed signal is significant is that it allows a layer 5 neuron, with apical dendrites in layer 1, to “act as a gate that opens only when a timing signal and a belief value are both available at its inputs” (George, 2008, p. 119). This process is the fifth Chunk in Fig. 6.3, represented via connections to and from intrinsic-bursting pyramidal cells. This timing feature may allow a different type of perception to be constructed e not just items arranged in sequence, but items timed in sequence. Examples are suggested by Fig. 6.3, albeit in an artificially simplified portrayal. For instance, the letters “b-e-g” occur in sequence among the first seven letters of the alphabet. But that is not the same as speaking them together with the right timing as the word-event, “beg,” or singing them in sequence with the childhood ditty, “a-b-c-d-e-f-g,” or inserting them in another word such as “begonia.” Similarly, the action of feeding wood onto a fire includes a sequence of closing the hand (around a piece of wood), moving and positioning it, and then opening the hand again, but if the opening does not occur with the right timing, one can easily get burnt. It would seem to be this issue of timing that distinguishes mere serial ordering, called a sequence, from sequences-timed-in-sequence, which could be called an event. There may be a further benefit from this aspect of perceptions gated according to time. Guillery (1995) speaks of higher order relay nuclei in the thalamus, which receive driving input not from peripheral sensors but from feedforward cortico-thalamic projections. Indeed, the connections come from layer 5 pyramidal cells, and the information is routed through the higher order thalamic nucleus to the layer 4 stellate cells of the next higher cortical area (Sherman & Guillery, 1998). The HTM description of how these cells work may serve the function of re-indexing the ascending perceptions according to time. This is a form of standardization needed at each cortical level, according to the HTM model. After all, if temporality is the main supervisor for forming meaningful co-occurrences at that next higher level, it would do no good for time to have been stripped out of the incoming data stream.

Sixth chunk: output the interpretation There is a further piece of George’s (2008) model, which is the operation of the layer 6 neurons. As discussed earlier, this layer forms the main downward projecting output layer for the cortical columns, and as such it is well situated to provide collated input to hierarchically-lower areas. This would include lower cortical areas, for instance the associative addressing of references

128 SECTION | B Models of brain and behavior

discussed as part of the third Chunk above. More centrally, it would also include a sign-inverted PCT reference projection through the thalamic reticular nucleus on a monosynaptic path to a thalamic relay nucleus (Guillery, 1995; Hill & Tononi, 2003; Ohara & Lieberman, 1985). The significance of this path, as argued in the online materials supplementary to Chapter 7, is for thalamic relay cells to serve as PCT comparators for references descending from the cortex. From the standpoint of HTM, George (2008) assigns to the layer 6 neurons the role of pooling the evidence from layer 5 ‘belief neurons,’ not only within the same column but from across several columns, whichever ones were implicated in the original ascending inputs involving coincidence patterns. In the HTM schema, the downward output contains that area’s overall degree of certainty as to how to best interpret the input received from below. It is the sixth Chunk in Fig. 6.3, represented there as pyramidal output. In a PCT schema, this would amount to conveying reference standards to the lower hierarchical areas, as to specific perceptions that would be preferred.

Compatible cortical mind-sets Constructing invariants: a summary In seeking to synthesize these two major theories, it is important to realize that HTM and PCT take very different approaches. Hierarchical Temporal Memory is an attempt to derive canonical computations that may be operative throughout the neocortex. It builds upon the highly regular columnar arrangement in the cortex to discern how the brain may be making sense of the temporal flow of input that the cortex receives. Perceptual Control Theory is a broader functional attempt to discern how the brain may be bringing about perceptual results that matter to it by means of cascaded control. Its feasibility test requires a completed circuit to see if stable control is being achieved. Closing that loop necessitates ascending and descending connections through subcortical structures, with the final link occurring through the environment itself. An immediately recognizable difference is that HTM works with the temporal input as it stands, and seeks to discern meaningful or at least regular patterns. PCT has the added benefit that it can help to create meaning, by taking actions that affect the state of its inputs. Its computational requirements are fairly minimal. Regardless of how a given perceptual input function is constructed, there simply needs to be a way to compare a perception to a reference standard and influence the perception to come closer to the reference. This means the sign of the output is primary; timing considerations are of secondary rank. If those functional components can be discerned among physiological structures in the brain, then control of perception is part of what

How the brain gets a roaring campfire Chapter | 6

129

is going on. A further PCT hypothesis is that the mismatch between perception and reference at a higher level sets the direction of needed change for perceptions at a lower level. The HTM theory limits itself to a temporal stream of input and whatever properties are implicit in that presentation of data. Four properties seem especially pertinent for the HTM theory: (a) Which inputs are happening together? (b) Which inputs are happening near to one another in a temporal sense? (c) What is the precise order of adjacent inputs? And (d) What is the precise timing of adjacent inputs? There is an implicit assumption for the HTM theory that the environment contains meaningful objects and events. However, the only real access the HTM brain has to those objects and events are the temporal regularities coming from constituent indicators of those environmental occurrences. This is very close to the implicit epistemology adopted by the PCT model. The four properties listed above are statistical degrees of freedom, which would be part of any temporal stream of input, and thus available to any neocortical level. There are both similarities and differences to the way PCT examines perceptual input. PCT considers that there may be many more qualitatively distinct ways to construct perceptions. In an earlier section of this chapter, following various proposals by Powers (1973, 1979, 1990), eleven classes of perception were briefly outlined, although at the lower levels of “intensity” and “sensation,” each sensory modality may actually have its own mechanisms for transducing environmental energy into neural currents and weighting the results accordingly. In terms of similarities to the HTM model, PCT has proposed types of perception that align with the four properties outlined above. For instance, the first property involving overlapping input has a parallel with the PCT category of “configurations,” based on co-occurring features. The second property involving nearby input has affinity for the PCT category of “transitions,” which can include the notion of the same input changing over time. The third property of the next immediate input aligns with the PCT category of “sequences,” dealing with the serial order in which inputs occur. And the fourth property of the next input according to an externally-supplied time index creates the substrate for the PCT category of “events,” where items are timed in sequence. These parallels are summarized in Table 6.2, along with examples drawn from the campfire scenario. The major challenge with synthesizing the respective approaches of PCT and HTM has to do with how hierarchical levels are conceptualized. In the PCT proposal, levels are virtually defined as different types of perception. Granted, this aspect of the theory is more descriptive of the phenomenology of human experience, and thus may serve simply as a first approximation for a more comprehensive model of cascaded control. Powers himself never insisted that his way of categorizing experience was the only way to do it. What is crucial, however, for a working model based on negative feedback control of

130 SECTION | B Models of brain and behavior

TABLE 6.2 Partitioning a temporal flow of input. Temporal property

Degree of freedom

Which inputs are happening together?

Overlapping input

Configurations, based on cooccurring features

Orientation of wood; patch of light & heat; sizzling sounds of boiling water;

Which inputs are happening near to one another in a temporal sense?

Nearby input

Transitions, including same input changing

Glowing embers; dwindling heat; flickering flames; ascending smoke;

What is the precise order of adjacent inputs?

Next immediate input

Sequences, based on serial order of inputs

Kindling, then sticks, then larger firewood; reach, grasp, move, release;

What is the precise timing of adjacent inputs?

Next input according to an externallysupplied time index

Events, based on items timed in sequence

Stoking campfire; curling fingers around wood; pouring boiling water;

PCT parallel

Campfire examples

Four properties implicit in a temporal stream of input are presented, which may be the basis for several distinct types of PCT perceptions. Examples from the campground scenario are suggested.

one’s perceptual input (if indeed it is structured as a hierarchy), is that any relatively higher level has to operate with a slower time constant. This is so it does not introduce instabilities into the lower level of functioning, by in a sense calling for results faster than they can be produced. In the HTM model, the same basic columnar circuitry is utilized to construct comparable input functions at whatever level of the cortex. What distinguishes hierarchical levels in this model is simply their degree of abstraction. As Hawkins and Blakeslee (2004) express it, “The input changes from representing mostly individual patterns to representing groups of patterns” (p. 165). Examples could be drawn from some of Glezer’s (1995) material, as a foretaste of Chapter 7. So then, co-occurrence of visual input at an early stage in the striate cortex might involve inputs from adjacent receptive fields in the thalamus with comparable degrees of contrast within the visual field. At higher levels of visual processing, say in the inferotemporal cortex (as discussed in the next chapter), co-occurrence may signify the superposition of

How the brain gets a roaring campfire Chapter | 6

131

spatial frequencies according to their harmonic components. Similarly, sequential input at a low cortical level might encode alternating on- and offcells from the lateral geniculate nucleus, whereas at a higher visual level the sequence may involve the shift of texture when moving one’s gaze from figure to ground. This gets to the heart of how to consider blending these two theories. The HTM model incorporates each of the four properties enumerated above into virtually every neocortical column. On the surface, that seems like a clash with the PCT approach. What brings these two conceptualizations together is to realize that a wide variety of perceptions can (a) co-occur, or (b) smoothly change in value, or (c) occur in serial order with other perceptions, or (d) require specific timing for their implementation and control. In other words, these four fundamental properties of a temporal stream of input can be applied to many different forms of perception. The situation is easiest to see with regard to sequences. Sequences can readily be observed with regard to a variety of higher level perceptions. For instance, there can be a sequence of categories, such as first kindling, then twigs, then larger pieces of firewood. There can be a sequence of relationships, such as bringing those types of wood near, then over, then upon the existing embers. There can be a sequence of timed behavioral events of first reaching, then grasping, then lifting, then moving, then releasing the pieces of wood. And there can be a sequence of body configurations and shifts of joint angles for arranging such firewood properly. What reconciles these perspectives is to realize that a sequence of other perceptions can be carried out at various speeds. So then, there can be a sequence of plans for a major project, which are carried out over a long period of time. What dictates relative placement within a hierarchy of perceptual levels would be the fastest speed at which sequences could be constructed as their own form of perception from their constituent components. Tracking simulations by Marken (2003) provide useful evidence for how fast such perceptions can be perceived and controlled, relative to other levels of perceptual control. These considerations support the instinct by developers of the HTM model that sequential regularities may well be part of a canonical circuit within the columnar arrangement of the neocortex, across many different cortical locales. So then, there are regular ways of constructing sequences laterally among adjoining columns, seemingly involving pyramidal neurons of layers 2 or 3, in many or most areas of the cortex. What is left indeterminate is the question, sequences of what? The type of sequence would be set by the ascending sensory input. For the striate cortex, that input might represent contrast changes of light and dark flickerings, which when placed in a sequential context become a sense of movement. In the peristriate cortex, the input might include orientation perceptions, where a sequence of similar orientations

132 SECTION | B Models of brain and behavior

would help to construct texture perceptions, while a sequence of different orientations might help to construct a sense of rotation. The logic is similar with regard to the other temporal regularities listed above. Powers (1973) specifically notes, “Nearly any perception of the first three orders, in fact, may apparently be detected at fourth order as change. The phenomenon I am calling here the ‘perception of transitions’ may explicitly involve rates of change of any perception of intensity, sensation, or configuration” (p. 131). By the same token, higher level perceptions can also involve change, as when the symmetrical gliding relationship between two figure skaters seamlessly flows into one supporting the other into a backward circular spin relationship. So then, transition as the perception of change itself can occur at various hierarchical levels, depending on the type of perception that is changing. Placement as its own level in a hierarchy would depend on the fastest speed at which changes could be constructed and detected. A similar situation can be seen with regard to what Powers (1990) calls the perception of events: “Even when we speak of an event which takes hours to be completed, at this level we make a single package of it. .an opera performance is an event, as is the first act, as is the aria which finishes it, as is the trill at the end of a passage” (p. 72). Despite the continual flow of experience, the perceiver creates temporal boundaries, giving an impression of more discrete events. As these examples show, the perceptions enclosed within such boundaries can be of a diverse nature. The campfire scenario offers additional illustrations of this level of perception. “Warming up” via a campfire is a series of timed perceptions, involving getting close to the heat source, monitoring the rate of heat transfer so as to not get too hot, keeping away from changing billows of smoke. Similarly, “making coffee” from the water boiling in a pot placed in a fire entails getting the timing right in a variety of ways, including perceptual indicators as to temperature, as well as a range of movements for holding and pouring carefully into cups, etc. These are perceptual events, comprised of sequences being controlled as to their timing. A passing consideration for a PCT model is that the ordering property necessary for sequences is distinct from the timing property necessary for events, and thus these two types of perception may comprise different levels in a PCT hierarchy.

Addressing reference signals Within the operation of the HTM model, invariant patterns are created at each level. As regularities are detected (PCT would say, constructed), it is important for the invariance-construction going on at higher levels that they are not swamped with all the ever-changing incidental details of the lower level patterns. So then, stoking a campfire is mostly about expanding the flame itself, not the minutiae of smoke or sound. The import of this consideration is that communication between levels happens in terms of a ‘name’ or index

How the brain gets a roaring campfire Chapter | 6

133

representing a given invariance. In George’s (2008) implementation of the HTM model, this is accomplished through a distinct class of layer 2 neurons, which essentially pool the outputs from each reliably occurring Markov chain sequence. It is this columnar name e one for each reliable sequence e that is sent to the next higher level. For an example involving kindling, a neural summation representing sequential inputs “catching fire” might be the kind of name transmitted to higher perceptual levels. It is at this point in the HTM model that a striking parallel occurs with some key features of Powers’ (1973) PCT model. Powers confronts a potential incommensurability problem between perceptions and references, by specifying certain conditions or features of reference signals in his model. In order to communicate in the right language and at the right locale as the perception being controlled through a given control loop, Powers proposes that references function as address signals for auto-associative memory. In other words, they only need to carry a fragment of a recorded copy of a given perception e what was described earlier as essentially saying, “Give me the rest of that” e in order to achieve commensurate communication at that key spot within the control loop. These are features that can be supplied in the HTM model by George’s (2008) pooling function in cortical layer 2, or what Hawkins and Blakeslee (2004) more colloquially call passing on the name of a predictable sequence. This name becomes the fragmentary address for a higher level to use as it sends it back down the hierarchy, when it wants a repetition of the perception(s) encoded within that cortical column. Repeating the name summons the perception, which sets the terms e whether by way of reference (PCT) or interpretation (HTM) e for even lower hierarchical levels. So then, “catching fire,” with all its attendant details, can be the shorthand goal sought by a higher level, whether it refers to a bunch of kindling, chunks of firewood, or the campfire more generally. The HTM model provides structural space, not just for PCT references functioning as address signals, but also for one form of gain parameter. There is a spot in the HTM model for tuning up the reliability and usefulness of the invariants being constructed. This is accomplished by determining the degree of belief that a higher level of the cortex supplies to a lower level, about the patterns being generated from below. “An HTM node calculates its degree of belief in a coincidence pattern by combining bottom-up, top-down and temporal evidences” (George, 2008, p. 117), into contextualized belief vectors. In the HTM presentation, these are modeled as Bayesian conditional probability equations, with layer 5 pyramidal neurons pooling that contextual input. This pooled evidence seems to correspond to adjusting gain, presumed to occur at every hierarchical level. For a higher level to supply this context becomes a statement as to reliability, thereby adjusting what seems to be perceptual gain for a PCT control loop implementation. High reliability is akin to raising the gain or amplifying

134 SECTION | B Models of brain and behavior

the salience of a given perception. It becomes a leverage point for shaping a perception’s relevance, or putting it to greater use. As a corollary, reliable patterns get utilized (or from an HTM perspective would be predicted to occur) across many settings. To continue with a campground illustration, wood catching on fire is a predictable and reliable occurrence, which is part of a range of useful goals, such as getting more warmth, enjoying a brighter fire, boiling water for coffee, or distinguishing adjoining campsites. Reliability in the face of disturbances e such as current weather conditions, or the firewood’s availability, dampness, or size e is one way to speak of the gain of the control systems involved. Reliable components appear to be necessary for shuffling sequences into what Powers in his later writings (1990) calls event perceptions. But for that full construction to occur, the sequences must happen according to the correct timing. In the HTM model, this timing mechanism is proposed to occur through intrinsically-bursting layer 5 pyramidal cells, which send projections to non-specific regions of the thalamus, where the signals get looped back through the receiving layer 1 of the cortical column. The co-occurrence of this timing signal with the belief signal being pooled in layer 5 serves as a gate for both upward and downward projecting signals. From a PCT perspective, this is the construction of timed sequences, which are necessary for generating and controlling behavioral events. There is further import to this provision of timing signals for when the HTM beliefs are to be operative. By means of this feature, ascending perceptions are re-indexed according to time. Recall that in the HTM conceptualization, the only supervisor giving direction to the construction of perceptual regularities is that of time. Perceptions that co-occur temporally have a greater chance of having mutual significance. So the time indexing that occurs via layer 5 cells of the cortical column may be an important step before those perceptions are relayed via higher-order thalamic nuclei to higher levels of the cortex. A final step in George’s (2008) instantiation of the HTM model is the collating of evidence from layer 5 belief neurons across numerous columns, and summating them via layer 6 neurons, for projection to hierarchically lower levels. Those levels would likely include both cortical and thalamic areas. This is essentially generating an overall interpretation from a small area of the cortex. From a PCT perspective, this step would constitute reference signals being conveyed down the hierarchy, for preferred values of perceptions at those lower levels. Fig. 6.4 portrays in schematic form how Powers’ hierarchical PCT proposals may be blended with the HTM schema of George and Hawkins. The diagram seeks to present the basic arrangement of cascaded control, suggesting how HTM nodes at different cortical layers may contribute to constructing perceptions in the brain. The abbreviation key in Table 6.3 helps to spell out the implications of various signals and connections.

How the brain gets a roaring campfire Chapter | 6

135

FIG. 6.4 Schematic synchronization of HTM and PCT hierarchical connections. Functional connections among the cortex, thalamus, and basal ganglia-striatum, are arrayed in terms of cascaded control. This arrangement is meant to synthesize cortical functions proposed by the Hierarchical Temporal Memory model with the type of hierarchical context that is so central to Perceptual Control Theory. Most of the thalamic connections derive from research reviewed for the online materials supplementary to Chapter 7, and they are presented here to assist with visualizing emerging portions of PCT control loops. Table 6.3 accompanies this figure, to explain the abbreviations used here.

Chapter 7 will expand upon these conceptualizations with a broader survey of research into the brain’s input and output functions. Specifically, it will draw upon Glezer’s (1995) presentation of how visual perceptions may get constructed, and examine parallels to the hierarchical mapping of Powers’ PCT proposals. It will also discuss functional implications of how the basal ganglia are structured (see online version of Chapter 7). Additional online materials supplementary to Chapter 7 will investigate subcortical areas of the brain, particularly the functioning of relay and reticular areas of the thalamus. Taken together, these companion materials will try to spell out how input and output functions may operate within control loops from a PCT standpoint.

Status of the brain’s campfire So how might an HTM cortical network of inputs experience a campfire? Obviously this chapter cannot do full justice to the complexity. In that sense the campfire has not been “modeled” but only “illustrated” by the concepts

136 SECTION | B Models of brain and behavior

TABLE 6.3 Key to abbreviations used in Fig. 6.4. Abbreviation

Meaning

Per

Perception, with index number for its relative level hierarchically

Ref

Reference, with index number for its relative level hierarchically

L2

Cortical layer 2

L2 & 3

Cortical layers 2 & 3

L4

Cortical layer 4

L5

Cortical layer 5

L6

Cortical layer 6

Gating

Gating of thalamic relay output according to burst vs. tonic modes of firing

Imag.

Proposed PCT imagination connection, as a cycling of layer 6 outputs back up as layer 4 inputs

Co-incid.

Co-incidence of inputs, constructed & perceived at cortical layer 4

Sequence

Temporal sequences & transitions, constructed & perceived at cortical layers 2 & 3

Name

A pattern of firing at cortical layer 2, compiled from sequence inputs, & used to communicate with a higher hierarchical level

Address

A reference indicator from a higher level, utilizing the name pattern of firing of layer 2 to access auto-associative memory in the lower-level column

Gain

A top-down signal conveying degree of reliability of a constructed perception

Belief

A degree-of-certainty appraisal constructed at cortical layer 5 from collated bottom-up & top-down inputs

Timing signals

Projections routed through non-specific thalamus & back to cortical layer 5, providing ways to sequence perceptions according to relative timing

Time indexing

The synchronization of ascending perceptions from cortical layer 5 with relative timing signals, so that temporal co-occurrence can serve as a supervisor as higher-level perceptions are constructed

Non-spec. thal.

Non-specific thalamic region, conveying a timing signal back to the cortex

Thal. retic.

Thalamic reticular formation, with circled plus & minus signs representing sign inversion from input to output

Thal. relay

Thalamic relay nucleus, with index letter indicating its relative hierarchical order, serving as a PCT comparator

How the brain gets a roaring campfire Chapter | 6

137

TABLE 6.3 Key to abbreviations used in Fig. 6.4.dcont’d Abbreviation

Meaning

BG-striatum

Basal ganglia striatum region, considered here the beginning of the output function for PCT control loops

(Arrows)

Multiple input arrows into the same node signify additional input from beyond the local column

Brief explanations are offered to clarify the abbreviations utilized in Fig. 6.4, along with their significance for a blending of HTM concepts with PCT-style cascaded control. The function and significance of the Thalamic reticular formation, Thalamic relay nucleus, and Basal ganglia striatum will be explained in online materials supplementary to Chapter 7.

presented here. Nonetheless, the functional chunking of perceptions summarized in Table 6.1 and Figs. 6.3 and 6.4 can provide some guidance, and a further brief review and summary is provided here. Representative examples offered in this section are grouped by scene in the campground, according to (a) the pre-flare up phase of the campfire, (b) the phase of controlled stoking, and (c) goals during the post-flare up phase. Recall that the issue for the first Chunk of co-occurrence is detecting temporal contiguity. During the first phase at the campground, the coincidence of orange light and charred wood might be noted, along with the fact that wood orientations at that point in the campfire would have been largely horizontal. Since that first locale deals with layer 4 cells in the cortex, we can consider that a layer 6 output routed back to that layer might enact what in PCT is called the imagination connection. In the campfire scene, admittedly shifting here from visual cortex to somatosensory cortex for this example, an imagined co-occurrence might include an implicit question about deriving sensations of hot water together with coffee and liqueur flavors. The second Chunk via cortical layers 2 and 3 seems to enact both transitions and sequences. Instances of the former would include the pulsating orange colors of the glowing embers, as well as the gradually dwindling sensations of heat coming from the fire. Perceptions occurring in sequence might include the alternating flame and shadow sensations of the fire, and the way the charred wood becomes progressively more horizontal as it burns down. The naming occurring via the third Chunk at layer 2 consists of a composite form of signaling, such as an overall sense of the campfire winding down, which could serve as an interface between hierarchical levels in the cortex by way of associative addressing. Such address signals could allow higher level references to summon the requisite columns, to bring about a change of input from the campfire. The fourth Chunk pools the ascending and descending evidence into a belief vector at cortical layer 5, which amounts to a judgment about reliability and salience of the unfolding perceptions. During that pre-flare up phase, this

138 SECTION | B Models of brain and behavior

may have been a determination about the implications of getting less heat from the fire, i.e., not enough to produce that imagined coffee. The fifth Chunk, also occurring at layer 5, seems to perform two functions. It adds a time index to the ensuing perception, which may help to supervise the construction of co-occurring invariants at the next higher level, as well as allowing for timed sequences such as are needed for coordinated movements going down the hierarchy. Time indexing during this phase of the campfire may have served as a pivot point, as the brain notices a new action sequence is needed and starts to identify the first step on that path. Specifically timed event sequences may have been ones for finding the proper firewood, and conducting a visual and tactile search. The sixth Chunk creates the overall interpretation from that local part of the cortex. This would be important for constructing reference signals to be projected to lower hierarchical levels for getting the right perceptions in place according to the higher level goals. At this phase in the scene, such layer 6 cells may have been part of generating a sequential program for obtaining the desired coffee. The brain would continue to register the temporal flow of experience, even as in the next phase its actions of stoking the campfire would help to bring about a different state of affairs. So then, subsequent layer 4 co-occurrences might notice wood being placed in various orientations, not just horizontal, while receptive fields adjacent to one another might signal an expanding patch of light, and if there were an associative overlap between visual and somatosensory cortex, the light might co-occur with greater heat. Transition perceptions in layers 2 and 3 would monitor the changing heat gradient against the skin, as well as an increase in sharply flickering flames. Constructing specific sequences in serial order might include a relationship sequence, such as “more” wood brought “close” to the fire and placed “upon” the embers. It may also involve a sequence of categories, first “kindling,” then “sticks,” then “larger pieces of firewood.” The composite name generated at cortical layer 2 and communicated to a higher level might simply be a fragmentary signal representing “stoking,” conveying the message that the sequence is underway. Pooling hierarchical evidence for reliable beliefs at layer 5 might convey a useful pattern of fingers being curled, whether they were around a piece of firewood or (later) the handle of a pot of water. The provision of a time index for other layer 5 cells would help to specify the timing needed for determining the next steps based on the current results. The descending behavioral cascade of timed and coordinated movements for arranging firewood would include: reach, grasp, move, release, along with their associated arm and joint accelerations and decelerations. Finally, the interpretive output from cortical layer 6 would be for the purpose of monitoring and guiding the progress of associated subroutines.

How the brain gets a roaring campfire Chapter | 6

139

In the final, post-flare up phase of dealing with this campfire, temporal co-occurrences at layer 4 in various cortical areas might involve the mess-kit pot actually in the flames, while steam and the sizzling sounds of water against the pot would start to gather. Layers 2 and 3 might encode transition indicators of boiling water, and vertically directed side-to-side alterations of light, as well as sequences such as “smoke,” then “flame,” then “heated water.” The fragmentary layer 2 name for this process might convey to higher levels that a bigger campfire is underway. Combining experiential evidence from below with contextual memory from above at layer 5 might verify that this is indeed a “roaring” campfire, and its salience may be that it is now “hot-enough-to-boil-water-for-coffee.” Checking off certain key conditions become the time-indexed decision points for the extended program underway. Various layer 5 perceptions would also be sequenced into timed behavioral events such as pouring boiling water carefully. The layer 6 output would keep providing the reference standards guiding these successive approximations toward “a nice evening around a campfire.” The ‘nice-ness’ of this campfire notwithstanding, it is certainly not a simple process that the brain is undertaking!

Remaining challenges An important challenge for the cortically-oriented theory of Hierarchical Temporal Memory is whether it can model and reproduce a fuller range of perceptions such as those proposed by Perceptual Control Theory. While certain low level classes of perception may get constructed at a more peripheral level than the cortex, there remain many qualitatively distinct ways of constructing perceptions e so argues PCT e presumably happening at neocortical levels. The presentation in this chapter suggests significant parallels with HTM for the substrates of what PCT calls configuration perceptions, as well as transitions, sequences, and events. That still leaves about seven classes of perception, according to hierarchical PCT, for which no clear parallel has been determined from the research examined for this chapter. The next chapter will explore several additional levels of perception, where physiological correlates are striking. The broader question is whether the PCT schema of hierarchical cascaded control can truly be integrated with the HTM algorithms. The HTM model largely deals with the perceptual input side of how PCT control loops would be arranged. Thus, HTM computations and simulations have not fully been tested as to whether they can implement perceptions being stabilized via PCT-style negative feedback. Some of the top-down descending connections are proposed in the HTM model. But without completing the loop through output functions that are ultimately closed through the environment, it is not known whether the Bayesian equations of HTM stabilize in the same way as PCT would suggest for its perceptions.

140 SECTION | B Models of brain and behavior

This issue of inserting perceptual algorithms into an actual closed loop, as per the PCT model, raises an opportunity that could be exploited within the HTM model. Recall that as it currently stands, time is the only supervisor for how HTM perceptions get constructed. The fact that certain inputs co-occur in temporal proximity is considered potentially meaningful, and given structural form within the HTM model as sequential patterns of firing. That temporal character is allowed to shape the construction of HTM perceptions in cortical columns at ever higher levels, as more abstract patterns of sequences get created and recognized. What PCT may provide in this respect is an additional supervisor for the construction of perceptual niches. PCT control loops act to stabilize their perceptual inputs at values specified by higher level reference standards. Whenever perceptions are stabilized in reliable ways, this becomes a controllable degree of freedom that a higher level could exploit to construct a different type of perceptual invariant. Perceptions being reliably controlled at lower levels become well-worn paths, which may be called perceptual affordances, for higher levels to utilize in constructing their own reliable paths. In addition to comparison with hierarchical PCT categories, a prominent form of overlap with HTM theory has been the similarity between how Powers (1973) speaks of references and how Hawkins and Blakeslee (2004) speak of prediction. Many of the features posed by Powers in the 1970s seem to overlap with cortical features investigated by Hawkins and his colleagues, and these have been detailed at some length in this chapter. If the neocortex not only predicts what it is sensing but actually affects it through negative feedback control, then functions contributing to stabilized control must be there, in some localized or distributed fashion, although obviously input and output structures would involve subcortical regions as well. A significant compatibility test therefore, as mentioned above, would be whether the lean form of PCT control equations could be integrated with HTM algorithms, such as the ones provided by George (2008). A deeper question is that of enacting PCT equations within the broader realm of what is known of neurophysiological properties and mechanisms. Powers (1973, 2008) provided a strong start in that direction. An important step for future development would be to test that proposition within the mainstream of neurophysiological research. A key test case, so argues this chapter, would be whether the proposals Powers (1973) made for references to function as associative address signals could be made to work by the brain’s actual wiring at multiple neural sites. The following is a preliminary list of what would seem to be requirements for this aspect of Powers’ theory. (1) A compression mechanism that can create a de facto address from a fragment of the initial input. It appears that sparse coding could achieve this feature (Olshausen & Field, 2004), as could the property described as “sub-sampling” (Numenta, 2011, p. 30) within an HTM framework.

How the brain gets a roaring campfire Chapter | 6

141

(2) A memory mechanism that can ask for something “again.” The HTM approach holds promise in this regard, because once patterns have been learned, they can then be reused or called forth in other instances. (3) An addressing mechanism that can summon a memory at the right locale. This need not be a supervised process, as the concept of associative addressing makes clear. According to further developments of the HTM model, it appears that sparse distributed representation could provide this feature, seeing as “you only need to match a portion of the pattern to be confident that the match is significant” (Numenta, 2011, p. 16). (4) A kindling mechanism that can expand a fragment into a full copy. It appears that prior Hebbian strengthening of synaptic weights could supply this feature. (5) A binding mechanism that can synchronize the timing and arrival of perception and reference signals, respectively. An alternative here may be a recursive or reverberating mechanism that can keep a signal going until the signal to be compared with it arrives. (6) Possibly, a gating mechanism that does not pass along a signal until the requisite computations have been performed. Potential candidates for this mechanism would include: (a) thresholds, (b) coincident signals, especially on distal dendrites (Numenta, 2011), and/or (c) disinhibition. (7) A relative timing mechanism that keeps the dynamics of a higher level operating at a slower rate than a lower level. It is unclear whether conduction delay or a differential number of processing steps would be sufficient to provide such a mechanism. The above features of a PCT approach to constructing reference signals become implicit predictions for features that will be uncovered by further research into the neural wiring of the brain. Because the broad functional approach offered by PCT can model actual behaviors to a very high standard, as spelled out earlier in this chapter, it is well worth considering whether there are in fact neurophysiological enactments of PCT components and equations still waiting to be found.

Acknowledgments I wish to thank Travis DeWolf, Henry Yin, Dominique Berule, and Warren Mansell for their heroic help in improving earlier versions of this chapter.

References Bateson, G. (1979). Mind and nature: A necessary unity. New York: E.P. Dutton. Calvin, W. H. (1996). The cerebral code: Thinking a thought in the mosaics of the mind. Cambridge, MA: MIT Press. Chevalier, G., & Deniau, J. M. (1990). Disinhibition as a basic process in the expression of striatal functions. Trends in Neurosciences, 13(7), 277e280. Available at: http://e.guigon.free.fr/rsc/ article/ChevalierDeniau90.pdf.

142 SECTION | B Models of brain and behavior George, D. (2008). How the brain might work: A hierarchical and temporal model for learning and recognition. Doctoral dissertation, Stanford University. Available at: http://alpha.tmit.bme.hu/ speech/docs/education/02_DileepThesis.pdf. George, D., & Hawkins, J. (2005). A hierarchical Bayesian model of invariant pattern recognition in the visual cortex. In Neural networks, proceedings: 2005 IEEE international joint conference on neural networks (Vol. 3, pp. 1812e1817). Available at: http://www.neurosecurity. com/articles/GeorgeHawkinsIJCNN05.pdf. George, D., & Hawkins, J. (2009). Towards a mathematical theory of cortical micro-circuits. PLoS Computational Biology, 5(10). https://doi.org/10.1371/journal.pcbi.1000532. e1000532. Glezer, V. (1995). Vision and mind: Modeling mental functions. Mahwah, NJ: Lawrence Erlbaum Associates. Groenewegen, H. J. (2003). The basal ganglia and motor control. Neural Plasticity, 10(1e2), 107e120. Guillery, R. W. (1995). Anatomical evidence concerning the role of the thalamus in corticocortical communication: A brief review. Journal of Anatomy, 187(Pt.3), 583e592. Available at: http:// www.ncbi.nlm.nih.gov/pmc/articles/PMC1167461/pdf/janat00131-0069.pdf. Hawkins, J., & Blakeslee, S. (2004). On intelligence. New York: Henry Holt and Company. Hawkins, J., George, D., & Niemasik, J. (2009). Sequence memory for prediction, inference and behavior. Philosophical Transactions of the Royal Society B, 364, 1203e1209. Hill, S., & Tononi, G. (2003). Thalamus. In M. N. Arbib (Ed.), The handbook of brain theory and neural networks (pp. 1176e1180). Cambridge, MA: The MIT Press. Jorgensen, E. (2005). A PCT approach to the frame problem [Online]. Available at: http://www. livingcontrolsystems.com/intro_papers/frame_problem.pdf. Koshland, D. E., Jr. (1980). Behavioral chemotaxis as a model behavioral system. New York: Raven Press. Llinas, R. R., Lesnik, E., & Urbano, F. J. (2002). Temporal binding via cortical coincidence detection of specific and nonspecific thalamocortical inputs: A voltage-dependent dye-imaging study in mouse brain slices. Proceedings of the National Academy of Sciences of the United States of America, 99(1), 449e454. Available at: http://www.pnas.org/content/99/1/449.full. Marken, R. S. (1983). “Mind reading”: A look at changing intentions. Psychological Reports, 53, 267e270. Marken, R. S. (1989). Behavior in the first degree. In W. A. Hershberger (Ed.), Volitional action (pp. 299e314). Marken, R. S. (1992). Mind readings: Experimental studies of purpose. Gravel Switch, KY: The Control Systems Group. Marken, R. S. (2002). More mind readings: Experimental studies of purpose. Chapel Hill, NC: New View Publications. Marken, R. S. (2003). A hierarchy of perception and control [Online]. Available at: http://www. mindreadings.com/ControlDemosJava/HP.html. Marken, R. S. (2014). Doing research on purpose: A control theory approach to experimental psychology. Chapel Hill, NC: New View Publications. Marken, R. S., & Powers, W. T. (1989). Random-walk chemotaxis: Trial and error as a control process. Behavioral Neuroscience, 103(6), 1348e1355. McFarland, N. R., & Haber, S. N. (September 15, 2002). Thalamic relay nuclei of the basal ganglia form both reciprocal and nonreciprocal cortical connections, linking multiple frontal cortical areas. Journal of Neuroscience, 22(18), 8117e8132. Available at: http://www.jneurosci.org/ content/22/18/8117.full.

How the brain gets a roaring campfire Chapter | 6

143

Mink, J. W. (2003). The basal ganglia and involuntary movements: Impaired inhibition of competing motor patterns. Neurological Review, 60(10), 1365e1368. Available at: http:// archneur.jamanetwork.com/article.aspx?articleid¼784785. Numenta. (2011). Hierarchical temporal memory: Including HTM cortical learning algorithms. (Version 0.2.1, September 12, 2011). Available at: http://numenta.com/learn/hierarchicaltemporal-memory-white-paper.html. Ohara, P. T., & Lieberman, A. R. (1985). The thalamic reticular nucleus of the adult rat: Experimental anatomical studies. Journal of Neurocytology, 14(3), 365e411. Olshausen, B. A., & Field, D. J. (2004). Sparse coding of sensory inputs. Current Opinion in Neurobiology, 14, 481e487. Plooij, F. X. (1990). Developmental psychology: Developmental stages as successive reorganizations of the hierarchy. In R. J. Robertson, & W. T. Powers (Eds.), Introduction to modern psychology: The control-theory view (pp. 123e133). Gravel Switch, KY: The Control Systems Group. Plooij, F., & van de Rijt-Plooij, H. (1990). Developmental transitions as successive reorganizations of a control hierarchy. American Behavioral Scientist, 34, 67e80. Powers, W. T. (1973). Behavior: The control of perception. Chicago: Aldine Publishing. Powers, W. T. (1978). Quantitative analysis of purposive systems: Some spadework at the foundations of scientific psychology. Psychological Review, 85, 417e435. Powers, W. T. (1979). A cybernetic model for research in human development. In M. N. Ozer (Ed.), A cybernetic approach to the assessment of children: Toward a more humane use of human beings (pp. 11e66). Boulder, CO: Westview Press. Powers, W. T. (1990). A hierarchy of control. In R. J. Robertson, & W. T. Powers (Eds.), Introduction to modern psychology: The control-theory view (pp. 59e82). Gravel Switch, KY: The Control Systems Group. Powers, W. T. (1992). Deriving closed-loop transfer functions for a behavioral model, and vice versa. In Living control systems II: Selected papers of William T. Powers (pp. 145e160). Paper originally presented at the American Society for Cybernetics Meeting, October 1983, Palo Alto, CA. Powers, W. T. (2008). Living control systems III: The fact of control. Bloomfield, NJ: Benchmark Publications. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960). A general feedback theory of human behavior: Part 1. Perceptual and Motor Skills, 11, 71e88. Sherman, S. M. (1996). Dual response modes in lateral geniculate neurons: Mechanisms and functions. Visual Neuroscience, 13, 205e213. Sherman, S. M. (2006). Thalamus. Scholarpedia, 1(9), 1583. Available at: http://www. scholarpedia.org/article/Thalamus_anatomy. Sherman, S. M. (2007). The thalamus is more than just a relay. Current Opinion in Neurobiology, 17(4), 417e422. Sherman, S. M., & Guillery, R. W. (June 1998). On the actions that one nerve cell can have on another: Distinguishing “drivers” from “modulators”. Proceedings of the National Academy of Sciences of the United States of America, 95, 7121e7126. Neurobiology. Available at: http:// www.pnas.org/content/95/12/7121.full.pdf. Wikipedia. (2014). PID controller. Wikipedia, the free encyclopedia. Available at: http://en. wikipedia.org/wiki/PID_controller. Yin, H. H. (2014). How basal ganglia outputs generate behavior. Advances in Neuroscience, 2014, 28. Article ID 768313.

Chapter 7

How the brain gets a roaring campfire: Input and output functions Erling O. Jorgensen Riverbend Community Mental Health, Concord, NH, United States

Introduction In the last chapter, a simple campground scenario was presented, with campers wanting a fire large enough to heat some water so they could have some flavored decaf coffee before heading off to their tent for the night. What seems like a straightforward task for the brain, upon closer examination proves to be quite a complicated situation, with multiple types of perceptions at work, many of them needing to be brought to different preferred states, so that the desired outcome would appear. A model of negative feedback control was presented, operating at multiple hierarchical levels, to explain how the brain may be obtaining its preferred type of campfire, and the higher level goals that might be in play. This model, now known as Perceptual Control Theory (PCT), was first proposed by William T. Powers and his colleagues in 1960 (Powers, Clark, & McFarland, 1960a, 1960b), and elaborated in subsequent works (Powers, 1973, 2008). Hallmarks of the PCT model include a very stable modular structure, hierarchical arrangement of cascaded control, and the flexibility of operating at multiple time scales. The elemental system of PCT equations is extremely robust. These features make it a promising candidate for integrating a broad range of functional findings about the brain. The previous chapter presented the four basic functions of an elemental PCT control loop (the boxes in the first figure of Chapter 6). Like the west, north, east, and south points on a compass, a clockwise listing of them is as follows: Perceptual Input function, Comparator function, Output function, and Environmental Feedback function. Entering the loop from the top at the point of the comparator is a reference signal, to provide a standard against which the perceptual signal is compared, generating in the process an error signal to The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00007-1 Copyright © 2020 Elsevier Inc. All rights reserved.

145

146 SECTION | B Models of brain and behavior

drive the loop’s output. Entering the loop from the bottom, in the environment, is a disturbance affecting the state of the perceptual input, but which can be counteracted by behavioral output via the effects of the environmental feedback function. This is the basic PCT module that can be applied at various degrees of scale and with a variety of hierarchical connections. The purpose of my chapters is to examine critical aspects of the classic PCT functions from the standpoint of neurophysiological research. In the previous chapter, the key issue was the commensurability of the reference and perceptual signals, so that they could be suitably combined in the comparator function. This entailed examining the metrics of how a reference signal would have to be constructed, in order to mesh with a perceptual signal of a different order. It specifically noted that what the cortical model of Hierarchical Temporal Memory calls the “name pattern” (Hawkins & Blakeslee, 2004, p. 153) could implement the principle of auto-associative addressing, which Powers (1973) predicted would be needed for the operation of reference signals. The present chapter takes this project further, by looking at the neurophysiological bases for other functions in the PCT model. It begins with some prefatory remarks about perceptual contours in PCT, and the predominant open-loop methodology in neuroscience. The second section examines a myriad of ways perceptual input functions may get constructed for the visual system of the brain. The third section begins to consider aspects of the motor output function, at least as it starts to take shape in basal ganglia structures (see the online version of this chapter). Because this chapter is investigating neuroscience research, it will not address the environmental feedback function, which closes the loop external to the person. The last of these three, available online, will outline a model of the thalamus, as considered through a metaphorical PCT microscope. Specifically, it will be proposed that thalamic relay nuclei provide a substrate for bidirectional comparators for perceptual control. A related matter is what type of addressing PCT reference signals employ, which seems to vary in different areas of the brain. In the previous chapter, a form of unsupervised addressing was discerned for the neocortex, through an auto-associative mechanism. This chapter touches upon a form of dynamic addressing, proposed for the basal ganglia, where the striatum there may form a bank of PCT comparators utilizing a temporal mechanism of co-incident detection from the cortex and thalamus. My third chapter will trace a form of structural addressing in the thalamus, using bi-directional mechanisms for PCT error signals, seeing as action potentials cannot go below zero. On the one hand, the band of thalamic reticular tissue serves to invert the sign of descending reference signals before entering comparators in the relay nuclei. On the other hand, inhibitory interneurons have a dendritic triad arrangement which can invert the sign of ascending perceptual signals.

How the brain gets a roaring campfire Chapter | 7

147

Prolegomenon: The what and the how A hierarchy of perceptions The PCT model makes a variety of predictions about the kinds of perceptions the brain may be producing. Table 7.1 presents a listing of perceptual types proposed by Powers (1973, 1979, 1990). These are mostly broad categorical descriptions, demonstrating some correlation with phenomenological experience, and thus offering heuristic guidance in understanding the brain. However, precise perceptual input functions and their respective physiological mechanisms have yet to be specified in the PCT model. Here, PCT must rely on domain-specific neurophysiological research to spell out the details of how the brain is constructing its perceptions. That is the kind of project that this chapter undertakes, with respect to visual perception in particular. The column entries in Table 7.1 may not sound very ‘perceptual.’ After all, where are the sounds and colors and textures that typically fill one’s perceptual world? And what does it mean to call something like a ‘program,’ for instance, a perception? Isn’t a program something that is done, more of an action than something perceived?

FIG. 7.1 Levels of Neural Processing. Forms of visual processing are presented as occurring in layered sheets, as one moves from peripheral retinal areas (the lowest five sheets), through thalamic areas (the next two sheets), and up through occipital (including striate, peristriate, and prestriate regions) and temporal cortex areas (the higher sheets). Reading across the diagram gives a functional name for each sheet, along with a brief description of the type of perceptual transformation involved. This manner of classification is gleaned from research presented by Glezer (1995).

148 SECTION | B Models of brain and behavior

TABLE 7.1 Proposed Hierarchical Levels of Perception (According to Perceptual Control Theory). Hierarchical level (from periphery)

PCT name

Type of perception

First Order

Intensities

Magnitudes, amounts

Second Order

Sensations

Attributes, weighted sums

Third Order

Configurations

Collections of attributes

Fourth Order

Transitions

Paths, rates of change

Fifth Order

Events

Temporal segmentations

Sixth Order

Relationships

Co-variations

Seventh Order

Categories

Class memberships

Eighth Order

Sequences

Serial orderings

Ninth Order

Programs

Networks of contingencies

Tenth Order

Principles

Guiding heuristics

Eleventh Order

System Concepts

Sense of organized unities

Perceptual Control Theory makes the case that whenever something is done by a living organism, it is in order to attain a perceptual result. Indeed, perception is all any organism can know of its environment. So if it is to monitor how it is doing within that environment, it must be via its perceptions. This holds true whether a cell is monitoring the ionic state of its nutritional requirements, or a person is monitoring the achievement of certain high level principles, like honesty or return on investment. To enact a principle requires some course of action, but of course choice of the ‘right’ program and how it is progressing are both perceptual results in their own right. And programs achieve their results by bringing about sequences of action in the proper order. So at this level of perception, sequence makes all the difference. It is at this juncture that there seems a tendency in some neuroscience literature to make quite a large leap. The logic seems to be as follows. Yes, people must carry out sequences, and of course that means sequences of muscle forces. So, what are the “motor commands” being generated high up in the brain to make those things happen? The leap being considered here, at least from the standpoint of the PCT enumeration of perceptions in Table 7.1, is that it appears that half a dozen levels of intervening perceptions are being straddled, as if they did not matter.

How the brain gets a roaring campfire Chapter | 7

149

While Table 7.1 is only a phenomenological list, not fully tested, its choice of perceptual levels is not arbitrary. The studious attempt has been made (Powers, 1973, 1979, 1990) to discern what type of perceptions would have to be present in order to construct each next higher level of perception. And typically e although this, too, is a hypothesis not fully tested e a higher level would not issue instructions for perceptions several levels removed from its own, at the risk of introducing instabilities into a smoothly graded system. So what is needed next according to Table 7.1 in a descending journey to lower orders of perceptual results? There is a need for sequences of the right categories of items in the right relationships. And it will be relationships among the proper macro events for segmenting the temporal world. And those events will need their constituent parts to change via the right transitions. And it will be transitions among the right joint angles and body configurations. And those postural results will only happen if the right sensations for muscle length are occurring among the stretch receptors of the respective joints. And only then can one speak of the right muscle forces being enacted so that the proper stretching can happen, so the joints move into the right configurations at the right rates of speed, and on up with the right perceptual results for each level in the hierarchy. The implementation problems raised by such a conceptualization cannot be ignored. There must be mathematical specification, across many degrees of freedom, with the right hierarchical dependencies (Cools, 1985). There is an inherent temporal nature to the brain’s input, which may require certain forms of standardization in the brain. There are questions of conscious attention, or whether predictive functions are necessary. The literature of PCT takes such issues quite seriously.1 This current chapter will examine a variety of input and output functions for possible control loops in the brain. This will include a detailed look at perceptual input functions as described by Glezer (1995) and other researchers for the brain’s visual system. Here it will become clear that while the proposed PCT hierarchy lists plausible classes of perception, they are much more nuanced when it comes to their actual neural embodiments. This chapter will also consider (see online version of this chapter) how the complicated structures of the basal ganglia may be understood in terms of gating functions, channeling the brain’s output circuitry for eventual closure of its loops of control through the environment.

Open-loop methods to study closed-loop features Much of computational neuroscience is an attempt to reverse engineer the operation of the nervous system. It does this by taking into consideration 1. The online version of this chapter suggests several distinct perspectives raised by a PCT approach to these questions.

150 SECTION | B Models of brain and behavior

questions of architecture (i.e., what connects to what?), dynamics (i.e., what are the interactions?), and procedures for modifying parameters (i.e., what degrees of freedom are available?) (Churchland & Sejnowski, 1992). In the process of deriving potential functions, it becomes a representational endeavor. As Churchland & Sejnowski (1992) note, computational functions provide “a mapping between the system’s states and the states of whatever is represented” (p. 62). The survey work for that mapping, however, tends to leave out an important feature of the landscape. There is limited discussion of the phenomenon of control, within the field of neurophysiological research. Perceptions in the brain do not just happen, they are made to happen. This is true in a double sense. The non-controversial portion is that functions for registering perceptual inputs are constructed. Understanding functions in the brain is difficult and meticulous work, by necessity involving a great many simplifications and approximations and assumptions. In particular, this task creates models of models of models. For instance, the net effect of thousands of synapses is modeled as gross patterns of graded potentials and firing rates. These simplified neurons are modeled as simplified mathematical processes. Population dynamics within certain structures of the brain are modeled as interactions among a small subset of such neurons. Those brain structures are discerned within certain experimental animals, which themselves serve as living models for similar species. These are accepted and arguably credible procedures within the field of neuroscience. The simplification that may not be warranted e at least according to PCT e is that these modeled processes occur with open loop causality. There is not a straightforward flow from environmental cause to peripheral sensor to information processing within the brain to behavioral output. That is not the way to understand the operation of perceptions in the brain. Rather, PCT proposes loops of circular causality, happening at every level of perception. This is to say, perceptions are not only constructed, they are controlled. The brain actively generates perceptual stabilities, and those stabilities are used to construct and then control higher levels of stability. Control, according to PCT, is not just some circular inner process of tweaking different neural values for the perceptions of note. Yes, those feedback processes do occur, but so does the process of actively regulating the values against a standard of reference. That is what distinguishes control from mere feedback. And the control proposed by PCT takes a particular trajectory. It is a matter of generating different perceptual values by way of the environment, specifically by affecting something ‘out there’ that then leads to preferred neural values ‘in here.’ That is the ultimate phenomenon that neuroscience must explain, in all its details. So argues PCT. With such a complicated mechanism as the nervous system to reverse engineer, the process necessarily proceeds in piecemeal fashion, on small segments of neuroanatomy at a time and with brief activities of

How the brain gets a roaring campfire Chapter | 7

151

neurophysiology. This makes it difficult to discern more global properties of the system. For instance, a global property that Perceptual Control Theory points to (or at least postulates) is the operation of complete negative feedback circuits, which extend from preferred goal states within the organism out through the physics of the environment and back into the perceptual machinery of the organism, where they are monitored to see if they have been achieved. What gets controlled is not the behavioral process itself, but the perceptual results at each sub-stage of the process. Such an arrangement is difficult to detect at the level of neuronal connections in some small section of the nervous system. There may be telltale signs on the more micro level, but they can easily be missed if one is not looking for them. This speaks to the value of PCT as a metaphorical microscope (see my online chapter), to raise such considerations and see if there is neuronal support for them. Here is an instance where a good model allows one to “make predictions about undiscovered local and global properties” (Churchland & Sejnowski, 1992, p. 133) in the broader system. Moreover, because of the local scale of operation of much of neuroscience, any macro circuits tend to get examined segmentally. This is especially the case with the black box nature of reverse engineering in general, where the workings inside the box are investigated by applying inputs and seeing what outputs emerge. This input/output methodology is essentially treating circuits in an open loop manner, consistent with a widespread open loop concept of causality. However, from the paradigm of circular causality reflecting cybernetic notions, it is essentially severing the loop. Care must be given to distinguish open-loop properties of the components from closed-loop properties of the circuit as a whole. Perhaps an example might demonstrate what an open-loop methodology has to offer. Perrone & Thiele (2002) have conducted a careful study to model how neurons in the middle temporal (MT) cortex of primates construct a perception of speed generated by a moving edge. They begin with several data sets gathered on the temporal and spatial frequency responses of neurons in the primary visual (V1) area of the striate cortex, fed by the contrast-sensitive neurons in the lateral geniculate nucleus of the thalamus. All this information has been assembled open-loop, with electrode arrays recording responses in a monkey brain, determining first the optimal orientation and size of spatial frequency gratings for each neuron, before varying the temporal frequency by systematically changing the grating speeds. From there, Perrone & Thiele (2002) conduct a variety of transformations of the data, each one biologically plausible, to see if they can construct a model that matches how MT neurons encode speed information. Speed could emerge as a ratio between neurons registering ‘transient’ (T) motion compared to ones registering a ‘sustained’ (S) static input, and indeed both types of neurons have been found in the V1 striate area. The technical difference is that the latter can be modeled with a low-pass temporal filter (i.e., information

152 SECTION | B Models of brain and behavior

passes through even from a static image with no movement), while the former is modeled with a band-pass temporal filter (i.e., slow movements are filtered out.) Responding in a log scale fashion is a common property among sensory inputs, so the values are scaled that way. The advantage this conveys is that the division needed for a ratio can occur via subtraction, a simpler neural transformation, where log (T/S) ¼ (log T e log S). Continuing with useful transformations, the authors introduce an absolute operation to make all the values positive (even when log S > log T), with the exception of the one lowest value when log S ¼ log T. Introducing an inversion operation makes that value the peak of the response curve, with other values falling away on either side. What this means is that there is now maximal response at the cross-over point where the ‘sustained’ neural response from an image is equal to a moving response from the ‘transient’ neural function of a V1 cell. Even though these two values are equivalent, they derive from functions representing a potential speed difference, so the crossover point into the band-pass region becomes an analogue of relative motion, i.e., speed. The feeding functions still register their responses independently, but their combination holds promise for encoding a different feature than either one alone captures. Perrone & Thiele (2002) note one implication as follows: “The additional absolute and inversion operations enable the two broadly tuned transient and sustained temporal filters to be converted into a narrowly tuned unit” (p. 1038). There are a few other refinements in their model, such as getting all the dimensions on the same 3D graph so that the plane of intersection between the ‘transient’ V1 neuron and the ‘sustained’ V1 neuron lines up on a diagonal, representing a ridge of optimal response of the MT neuron. The slope of this ridge is what corresponds to the speed of a moving edge. A parameter determining the degree of relative band-pass filtering in the ‘transient’ V1 neuron allows tuning MT neurons to different speeds. They also have parameters governing the width and length of the ridge, respectively. The extended ridge of intersection between the V1 input functions means that a given MT neuron is tuned to a particular speed. This is the case regardless of which combinations of temporal and spatial frequency edges may be generating that speed. This aspect may sound counterintuitive. But note, a moving spatial frequency of four narrow gratings will appear to pass the center of a receptive field more times than a single wide grating, even though they are actually moving at the same speed. So the temporal frequency input must be scaled according to the width of the spatial frequency grating. The result is multiple MT neurons each tuned to a specific speed of moving edges, reflected in the slope of the model’s graph. They designate their proposal as the Weighted Intersection Mechanism model, WIM, (Perrone & Thiele, 2002). The real payoff of this series of open-loop methods comes when they then test this WIM model against actual response data from MT neurons. They refer

How the brain gets a roaring campfire Chapter | 7

153

to six other published data sets of the response properties of neurons in the middle temporal cortex of primates, noting they are “from a large number of MT neurons tested with a broad range of speeds” (Perrone & Thiele, 2002, p. 1040). For instance, when the five parameters of the WIM model were optimized for each neuron in a set of 84, and then tested with 30 different combinations of temporal and spatial frequencies, the match between the WIM response profiles and that of MT neurons in the comparison study was as high as r ¼ 0.99, with a mean correlation of r ¼ 0.83 (p. 1044). This is a striking result, especially given that the model was constructed “using an economical mechanism based on the inputs from just two V1-like neurons” (Perrone & Thiele, 2002, p. 1044). What it shows is that a clearly specified quantitative model, that closely matches a range of neural response data, can be constructed from careful input/output studies using an open-loop methodology. It also shows that the emergent property of speed-tuned MT neural sensors can be generated from V1 neural sensors measuring quite different properties. From the standpoint of the PCT perceptual hierarchy discussed above, this study suggests that the speed of moving edges may get calculated and ‘perceived’ in the visual MT cortex. The purpose of this extended example has been to show that open-loop methods have a role to play, especially when it comes to discerning perceptual input functions that may be at work in closed-loop arrangements of negative feedback control. After all, each component within a PCT circuit either computes or conveys a signal that the circuit needs. Open-loop measurements that are within the time constant of a given level of perception are essentially gauging effects within the transport lag of that portion of the loop. The key is to consider whether other aspects of the closed-loop functioning2 may be changing the open-loop properties while they are being measured. There is a further issue with respect to time, when considering closed loops hierarchically, as with PCT’s proposals for cascaded control. Essentially, each level operates with its own time constant, where each relatively lower level has a faster time frame for enacting its results. It does not work for a higher level to call for results faster than they can be produced. That typically would lead to oscillation or instability, as a system would “sense its own actions as disturbances and try to correct for them” (Powers, 1973, p. 52). Rather, measuring events on a slower time scale as one moves up the hierarchy allows action to begin to move a perception in the proper direction at the lower level, before new instructions are issued. The hierarchical notions of PCT lead to another potential benefit. By adopting a mechanistic engineering approach, where each signal and function 2. Certain implications of closing the loop are presented in the online version of this chapter, including a stability factor used in testing for controlled variables (Powers, 1978; Marken, 1983).

154 SECTION | B Models of brain and behavior

in a loop must be accounted for, there is less danger of skipping over intervening steps. Indeed, on a macro level, PCT proposes nearly a dozen distinct types of perception, aside from specific instantiations by sensory mode at lower levels of the hierarchy. Each type of perception would need to be constructed by actual neural mechanisms, and arrayed with proper timing relationships, if higher levels are to achieve their own more abstract ends. This is quite a formidable project of reverse engineering for understanding the nervous system. A key role of PCT may be to serve as a rich heuristic source of what functions to look for, amidst all the tangled connections in the brain.

Constructing visual input functions: This way in This chapter turns now to forms of processing that seem clearly input related, from the standpoint of PCT control loops. Actual working models of a given physiological phenomenon in the brain require specification of the perceptual variables being controlled, values for the parameters involved in any given simulation, and constraints on which hierarchical levels will be included among the proposed control loops. Even that would be a radical simplification of the actual biological mechanisms within a living control system. PCT is a rigorously defined macro theory, providing a predictive map of the kinds of functions that would be expected, at various levels of scale. But to turn the modular functions of PCT into domain-specific instantiations of negative feedback control in a biological system requires neurological estimates of potential enacting mechanisms. The greatest need for those estimates, from the standpoint of PCT, is that of plausible perceptual input functions, which retain some of their computational rigor according to neurophysiological studies. What follows is a beginning in that direction. Glezer (1995), in a comprehensive presentation,3 presents some very intriguing findings as to what may constitute various perceptual input functions within the visual system of the brain. It is a detailed account, encompassing six broad areas, from the retina, to the lateral geniculate nucleus (LGN) of the thalamus, to the first visual area of the striate (occipital) cortex, to additional visual areas of the peristriate and prestriate cortices, to areas in the temporal cortex, and the posterior parietal cortex. He proposes the types of functions that may be enacted by different types and combinations of cells in these various areas. To give some sense of the richness of his account, an extended description of the construction of numerous types of perceptions is presented below. It is interlaced with occasional other models of visual functions developed by other researchers. 3. One indication of the scope of Glezer’s work is the inclusion of over 400 references, a quarter of which were Russian language works, which may not be readily available to an English language reader. Some four dozen citations were authored or co-authored by Glezer himself, and 70% of those were as lead author.

How the brain gets a roaring campfire Chapter | 7

155

As a foreshadowing of the argument to come, Table 7.2 suggests the range of neural models or semi-models that have been proposed for how visual input is processed in the brain. Examination of a visual scene, by adult viewers with a fully developed nervous system, gives the impression of a seamless integration of the visual elements. What is easily missed is that the result has actually been constructed from various modular components. Table 7.2 gives a beginning sense of some of those components. The basis of the first column in Table 7.2 is what Bateson (1979) calls “news of difference” by means of double description (p. 68.) It is the creation of new perceptual niches by new combinations of what has already been constructed. The resulting distinctions are listed in the second column, some of which do not have a very perceptually-sounding label. The third column points to the types of neural models in the literature that may generate those perceptual results. The fourth column lists either a primary or secondary source of the information about the respective models. So then, Table 7.2 serves as an advance roadmap of the journey ahead. This chapter will argue that it takes this many neural models and more to explain how campfire-related perceptions are constructed in the brain’s visual system.

Signal-to-noise sensitivity While the retina is the most peripheral area of the visual system, its transduction of light energy into neural signals is far from simple. One form of perception that it constructs, quite obviously, is that of brightness, which Powers (1973) according to his schema would classify as “Intensity”. The distribution of luminance meeting the eye is conveyed in analogue style as graded membrane potentials by the photoreceptors of the retina. It is then parceled according to a regular arrangement of receptive fields (RFs) in the retina. It is integrated within the central summation zone of each RF, and processed via graded membrane potentials by what Glezer (1995) calls “the triad, photoreceptor e horizontal cells e bipolar cells” (p. 2). In fact, the signal-to-noise sensitivity is actively improved by patterns of inhibition within the triad, so that random discharges of the cells do not signal light events. Rather, against a curtain of inhibitory effects, it takes a sequence of disinhibition to reliably signal light entering the eye. Based on such early retinal processing, the initial impression from the campfire would simply be the faintness of light from the original fire. Specific details of this triad and what each kind of cell contributes will not be spelled out here. The basic arrangement is overlapping zones where inhibitory effects on the bipolar cells are four times wider than excitatory ones (Glezer, 1995). The bell-shaped distributions, both inhibitory and excitatory, are superimposed on one another in what is called a Difference of Gaussian arrangement. The general effect is to perform a spatial type of analysis

156 SECTION | B Models of brain and behavior

TABLE 7.2 Construction of Perceptual Results by Layering of Neural Models. This non-comprehensive table illustrates how many models of neural functioning may be needed to generate a given type of visual perception. The best approximation of a perceptual label is italicized in the second column. Opportunity for perceptual niche

Constructed Result

Neural model or Semi-Model

Citation or source of information

Differentially route neural impulses to different areas of receptive fields

Improved reliability, a “clearer” signal

Center versus Surround

Barlow (1953), Kuffler (1953)

Subtract out average brightness

Local contrast between visual areas

Ratio of Adjoining Reflectances

Glezer (1995), Shapley et al. (1993)

Adjust sensitivity according to available light

Standardized contrast independent of level of illumination

Retinal Adaptation

Shapley & Enroth-Cugell (1984)

Only go ‘on’ or ‘off’ from a change in intensity

Onset/Offset signals

On-Center & Off-Center Cells

Kuffler (1953)

Quickly go off to a sudden change

Motion detection

Adaptive Cascade Model

Chen et al. (2013)

Go on in sequence from a moving stimulus

Detecting orientation

On-Signals in Sequence

Hubel & Wiesel (1962, 1974), Glezer (1995)

Overlap on- & off-zones

Selective response to large uniform area

Push-Pull Model of Thalamocortical Projections

Glezer (1995)

Differentiate contrast areas as sinusoidal gradations (i.e., spatial frequencies)

Standardized contrast signals independent of retinal position

Multichannel Model of Spatial Frequencies

Campbell & Robson (1968), Blakemore & Campbell (1969)

Capitalize on phase differences between similar inputs

Depth perception

Local Phase Disparity

Poggio (1995), Gonzalez & Perez (1998)

Collate areas of neural transformations yielding similar results

Figure versus ground distinction

Glezer (1995)

How the brain gets a roaring campfire Chapter | 7

157

TABLE 7.2 Construction of Perceptual Results by Layering of Neural Models. This non-comprehensive table illustrates how many models of neural functioning may be needed to generate a given type of visual perception. The best approximation of a perceptual label is italicized in the second column.dcont’d Citation or source of information

Opportunity for perceptual niche

Constructed Result

Neural model or Semi-Model

Autocorrelate extent of similarity versus changes in gradient

Differentiate texture

GaborGaussianLaplacian Algorithm

Fogel & Sagi (1989)

Differentiate borders between contours

Distinguish shape

Chains of Nonlinear Inputs

Glezer (1995)

Compare transient responses to sustained ones

Speed, Rate of change

Weighted Intersection Mechanism

Kulikowski & Tolhurst (1973), Perrone & Thiele (2002)

Notice the directional sign of foreground versus background movement

Distinguish depth by motion parallax

Depth-Sign Indicator

Maunsell & Van Essen (1983), Nawrot & Joyce (2006)

Overlap filters at harmonic intervals

Superposition of standardized relative contrast differences

Modules of Harmonic Channels

Glezer (1995)

Compile sinusoids & harmonics from multiple spatial frequencies

Visual objects grouped by harmonic composition

Piecewise Fourier Analysis by a Model of Modules

Glezer (1995)

Jointly encode location in space & direction from the body

Spatial relationships among recognized objects

Basis Functions with Blended Coordinate Systems

Pouget & Sejnowski (1995, 1997)

(Dowling, 2007), by establishing receptive fields with a characteristic excitatory center and inhibitory surrounding region (Barlow, 1953; Kuffler, 1953). It is worth noting, on the behavioral level, that a global perception of brightness is subject to negative feedback control, apart from what the neural computations are doing. The simple act of squinting with the orbicularis muscle shows that it is possible to regulate how much brightness enters the

158 SECTION | B Models of brain and behavior

eye. There is also a fast-operating, autonomic feedback response using iris muscles to narrow or dilate the pupil, to regulate the amount of light intensity. The campfire watcher at the beginning of the previous chapter, by an even slower sequence of behaviors, sought among other things more brightness than the dying embers could provide. On a more mundane level, a way to heighten and thereby regulate the reliability of the visual signal’s range of information is to provide differential attention by means of the central foveal position of the eye. Even this early discussion of how perceptions are constructed in the retina brings up a distinction that should be considered from the outset. There is a difference between structural reliability as to how a function is fashioned, and enacted reliability as to how a function is used. The latter is a behaviorally enacted reliability achieved by doing something in the environment, to achieve a specific change in the values the perceptual cells are reporting. When the term “control” is used in these chapters, it refers to this latter process of enacted reliability. Essentially, each type of perception is exploiting a perceptual niche created by the structural reliability of lower levels, and often stabilizing the niche further in a behavioral manner by the enacted reliability described above. As each new type of perception is constructed, as a form of emergent reliability, it creates a perceptual niche to be further exploited and stabilized by higher levels of perception.4

Early forms of contrast Returning to the brightness signal, there is another form of preliminary processing. When global illumination changes, such as when the sun appears after being obscured by some clouds, the various local retinal ‘pixels’ registering those changes will increase by comparable proportions. This means that relative reflectances between areas will stay mostly constant. It appears that local reflectance is comprised of whatever change in reflectance occurs plus the average reflectance5 of background areas conveyed via neighboring cells (Glezer, 1995). This allows the contribution of global brightness to be parceled out. This is the construction of relative brightness, or contrast. In the context of the campfire scene, it means differential patches start to emerge. Such a standardized perceptual invariance, which does not fluctuate wildly with every transient shift in light level, is necessary before local changes can have meaning in their own right. This illustrates the PCT conception of perceptual levels, where prior perceptual processing provides the substrate upon which the next level of 4. There is a further discussion of these types of reliability in the online version of this chapter. 5. The average reflectance component may contribute to the need for a higher level parameter for constructing spatial frequencies, as discussed in the online version of this chapter.

How the brain gets a roaring campfire Chapter | 7

159

processing is built. Constructed as it is by differential weighting of its inputs, contrast seems an example of what Powers (1973) would characterize as a “Sensation.” However, that construction is far from straightforward. For one thing, relative brightness is mostly dependent on the contrast at the border between two areas (Shapley, Kaplan, & Purpura, 1993), as is demonstrated by many perceptual illusions where identically bright patches look different against varying backgrounds. And attempts at control involve more eye saccades directed to the borders between patches. That may be part of the allure of a flickering fire. For another thing, “the perception of contrast depends upon retinal adaptation” (Shapley & Enroth-Cugell, 1984, p. 266), which refers to anything that can change the physiological sensitivity of the retinal cells. In total darkness, the retina is most sensitive in its response to light. But the presence of light can change that sensitivity, for instance, through the breakdown of photo pigments (Shapley et al., 1984), or through depletion of presynaptic vesicles of neurotransmitters in the bipolar cells (Jarsky et al., 2011). Thus, retinal adaptation serves as a negative feedback form of gain adjustment, where sensitivity is reduced as the amount of light is increased. In this way, a standardized signal of relative contrast gets transmitted to higher areas of the brain across a wide range of potential illumination levels (estimated by Shapley et al. (1984), as ten orders of magnitude). Here again, adjoining RF properties and alteration in cellular properties (both forms of structural reliability) contribute to an emergent reliability with contrast as its own type of perception. Or in other words, borrowing from the campsite scenario, the gathering twilight does not interfere with the contrast and thus enjoyment of a re-kindled campfire. Once constructed, degree of preferred contrast can then be controlled, as a form of enacted reliability. For example, one can reposition oneself so as not to stare into the glare of the sun when trying to distinguish nearby details. Or one can stay in the dark for half an hour prior to star-watching.

On and off signals Just four layers of cells beyond the photoreceptors, an important elaboration of the perceptual signal takes place, via the further retinal triad of bipolar cells e amacrine cells e ganglion cells. There are specific on-center and off-center cells (Kuffler, 1953), which respond to increases and decreases of light, respectively, in the center of their RFs (Glezer, 1995). That is to say, not just light intensity but changes in light intensity, either up or down, are registering here. Dowling (2007) notes that, in addition to the spatial arrangement and analysis of the previous triad, this triad involving the ganglion cells amounts to a temporal analysis of the changing distribution of light entering the retina. Registering temporal changes in light distribution means the campfire is glowing.

160 SECTION | B Models of brain and behavior

This involves constructing perceptions of visual on-ness and off-ness. The perception of “off” here does not mean the lack of a signal. Rather, it is a specific signal designating that the visual input to an area of the retina has diminished. Glezer (1995) notes these signals are projected from the retina onto the linear RFs of X-cells within the thalamic LGN. The on-zones and offzones of those thalamic RFs are ways to represent a moving and changing visual world. On and off signals are both necessary for certain linear and nonlinear operations further upstream within the striate cortex. It is worth considering whether this is an example of what Powers (1973) would call a “Transition,” or the perception of change. Certainly there is detectability going on, when a neuron fires as a result of the onset or offset of light intensity. However, that argument could be made for any perceptual level. Just opening one’s eyes leads to a change in neural firing of one’s visual sensors. As Bateson & Bateson (1987) note with their definitions, a sensory organ is “a device that responds to difference” (p. 17). So not all differences should necessarily be called transitions. Indeed, as discussed by Powers (1973, chapter 10), transitions are best thought of as rates of change, signals that vary as a function of time, whether as derivatives, durations, or even partial derivatives. That seems not to be the case for the ‘on’ and ‘off’ signals discussed here. Registering such changes would certainly be an important feature for a system sensitive to the magnitude of its stimulating energy. Orienting movements of the head and eyes, when there is a sudden local change in light intensity, serve to demonstrate that such a feature can be subject to negative feedback control. It can also be controlled in more extended ways. In stoking the campfire, the process of making dry kindling catch fire was an interim controlled result, likely signaled by an array of on-center perceptual functions. Chen et al. (2013) propose a mechanism to account for the onset of motion in retinal receptive fields. When a bar moves in the RF center of retinal ganglion cells, the onset of that motion is selectively captured by the fast-Off version of those cells. This is verified by peak response from the leading edge of a dark bar, or the trailing edge of a bright bar, either one of which elicits an off-response. They proceed to develop an Adaptive Cascade Model of ganglion cell firing, with components such as a sequential summation of bipolar cell inputs, and gain control sensitivity at both the bipolar and ganglion cell levels. This model closely matches differential results whether from a bar’s appearance in the RF, smooth motion across the RF, or sudden onset (at various speeds) from the RF center. In all these situations, the key contributor is the fast-Off type of ganglion cell. This is likely the neural signal operating when a spark arches out and away from newly kindled sap within a campfire. The nervous system appears capable of discerning potentially relevant motion, even when the nature of that motion is unclear. The physiological evidence includes the following. To anticipate the discussion under the next

How the brain gets a roaring campfire Chapter | 7

161

sub-section below, there are concrete findings that Glezer (1995) presents regarding flickering stimuli. At low degrees of contrast for spatial frequencies, Glezer (1995) states, “In the low-frequency range, however, if the eye is stimulated by a flickering grating at threshold, at first only the flickering is seen; it is only when the contrast is increased that the grating itself is seen” (p. 90). An example might be the way steam rising from a cup of coffee e or a pot of water placed within a flaming campfire e is detected as a form of motion, even if the specific swirls are not yet discernible. As discussed above, this is still a matter of perceiving the fact of change, not measuring the rate of change. Thus, the on and off alternation, especially the off component constructed as it is by weighted inputs from ganglion to thalamus, would best be classified as a “Sensation” under the PCT formulation. There is subtlety, however, in how the retina and thalamus interact. Nonlinear Y-cells, a designation applied to certain retinal ganglion cells as well as their corresponding version in the lateral geniculate nucleus (LGN) of the thalamus, are comprised of large RFs with overlapping subunits tuned to detecting movement (Glezer, 1995). The inputs to these cells come proportionately higher from the peripheral areas of the retina, rather than the fovea. Thus, even indistinct peripheral movement such as a bird flitting off to the side can catch one’s eye, before added discrimination is supplied by the more linear properties of the fovea and their projections to linear X-cells of the LGN. Movement then, as signaled by the nonlinear outer portions of Y-cell RFs, starts to become an additional type of constructed PCT perception. The construction of on-signals and off-signals appears to make a different type of perception possible, namely spatial orientation, which might be classified as a “Configuration” according to PCT. Classic research by Hubel & Wiesel (1962, 1974) demonstrated the presence of cortical cells sensitive to orientation of the stimulus bars, i.e., a drifting sinusoidal grating (Sompolinsky & Shapley, 1997). Such a perception could be constructed by a series of thalamic on-signals that go on in sequence during visual scanning, without adjoining off-signals occurring, as the scanning of the bar moves lengthwise across the visual field (Law & Cooper, 1994; Lee, Blais, Shouval, & Cooper, 2000; Jin, Wang, Swadlow, & Alonso, 2011). This is essentially combining a temporal distinction with a spatial one. The ascent of the smoke from the campfire may well be registering through such a series of signals with a vertical orientation. In potential support of this notion, Glezer (1995) notes that there appear to be complex RFs of the striate cortex collating LGN output of on-cells only (p. 49), as well as other responses proportional to the length of a bar moving through the RF (see also Cudeiro & Sillito, 2006). In a sense, each on-signal cell might represent an “edge detector” for the leading edge of a stimulus, whereas the series of on-signal cells may represent a “bar detector” delineating orientation and perhaps directional movement (Glezer, 1995, p. 42). Reading a page of text, rotating and aligning it if need be, is one way to control the

162 SECTION | B Models of brain and behavior

corresponding perceptions of orientation. In a sense, a line of text with white line spaces above and below it could be treated akin to a bar to assist with horizontal tracking of the eyes, even as each letter presents a variety of edges to be deciphered for meaning. The above series of perceptual constructions from the retina, thalamus, and early projections to the striate cortex only go a portion of the way toward actually experiencing the campfire. The preceding sub-sections in this chapter merely set the stage for higher levels of perceptual processing, as can be noticed using instances from the campfire scenario. For instance, the initial intensity conversion of light energy into polarized membrane potentials is at most a sense of the faintness of the energy from the existing fire. And the improved signal-to-noise sensitivity of the centersurround arrangement of RFs merely constructs a perception of clear versus indistinct. The opponent weighting of retinal cone responses starts to allow a sense of orange versus yellow. To contrast one area of brightness against another is to start to construct patches, while retinal adaptation allows reliable registering of contrast despite the gathering twilight. The onset and offset of ganglion cells could be called a perception of glowing. The fast-off ganglion response from the trailing edge of a moving bright spot allows noticing a spark. Linear on/off projections to the thalamus become the flickering wisps of smoke, while a series of on-signals in a particular orientation becomes a sense of ascending. All that can be known about the environment is what perception says is there. Consequently, as of this stage of visual processing, only these disparate aspects exist, not the campfire in its own right.

Figure-ground contrasts One of the forms of emergent reliability that comes about as these perceptual functions are built up, layer by layer, is that of figure-ground contrasts. Indeed, that distinction seems so axiomatic as one looks around, that it is easy to forget it is actually a perceptual accomplishment for a developing brain. That is to say, the distinction between a foreground figure and its visual background is a construction, built upon prior constructions the brain has already made. Some of those distinctions have already been discussed. For instance, retinal pigments sensitive to different wavelengths transduce that reliable property of light into a reliable property of neural firing. The notion of a receptive field, from retinal cells on up through networks in the cortex, generally makes a distinction between the central response and that on the periphery. Global mean illumination gets parceled out, so that the local difference in reflectance between adjoining areas becomes a reliable distinction to build upon. As a result, perception of contrast is mostly independent of level of illumination. In the campfire illustration, finding the brownish-gray firewood on the brownishgray ground, whether in the ambient light of dusk or the growing light of a

How the brain gets a roaring campfire Chapter | 7

163

rising fire, may have drawn upon the invariant ratio of contrasts between the reflected surfaces. Contrast, by definition, is a comparison between different portions of the visual field. It therefore has high salience for the issue of distinguishing a figure from its background.6 Because the eventual construction of spatial relationships among recognized objects is an important one for the nervous system, this matter of figure-ground distinctions gets increasing attention, as ascending perceptual signals get processed. There are several forms of perceptual means available for highlighting visual contrasts and making those distinctions. These would include differences in orientation, depth, color, texture, and direction of movement. Most of these involve their own set of channels and complex computations in the striate and higher cortical areas. Drawing upon algorithms by Fogel & Sagi (1989), Glezer (1995) presents an extended discussion, which will not be recapitulated here, of a model for how texture and textural borders may get constructed in the peristriate and prestriate cortices (pp. 120e129), utilizing autocorrelation functions in particular. A certain standardization takes place as signals project from the thalamus to the striate cortex. RFs in the early relay areas of the thalamus preserve a retinotopic arrangement. To increase structural reliability, however, a method is needed that is not dependent on changing retinal position. The concept of spatial frequency fulfills this requirement. Images constructed in terms of cycles of light-to-dark shifts are known as spatial frequencies, referring to the number of those cycles occurring within a given visual angle. These can be modeled in terms of what are called Gabor functions, which multiply an undulating sinusoid function by a bell-shaped Gaussian function (Glezer, 1995). The brain takes on a mosaic character in such regions, as functions at one level get combined with subsequent functions, to create a greater whole. A lattice of these Gabor filters can represent various spatial frequencies arranged by different directions of orientation, a combination that helps with gradient discrimination. At the campsite, gradients of illumination mean the differential shading among the light shadows. A low spatial frequency is used for detecting large, relatively uniform areas of contrast, and this starts to occur in the striate cortex in the occipital lobe of the brain. Subjectively, this is the first indication of the size of the fire. What was previously captured in multiple sub-images of retinal ‘pixels’ now takes on a broader perceptual sense independent of shifts in retinal position. General size is first constructed by the lowest spatial frequency that fits the overall degree of contrast of a visual patch from its adjoining areas. Higher spatial frequencies mean there are more alternating light-to-dark cycles within the same visual angle, although these are mostly constructed at

6. The structural and emergent reliability of visual contrast allows for enacted reliability by means of negative feedback control, as noted in the online version of this chapter.

164 SECTION | B Models of brain and behavior

higher cortical levels. However, situations of high contrast create an early substrate among thalamocortical projections. Linear on- and off-cells of the LGN were discussed above in connection with flickering stimuli; an example would be watching the sun glint upon gently rippling water. There are also nonlinear cells where, at higher contrast, the center of the RF decreases in size. This is similar to retinal adaptation, where the properties of the RF are shifting. What ends up happening is that these cells convey dual or even multiple peaks, a spatial frequency fundamental in the RF center and a second or higher harmonic in the RF surround. This kind of combination of lower order sensation vectors would likely be considered a “Configuration” by Powers (1973), and it sets the stage for the cortex standardizing its input in terms of spatial frequencies. As will be shown below, spatial frequencies are pivotal in Glezer’s (1995) conception of how the visual cortex is organized. Glezer (1995) argues that there are arrays of complex striate RFs, with neighboring cells having an overlapping phase offset of ninety degrees in their optimal responses. One can picture it as the way sine and cosine waves co-vary by ninety degrees. These cortical RFs are comprised of pairs of on-zones and off-zones projected from the LGN, some of which converge via inhibitory interneurons. Such an arrangement is consistent with thalamocortical connections that other researchers have reported (Llina´s, Lesnik, & Urbano, 2002; Inoue & Imoto, 2006). These RFs begin to set up a system for perceiving, not simply luminance contrast, but contrast between different widths and spacing of spatial frequencies. This is the construction of demarcation itself, as bands and striations amidst the firewood start to be distinguished. Mundane instances requiring control of such perceptions occur when studying the wood grain to decide about splitting, or looking through a slatted grill above the fire, or even manipulating the tongs of a fork. Glezer (1995) claims the chief purpose of such complex RFs in the striate cortex is to extract a figure from its background. Again, this is happening with the feature of contrasting spatial frequencies. Other channels are busy collating their own areas of comparable orientation, depth, color, texture, or directional movement, as distinguished against adjoining areas. For instance, depth perception from binocular parallax seems to be constructed by local phase disparity among RFs in striate and extrastriate areas (Anzai, Ohzawa, & Freeman, 1999; Parker & Cumming, 2001; Gonzalez & Perez, 1998; Poggio, 1995). Distinguishing depth creates a perception of obstruction, whether as stacked pieces of firewood, or as flames that flicker and recede from underneath half-burnt pieces of wood. Perceptions along a distance dimension can be controlled via foveal eye vergence movements, which shift the eyes’ point of near-versus-far fixation. In the language of Powers’ (1973) classification, these features would be utilizing various forms of “sensations” to begin to construct “configurations” as a new form of figure-ground invariant. In the analogy of the camp setting, the figure of the fire was set off in several ways, both from the background of

How the brain gets a roaring campfire Chapter | 7

165

the unburnt wood, and from the broader background of the campsite surroundings. And the perceptual status of that foreground fire was definitely a feature the campers were trying to control. There were figure-ground distinctions involving brightness, color, orientation of the flames, directional movement, and perhaps texture. Likely, the bark and grain within the wood were distinguished more by texture than simply a figure-ground difference. However, it is worth noting that a shimmering flame has quite a different texture than does firewood. Highlighting the borders between such textures seems to be the way the shape of those features gets perceived. So then, ignited areas of the flames now have a general shape as well as size. And each piece of firewood also has, in this area of the brain, a perceptible shape.

Motion and transition In the discussion above. about on- and off-signals, the groundwork was laid for the perception of motion. At that stage of retinal and thalamic processing, it was the fact of motion not the rate of motion that was being constructed. For instance, research by Chen et al. (2013) demonstrates the importance of the fast-Off version of retinal ganglion cells for signaling motion onset. Their Adaptive Cascade Model appears able to reproduce ganglion cell firing, whether from sudden movement from the center of the RF, smooth motion across the RF, or the sudden appearance of a dark bar within the RF. Another early mechanism in the visual system is the operation of Y-cells in the thalamic LGN. It appears that the large nonlinear surround of such RFs, with selective inputs from peripheral areas of the retina, allows for signaling movement (Glezer, 1995). A similar mechanism appears to be at work in the peristriate and prestriate cortices, where large RFs distinguishing low spatial frequencies are combined with derivative information from the striate cortex. These then construct perceptions of large size and fast speed (Glezer, 1995), a joint perception that could be especially useful for prey animals evading predators. Perhaps a sudden flare-up from over-dry wood might be an example at the campsite. Such combinations set the stage for further differentiation of figure-versusbackground based upon their relative motions. In general, the construction of a true rate-of-change “transition,” to use the PCT classification, requires both spatial and temporal information, such as distance over time. Kulikowski & Tolhurst (1973), in an influential article, discuss the presence of two types of detectors in the human visual system, which they term “movement-analysers” and “form-analysers,” respectively (p. 159). While movement and form have been found to be constructed by much more complicated functions than two types of neurons in the visual system, nonetheless, the distinction between transient-response neurons and sustainedresponse neurons has been a useful one (e.g., Foster, Gaska, Nagler, & Pollen,

166 SECTION | B Models of brain and behavior

1985; Watson & Ahumada, 1985; Hawken, Shapley, & Grosof, 1996; Perrone & Thiele, 2002). Indeed, those two temporal channels, one responding best to moving patterns e “primarily at stimulus onset and offset (transient),” according to Perrone (2006, p. 11987f.) e and the other to static patterns, seem sufficient to generate a rudimentary perception of speed. This does not mean output in the form of a speedometer. As Perrone (2004) notes, “physiological evidence from a wide range of species indicates that a metrical readout of image speed is not a fundamental property of motion sensors in biological systems” (p. 1735). Rather, the notion is that various neurons in the middle temporal (MT) cortex are tuned to preferred directions and speeds. Perrone & Thiele (2002) have shown that the ratio between the ‘transient’ neural output and the ‘sustained’ neural output, among complex RF cells in the striate cortex, is the key component for generating speed-tuned motion sensors higher up in the temporal cortex. The details of their neuronal motion detectors, called the Weighted Intersection Mechanism (WIM) model, are spelled out above under open-loop ways to study closed-loop phenomena. Here then is a way for the rate of spatial change in a moving image to become part of the brain’s repertoire of perceptions. For instance, this is the level of the brain that can distinguish a smoldering campfire from a roaring one. Later work by Perrone (2005) shows that a particular type of sensitivity parameter, when applied to the ‘transient’ output prior to the ‘transient’/ ’sustained’ ratio function that sets up the WIM model, allows for variable speed tuning, as demonstrated by shifting the slope of the optimal MT responses. This would mean a significant reduction in the striate cortex resources needed for constructing perceptions of speed in the MT cortex. Perrone (2004) suggests the reason this is important: “psychophysical evidence has consistently supported the idea that there are just two (or at most three) temporal ‘channels’ in human vision” (p. 1736). The resource-conserving economy of the WIM model is built upon taking these two channels, a lowpass one for static images and a band-pass one for moving images, and combining them in terms of a ratio. In widely cited early research, Maunsell & Van Essen (1983) show that many MT neurons demonstrate tuning for depth disparity, in addition to having preferred speed and direction of movement tuning. Thus, the middle temporal cortex “is well suited for the analysis of motion in three-dimensional space” (p. 1148), such as tracking the 3D movements of the flames.7 Palanca & DeAngelis (2003) show that motion is not absolutely essential for MT neurons to signal depth via binocular disparity, even though movement disparity increases discriminability of depth. This is especially the case during the first 200e300 msec of response of MT neurons, which they note is “a normal time

7. There is a brief discussion of self-motion and of motion parallax in the online version of this chapter.

How the brain gets a roaring campfire Chapter | 7

167

interval between saccades” (p. 7647). So here, the form of movement can just be via the eyes, with their focus jumping from the near to far side of the fire, while examining where firewood might be placed. Thus, saccadic scanning of static images still demonstrates a depth component at that middle temporal level of the brain.

Modeling with spatial frequencies In order to discuss the next methods of perceptual processing in this survey of Glezer’s formulations, it is necessary to describe in more detail the potentially confusing notion of spatial frequencies. Campbell & Robson (1968), in a classic and influential article, proposed that the human visual system may process size and contrast patterns of luminance in terms of sensitivity to spatial frequencies. Any visual scene has contrasting patches of light and dark. Spatial frequency refers to how many cycles of light-to-dark shifts occur across the image, relative to a fixed visual angle. An analogy from a different sense modality might help to picture what is going on. Moving one’s fingers across a computer keyboard gives a sensation of smooth areas alternating with ridges. Cascio & Sathlan (2001) demonstrate that both spatial and temporal cues combine in discriminating roughness, when there is tactile movement across gradients of varying ridge width and groove width. When it comes to a computer keyboard, the tips of the fingers even detect slight concave depressions in the center of the keys, which is part of controlling for having one’s fingers aligned properly. If while typing there is a sensation of pushing down on the edge of a ridge, it is likely that an error will appear on one’s monitor as an extra key gets pushed. The brain is controlling the position of the fingers with regard to tactile spatial frequencies, embodied in the alternating gradations of texture in moving across the keys. In the same way, it appears that the receptive fields of the brain’s visual system can tune into the alternating contrasts of illumination as the eyes scan across a scene. Using Powers’ (1973) nomenclature for that phrase, illumination is an “intensity,” contrast is a “sensation,” and the alternating pattern is a “configuration.” The manner of researching this has been to use gratings of alternating light and dark bars. Wide bars amount to a low frequency of alternating cycles per specified degree of visual space. Narrow bars constitute a high frequency of cyclic contrasts, since more narrow bars than wide could fit within the same visual angle (or the same size RF.) Such bars may be oriented vertically, horizontally, or at various in-between angles, and of course most naturalistic scenes have combinations of multiple orientations in their illumination patterns. If each bar is smoothly graded away from the center as to brightness, the pattern perpendicular to the bars takes on a sinusoidal profile. Research subsequent to Campbell & Robson (1968) seems to have confirmed the presence

168 SECTION | B Models of brain and behavior

of neurons or neuron clusters utilizing such forms of encoding (e.g., Blakemore & Campbell, 1969; Blakemore, Nachmias, & Sutton, 1970; De Valois & De Valois, 1988; Klein, 1992; Plainis, Parry, Panorgian, Sapountzis, & Murray, 2009). Glezer (1995) makes the claim that, whereas in subcortical areas visual processing preserved a spatial retinotopic representation among the RFs, in the cortex there is a transition to a distributed representation in terms of spectral frequencies. Klein (1992) emphasizes the need, given constantly moving input on the retina, for the brain “to recode the information from the position domain to the relative position (size) domain to remove the effects of image motion” (p. 11). The reason these are significant findings for how perceptual input functions may get constructed is that they allow computational modeling in terms of Fourier analysis.8 Fourier analysis demonstrates that any wave form can be modeled as the sum of a series of sinusoid waves, with different periods from crest to crest, superimposed on one another.9 This would be a sizable reduction in dynamic complexity, for analyzing the brain. Superposition of inputs upon neurons representing a common receptive field is itself a basic property of neuroanatomy that may get exploited here. Each Fourier term would have three parameters, as follows (Lehar, undated). Spatial frequency is the number of full light and dark cycles, per degree of visual angle. This corresponds to the period of each wave. Magnitude refers to the contrast in brightness between the lightest and darkest peaks of the sine wave. And phase is the relative lateral offset of a fixed point on the wave compared to a reference point. There is also a separate term of average brightness across the image, considered to be zero frequency (i.e., there is no change in brightness within that term.) The point needs to be emphasized that this is a process of comprehension through modeling. This part of the brain is not conducting a Fourier analysis in a literal sense; it is being modeled as if that were the case. As discussed above, neurophysiology is often creating mathematical models of biological models of physical models. The starting point is the dictum attributed to Korzybski: “The map is not the territory” (Bateson & Bateson, 1987, p. 21). So then, Fourier analysis is a mapping of how multiple sinusoid waves could be understood, which is a mapping of visual contrast gradients by orientation, which is a further mapping of the distribution of brightness entering the eye. And if the premise of this chapter is valid, there are a number of sub-maps intervening at each step along the way.

8. For those less familiar with Fourier analysis, a useful introduction is provided by Steven Lehar (undated), “An Intuitive Explanation of Fourier Theory,” available at: http://cns-alumni.bu.edu/ wslehar/fourier/fourier.html. 9. This is not to suggest that there is some receptive field in the brain that subjectively looks like a Fourier transform or an inverse transform. It simply notes that Fourier analysis may be used as a modeling tool, to manage the near intractable computations of neural functions.

How the brain gets a roaring campfire Chapter | 7

169

The basic argument here is that different aspects of RF properties in the visual cortex capture these features which are here called parameters. With such features, it is possible to deconstruct10 a visual image: (a) into different spatial frequencies, (b) existing with different levels of contrast, (c) sensitive to different angles of orientation, and (d) laterally arranged at different phase offsets.11 These aspects are then combined through superposition at a higher level of processing. Low frequencies of changing brightness capture large uniform areas of light and dark, but at the cost of blurring the contours. High frequencies capture the sharp contrasts along edges, including those of small crisp details, with little response to broad gradients of transition within the image.12 In the campground setting, the bright but blurry status of neighboring campfires would register via low spatial frequencies, because their entire image at that distance might fill the visual angle. Over and against that level of perceptual detail, it seems the campers wanted their own close-up version of a fire with sharp, flickering points and angles, utilizing higher spatial frequencies, so as to enjoy the shifting and dancing contours. A more detailed example may be helpful to get a better feel for the parallel processing of such frequencies within an image. Consider a page of text in a book or on a computer screen. When viewed from a distance, it may only be the paragraph structure relative to white space on the page that stands out. This is the use of low spatial frequencies to detect a block of text, and slightly higher spatial frequencies to detect that it is comprised of semi-black lines of text interspersed with white spaces above and below the lines. These are primarily noted with regard to a particular orientation. To detect that the words themselves are separated by spaces across the line takes a different orientation. To detect that those words are comprised of letters requires even higher spatial frequencies distributed across multiple orientations, because the white space portion for many letters often occurs inside the letter proper. For each of the space-versus-object distinctions listed above e i.e., paragraph, line of text, word, individual letter e the spatial frequency across a given visual angle has a different period. It would also take different phase offsets to detect that certain letters project above or below the customary 10. Deconstruct is not quite the right term here, because it gives the impression that the brain captures a visual image and then deconstructs it into these parameters. Actually, the perceptual action is in the opposite direction. The brain actively constructs illumination input layer-by-layer, creating features by means of the receptive field properties at each respective level of processing. These are then combined by superposition onto higher level modules, which represent the broader image. 11. The online version of this chapter includes as “Fig. 7.1” a simplified way to visualize these different parameters. 12. Pictures of how the Mona Lisa’s face changes, relative to filtering out different spatial frequencies, can be found at: http://www.mindsmachine.com/asf07.04.html (Watson & Breedlove, 2012).

170 SECTION | B Models of brain and behavior

height for most letters. So, the image is processed with a mosaic of RFs tuned to different frequencies and orientations. The high spatial frequency needed for the narrow lines comprising individual letters serves as a fundamental wave of particular amplitude and cycle, supplemented with its own harmonics (explained in the next subsection). The odd harmonics in particular sharpen the edges e in every which direction e so the individual letters may be read. All of this is simply the preliminary visual processing in the cortex, before any sense of the actual meaning in the words can be addressed.

Handling stark edges To consider the matter of perceiving in terms of visual spatial frequencies is to realize that people see gradients of light and dark shadings all the time. Graded contrasts are the norm in most visual scenes. At the campsite, surfaces away from the campfire were generally darker, whether rocks or trees or areas in shadow behind the tent. They became progressively lighter as their surfaces curved in the fire’s direction. Such gradients are continually flowing into one another, and indeed are constructed by the overlapping and adjoining frequencies of light intensity and contrast, which are then compiled into additional perceptual functions in the brain. What may not be realized, because it is so habitual and automatic, is that the visual focus of attention is generally segmented and circumscribed. In other words, not every collection of gradients is noticed at the same time. Rather, one focuses on a portion of the visual field, experiencing a group of spatial gradients together. It is this that comprises what could be called a visual “Event” perception, to use Powers’ (1973) classification. A stark edge is described below as a particular example of this composite perception. There is a further advantage to the visual system processing in terms of spatial frequencies. If the light energy coming into the eye is not distributed according to a perfect sine wave, (which would be very unusual), then some of the energy would be piled up in harmonic components relative to that fundamental frequency. Harmonics are waves of increasingly shorter period that nonetheless complete their respective cycles at the same endpoint as the original cycle (which is also called the first harmonic.) So the second harmonic completes two cycles within the wave length of the original, while the third harmonic completes three cycles within that same span, and so forth. To perceive in terms of harmonics is a further refinement of an image partitioned into spatial frequencies, because it deals with frequencies relative to one another. A wide grating and a narrow grating have no intrinsic relationship, whereas a third harmonic gets its very definition vis a vis the fundamental frequency. Standardizing the components with reference to the fundamental period allows certain features of the image to take on greater salience. What this means is that the energy (brightness in this case) can be modeled and partitioned using integer multiples of the initial frequency. This offers

How the brain gets a roaring campfire Chapter | 7

171

succinct ways to construct even non-sinusoidal functions, such as a saw-tooth wave or a square wave, which might approximate how brightness is distributed near a sharp edge. Whenever an edge is formed via two diverging planes, there is often a line of maximal brightness (or maximal shadow) right along the intersection of the two planes. As the eye scans in either direction away from the edge, there is often a changing gradient of brightness. Such a sharp linear corner is not the typical way that a sine wave gets viewed. Nonetheless, a superposition of sine waves with decreasing periods can basically simulate even that kind of stark edge. For instance, the superposition of a sine wave plus its next four or five harmonics can give the appearance of a saw-tooth wave. And a sine wave fundamental together with two or three of the odd harmonics e i.e., including periods at one-third, one-fifth, and perhaps one-seventh of the fundamental frequency e is enough to approximate the square wave of a step function. The example of a square wave is intriguing. Glezer (1995) presents a simulation testing a Gabor filter weighting function e i.e., a sinusoid function modulated by a bell-shaped Gaussian function e for complex modules of the visual cortex. When tested against square wave inputs of dark and light bars, not only does the module’s output show the characteristic semi-squared off waveform of a fundamental with its third harmonic. At the edge of each contrast demarcation, “Mach bands appear” (p. 86), as small horns of additional brightness or darkness. Mach bands are a perceptual phenomenon, created by lateral inhibition in the visual system, where edges are heightened by faint lines of additional contrast. It is striking that Glezer’s model of superposition of two harmonic components yields a feature that represents Mach bands. So it seems RFs capturing and then summating a range of spatial frequencies and harmonics can handle sharp boundaries in the visual field that look nothing like smooth gradations arranged in sine wave distributions. According to Glezer (1995), the higher harmonics are engaged in situations of high contrast, which would certainly be the case as a campfire flares up and casts stark step-like shadows on and beyond the surrounding rocks of the fire ring. In the model Glezer (1995) presents, the visual features extracted at lower levels become in effect spectral coefficients for what is known as a piecewise Fourier analysis of the image, as it is processed at successive levels on up to the inferotemporal cortex (ITC) of the brain. There is significant analytical gain if nothing else to this formulation, by simplifying what can seem like an intractable complexity in the workings of the brain. A Fourier analysis model offers a concise way to conceptualize the functional dynamics in terms of a small number of concepts and parameters. For instance, Glezer (1995) uses spatial-frequency terms to distinguish between the functions of detection and recognition as follows: “The first harmonic with a broad bandwidth is enough for detection. The narrowly tuned harmonics are needed for recognition” (p. 102). In the setting of the campground, this speaks to distinctions between

172 SECTION | B Models of brain and behavior

detecting vague pockets of light cast by distant campfires in the growing twilight, versus analyzing the singular character of this campfire, with high frequencies spelling out the best angles for positioning the mess-kit pot filled with water. This is the scheme that emerges from the discussion so far. Returning to the analogy of pixels, there is a massively parallel processing of perceptual input, horizontally at each level of the nervous system and vertically as the input moves between levels. Horizontal processing can be pictured as occurring in layered sheets, whereas vertical movement can be thought of as successive filtering of the input, as it moves up the layers from the retina to the thalamus and then to various areas of the cortex. This arrangement is schematically portrayed in Fig. 7.1, entitled Levels of Neural Processing, based on neurophysiological data and models presented by Glezer (1995).

Objects recognized by harmonic composition Glezer (1995) proposes a kind of successive layering in the visual cortex, where contrasts in the striate area, textures in the peristriate area, and harmonics in the inferotemporal area each get processed in a distinctive manner. He highlights the import of the multichannel, spatial frequency model of Campbell & Robson (1968), as follows: “The perception of a complex stimulus is defined, according to this model, by the contrast of harmonic components, rather than by local contrast” (Glezer, 1995, p. 89). For instance, the way the harmonics pile up for a rounded cup are quite different from the starkedge harmonics of a book. As perceptual signals functionally ascend in the neocortex, they get filtered in terms of relative spatial frequencies, i.e., harmonics, rather than absolute frequencies. When these are collated by modules at the higher levels, they comprise images with unique harmonic compositions. In other words, they become visual objects that cohere harmonically, in spite of other perceptual transformations going on. For example, orientations will vary as an object is rotated. But the same relative positioning of certain harmonics at a certain orientation will still register in tight concert, even if it is by a different neural module than the original one. According to this proposal, it is the harmonic composition that allows for object recognition, which remains invariant despite other apparent changes of size, rotation, or depth. On the surface, this seems to be what Powers (1973) would call a “configuration.” However, that designation has proven useful at a lower level, to capture what is happening perceptually when a figure is differentiated from its background, whether because of color, texture, motion, or depth disparity. Perhaps a different term such as “Form Invariance” is needed at this present level, to convey the fuller sense of object recognition that Glezer talks about. After all, invariance with respect to size, rotation, or depth necessitates the perceptual system registering those variations and then disregarding them with

How the brain gets a roaring campfire Chapter | 7

173

regard to a higher level perception being constructed. To continue the campground analogy, it is still the same campfire, and recognized as such by its general shape, collection of colors, motion relative to its immediate background. This is despite its constant flickering, being seen as smaller when backing away from the smoke, its shifting orientations when approached from a different angle, or partial obstruction when leaning forward to get a pot of water boiling for coffee. The harmonic composition of “The Campfire” remains the same, or slowly evolves, despite other forms of perceptual transformation from those other vantage points. The ever changing campfire is the same ongoing campfire, which the campers are utilizing for their higher level purposes. Some of those higher level purposes get enacted in terms of the spatial relationships between objects within the campground scene. Gathering firewood, stoking it atop a fire, positioning a pot of water, drinking from a cup, these all involve controlling spatial and sequential relationships. This is where other areas of the brain’s visual system come into play. An influential hypothesis is that the brain processes visual information via two semi-distinct pathways, a ventral stream and a dorsal stream (Ungerleider & Mishkin, 1982; Mishkin, Ungerleider, & Macko, 1983; Van Essen, Anderson, & Felleman, 1992). The ventral and dorsal streams have been termed the “what” and “where” pathways, respectively, with object recognition occurring via the former, and spatial relations occurring via the latter (Ungerleider & Passoa, 2008; Kreiman, 2008). Glezer’s (1995) proposals above about functions leading to object recognition have essentially been detailing the workings of the ventral stream, as projections pass from area V1 of the striate cortex, through areas V2 and V4 of the peristriate and prestriate cortices, and on into areas of the medial and inferior temporal lobe. This occipito-temporal path seems to identity “what” an object is. There is also a dorsal, occipito-parietal path dealing with “where” an object is, in terms of spatial relationships including visual guidance of movements (Ungerleider & Passoa, 2008). Whereas classification and object recognition take place within the inferotemporal cortex (ITC), concretization of the spatial relationships takes place within the posterior parietal cortex (PPC). Glezer (1995) raises a telling distinction between these two functions: “When a learned figure is transformed, the invariance mechanism indicates that this is the same figure, whereas the mechanism of concretization discerns a difference” (p. 147). In other words, the ITC says it is the same, while the PPC says it has changed.

Sensorimotor coordination: This way out This chapter now begins a slow pivot toward the matter of output functions for the brain. The sense here is not that of I/O devices, where every link along a causal chain receives an input and provides an output. Rather, the notion is that

174 SECTION | B Models of brain and behavior

of output functions that can operate as part of a closed loop of control. The discussion in the previous section of this chapter has been portraying the steady ascent of perceptual input functions as they get constructed at successively higher regions in the brain. That perspective is admittedly partial, because in a PCT hierarchy of control, any given perception once constructed can be subject to preferred values for that perception, generated by a level higher up. That is the purpose of sending a reference signal back down the hierarchy, to compare against the current state of the perception and generate an output around the loop geared to change it to some preferred state. Such can happen at any level in the hierarchy. The discussion in this section is simply more explicit about turning the corner and sending outputs back down the hierarchy. It shows in greater detail the role of self-generated movement in affecting and altering the perceptions the brain is receiving. Active search and engagement with one’s environment is the norm, rather than the exception. The reason this part of the discussion has been reserved until now is that it is easier to visualize and consider the perceptions that the temporal and parietal cortices deal with, in order to see how movement must be involved in the process. Not until the temporal cortex does full object recognition happen, and not until the parietal cortex do spatial relationships come into play. These are the “what” and “where” pathways, respectively, just mentioned. Objects arranged in space are easier to imagine than spatial frequencies or gradients of contrast. Getting the body to do something is the main method that higher levels of the brain have of bringing about the perceptual input they are specifying. And so, the instructions that are relayed down the hierarchy about what to perceive13 must eventually get structured into motor-related perceptions, such as head orientation, posture, joint angles, muscle lengths, and muscle tension. The level of spatial relationships is a key place for getting a glimpse of this sensorimotor character of the output pathway. By definition, an object’s relationships are constituted with regard to something else. So the question becomes, relative to what? Spatially, it could be relative to other things in the external space, or relative to the body. It is worthy of note that the parietal cortex of the brain appears to jointly encode these two dimensions, both an object’s location in space and its direction from the body. 13. This is the sense used in this chapter for so-called “motor commands.” Because any given motor action must be combined with the effects of gravity and other disturbing forces, including many not known ahead of time, it is questionable whether the brain predicts the precise amount of muscle contraction needed for a given effect. It is also unnecessary to predict those amounts. It is sufficient to determine the sign of the output and then measure the net effects perceptually, adjusting in real time as one goes along. There are many intervening perceptual levels between the output of a given control loop and its eventual emergence as muscle tension exerting a force in the environment. So commands at every level are not instructions about what to do, but instructions as to what the next lower level is to perceive. This is a major distinction between PCT and many theories of motor action in the literature.

How the brain gets a roaring campfire Chapter | 7

175

As delineated in my online chapter, the thalamus may be an important transit point for references, to help calculate error signals for projection down the hierarchy. It seems that the output path then proceeds by way of the basal ganglia. These two sets of structures may be initial locales for the brain’s “This Way Out” exit sign. There is debate over the piece that the basal ganglia may contribute to the process (see online version of this chapter). It appears that the striatum of the basal ganglia stands at the head of a motor hierarchy, either orchestrating subsequent steps for motor programs, or in a more specialized role vis a vis control of transitions for body positions further below that. As error signals percolate down such a motor hierarchy, it may be that each descending level adds its own constraints on the degrees of freedom for what is ultimately possible behaviorally.

Perceiving objects and their relational flux The role of the posterior parietal cortex (PPC) appears to be that of constructing and delineating the spatial relationships between the objects distinguished by the ITC, and uniting those objects and patterns into a scene. Glezer (1995) notes, “The cells of the parietal cortex function as a command apparatus for investigating the external world by hand and eye” (p. 179). This deals with the level of perception that Powers (1973) calls “Relationships,” but it also draws in a higher programmatic level. An important clarification needs to be made at this point. Many of the neurophysiological properties and functions described to this point have been investigated by variations on a particular research paradigm. While the animal is passively or chemically restrained in some way, with micro-electrodes carefully placed in specific areas of the brain, various lines or bars or gratings of light and dark bars are systematically moved across the animal’s visual field, and the neural responses are recorded and analyzed. But of course, this is not how animals themselves do it. They do not simply wait for a world of illumination to pass in front of their eyes. Rather, they actively scan the visual world. This is the insight noted by Dewey (1896/1998), in his classic critique of the reflex arc concept: “(T)he motor response determines the stimulus, just as truly as sensory stimulus determines movement. Indeed, the movement is only for the sake of determining the stimulus, of fixing what kind of stimulus it is, of interpreting it” (p. 151). So then, most animals including humans have various means of movement available to them, to actively engage in searching or scanning their world. Among eye movements alone, a variety of methods exist. There is focusing the lens of the eye, by adjusting its curvature to fine-tune focal length according to distance. There is the so-called pupillary reflex, which adjusts the illumination level for optimal contrast. There are vergence eye movements (convergence or divergence), where the eyes rotate in opposite directions to

176 SECTION | B Models of brain and behavior

adjust the depth of their binocular fixation point. There are random saccadic eye movements, while the image on the retina is temporarily paused, for sampling the visual field. There are also intention saccades, which appear to give extra attention to areas of high contrast.14 And there are smooth pursuit eye movements, for following sensations that are moving within the visual field. Each of these is a process for actively engaging and affecting the content of the stimulation entering the eye. In that respect, they are each enacted by means of negative feedback control loops. Glezer (1995) proposes that object recognition itself consists of a twofold process, initially structured by innate mechanisms for extracting simple features, but then proceeding as sequential scanning by way of learned complex features. This is consistent with what Loftus & Harley (2004) more broadly call “global-precedence” theories of information acquisition, with a temporal distinction occurring, where first the low-spatial-frequency “global information is acquired initially, followed by local information” (p. 105). The latter aspect in Glezer’s (1995) proposal amounts to “a successive calculation of the parameters of the discrimination features” (p. 164), essentially checking the perception against learned features. This second process appears to be regulated by the PPC. So even as spatial relationships are being worked out in the parietal cortex, the brain is asking the question, ‘Relationships between what?’ Glezer models it as a rapid sequential “movement along the decision tree in which the links of the complex features are checked at each node” (p. 168). This notion of checking each node against a condition makes it into an activesearch, template-matching control process. It is worth noting that such a description is very similar to what Powers (1973) proposes as the structure of a “Program” form of perception, comprised of conditional decision nodes, sequentially carried out. Programs having that structural feature can be enacted according to various time scales. The program Glezer (1995) envisions for object recognition starts with using low spatial frequencies to allow fast orienting and “crude, schematic description of the object” (p. 171). Subsequent steps incorporate higher frequencies, which aid in recognition but require more time for extracting information, because the right modules for analyzing the harmonics must be settled on. The additional time needed is an indication that this form of perception happens at a higher level than the figure-ground distinction of simple configurations. Finding the right modules is thus key for this process. And under the spatial-frequency hypothesis, this means finding modules that capture each set of harmonics involved with each object within a scene.15 This appears to be 14. Young & Illingworth (1998) present a PCT-informed model for fixation control, based on simple contrast or multi-colored input. 15. The combination of an internal and external active search process is described in the online version of this chapter, where movement helps to filter the visual input, while memory templates are searched by successive approximation until a suitable perceptual match occurs.

How the brain gets a roaring campfire Chapter | 7

177

where the ‘re-cognition’ part of recognition comes in. Partially formed patterns (i.e., Glezer’s ‘crude, schematic descriptions,’ mentioned above) are checked against previously learned patterns to see if the fit is right. The brain successively recalculates parameters for discrimination as the evidence accumulates. Here, action and stimulation continually feed one another, as Dewey noted. For instance, fast-acting saccades of the eye keep refining the visual input. Intentional saccades gravitate to areas of higher contrast to keep gathering potentially relevant details. A variety of spatial frequencies and harmonics arrayed in various orientations, and embodied in modules tuned to those several features, are successively checked against the data. Such is the picture that emerges from Glezer’s presentation of the ITC and PPC mechanisms, to generate a sense of form invariance. In sum, what begins as mere feature extraction continues as an active search process.

Blended frames of reference In parallel with the ITC, the PPC of the brain, according to Glezer (1995) “creates extrapersonal and intrapersonal space using frames . that enable one to describe relationships within situations” (p. 242). That is to say, it can construct and perceive relationships both relative to the body and relative to the external space itself. It is with the PPC that visual input starts to take on an active sensorimotor character. Programs for clarifying what is being seen are generated, as well as programs for specifying and altering the spatial relationships within the visual scene, among those objects that have been identified via the ITC. In working with those spatial relationships, positions with respect to oneself are especially important, captured in the phrase, eyehand coordination. So then, it appears that the PPC provides mechanisms for carrying out visual relationship control. Several researchers have explored the nature of the coordinate systems being encoded in the parietal lobe (Andersen, Essick, & Siegel, 1985; Pouget & Sejnowski, 1995). There is evidence that the PPC handles the spatial arrangements in terms of multiple coordinate systems (Xing & Andersen, 2000), integrating them with eye-centered, head-centered, and body-centered representations. The issue is this: “Since the positions of the eyes and head vary from moment to moment, the retinotopic locations of targets likewise change, although their spatial locations with respect to the head or body may not change” (Andersen et al., 1985, p. 456). Therefore, the integration of sensory and motor components being initiated in the PPC needs to be sensitive to any and perhaps all of the following: position of images on the retina, what direction the eyes are gazing, and the angle of head position for either visual or auditory variations in the perceptual input. It appears that these requirements could be met in terms of PPC neurons displaying receptive fields that are multiplications of other functions. For

178 SECTION | B Models of brain and behavior

instance, Andersen et al. (1985) analyze RFs in parietal area 7a, deriving what they call gain fields. These are formed as the product of a bell-shaped Gaussian function, a common neural function representing retinotopic RFs, and a linear slanted planar function, representing how the response of the neuron changed depending on the direction of eye gaze. Thus, in the parietal lobe, there are cells that jointly encode location of a stimulus and direction of gaze, a combination that would be important in tuning and coordinating subsequent movements. Pouget & Sejnowski (1995) utilize a similar approach to modeling parietal neurons, but with a sigmoid function (roughly an S-shaped ramp) replacing the planar slope, to represent eye position. In their case, multiplying the sigmoid function by the Gaussian function, distributed over a large set of neural RFs, leads to what they call basis functions, which when plotted resemble the front portion of a saddle. They note, “A basis function decomposition is a wellknown method for approximating nonlinear functions, which is, in addition, biologically plausible” (p. 161). They suggest that such transformations are needed “downstream of the parietal cortex in the premotor cortex and superior colliculus, two structures believed to be involved in the control of, respectively, arm and eye movements” (p. 163). In a later article, these same authors describe the import of their approach this way: “(P)arietal neurones (sic) compute basis functions of their inputs and, as such, provide a representation of the sensory inputs from which motor commands can be computed by simple linear combinations” (Pouget & Sejnowski, 1997, p, 1451). The sensorimotor character emerges as essentially a combined mapping of visual signals and posture signals, whether eye positioning or head inclination, in a way that facilitates coordinated control of eye gaze direction along with head and/or body orientation. It is tempting to consider that the sigmoid portion of the basis function would represent the acceleration and deceleration that would be needed for any smooth movement of the joints. However, that would be leaping over many intervening layers of controlled perceptions, as discussed in an early section of this chapter. A better approach is to consider how the situation looks through a PCT microscope. First, consider the nature of PCT output functions. It is true that both acceleration and deceleration are needed for smooth biological movements. But those dynamics emerge as offshoots of the standard way negative feedback control loops are set up, using proportional-integral (PI) controllers in their output functions. Specifically, the integral lag of the PI function creates the delay in getting acceleration going. This is because the error signal needs to build up an initial history as the output gets started. In a similar manner, output is also proportional to the error signal. So as error decreases upon a perception’s approach to its final reference value, there is a natural deceleration built in to the very way the PI controller operates. If a sigmoid function is not needed for those aspects of low-level perceptual control, what then is its purpose? Consider a PCT

How the brain gets a roaring campfire Chapter | 7

179

conceptualization of the types of perceptions that may be controlled at these levels in the brain. Glezer (1995) presents evidence that the posterior parietal cortex handles spatial relationships. What neural components might be part of how a relationship is constructed? It appears that a sigmoid function could represent relative degrees of co-variation. The asymptotic top and bottom portions of the sigmoid curve would seemingly represent ‘achieved relationship’ versus ‘no relationship’, respectively. As Powers (1990) notes, “Any relationship is a variable attribute; as we place a cup ‘on’ a saucer, the ‘onness’ grows as the cup nears its final position, being correct when we release the handle” (p. 74). Such a conception is consistent with “categories” being the next higher level of perception, according to some PCT proposals (Powers, 1979, 1990), as listed in Table 7.1 above. The category level is where class membership is constructed, and where either-or determinations are made. A binary step function, representing 0 or 1, is a useful way to consider coding such classification schemes. It can even be combined with hysteresis, implemented by lateral mutual inhibition at the boundaries of category space, so that reclassifications are not made until excess evidence has been built up. In any event, categories are fruitful ways of providing reference specifications for relationships, the next lower level in a PCT listing. For instance, “combine ‘this’ with ‘that’, along ‘such-and-such’ a dimension.” It is worth noting that a sigmoid function (at the relationship level) can approximate a step function (from the category level) for outputs that require time for their enactment. Indeed, “the integral of any smooth, positive, ‘bump-shaped’ function will be sigmoidal” (Wikipedia, 2016, Sigmoid function), as may be produced with a proportional-integral output function. Consider one more aspect of a hierarchical arrangement of PCT types of perception. While “categories” provide reference specifications for “relationships,” the latter provide references for what are called “events” in PCT. “An event is defined as a unitary package of transitions, configurations, sensations, and intensities, having a familiar pattern in time” (Powers, 1990, p. 72). Examples within the present context might include: ‘rotate the head,’ ‘fixate the eyes,’ ‘pull something closer.’ It is a broad targeted movement, with details to be specified by lower levels as they carry out their portion of the control task. The reference for such events needs to initiate the instructions and keep them in place until completion. Plotted over time, a sigmoid function essentially moves from a low initial asymptote of ‘no need for change,’ through sustained output on the slope of the curve, and to a high asymptote of ‘no need for further change.’ The slope guarantees that some error will persist until the completion of the event. Returning to the matter of basis functions discussed above, while the sigmoid portion may initiate a coordinated perceptual event at the next lower level, the Gaussian portion would make it a spatially targeted event. For instance, not just ‘rotate the head,’ but ‘rotate the head there.’ The foveal

180 SECTION | B Models of brain and behavior

center of gaze is well represented by a bell-shaped reference function, as is the head or body’s directional orientation. Any of these systems could be supplied with a target via a Gaussian function, with a network of different PPC neurons specifying the range of targets within the visual field. So then, the same set of parietal cells can simultaneously provide reference outputs for different body systems, to coordinate the targeting of head, eyes, and body.

Summary and implications It may be useful at this point to summarize some of the findings and implications of this chapter in light of categories of perception proposed by Powers (1973). Such a comparison is presented in Table 7.3. This enumeration lists two dozen forms of visual input functions, many of which have been described by Glezer (1995). They are listed in groups showing the best approximation according to a PCT classification system. A brief indication of the neurophysiological mechanism is provided for each one, based on the more detailed discussion in the body of the chapter above, along with the general locale within the nervous system. An example of how perceptual control may be enacted is also offered for these visual functions. It is worth noting that causing one’s attention to purposely linger upon a perceptual feature is still an instance of control, because it creates a temporal stability that goes beyond what would be expected by chance. Finally, a perceptual example present in the campfire scenario is listed for each one. Table 7.3 captures some of what was discussed above about different forms of reliability. The column labeled “Mechanism” summarizes various forms of structural reliability, in terms of the neural properties used to construct different perceptual functions. The column labeled “Example of Control’ gives instances of enacted reliability, using behavioral means to bring about certain values of a perceptual function out of the range of permissible values. From the standpoint of the person, structural reliability governs what can be perceived, while enacted reliability governs what will be perceived. The interlaced headings according to a PCT classification scheme are meant to convey some of the emergent reliability that may arise hierarchically, as qualitatively distinct types of perceptions get constructed. There is admittedly more gray area in this last designation, than in the former ones. What PCT would call intensities and sensations make up the first third of this table, and one is barely beyond the retina. This shows that a lot of neural processing can take place with just a few layers of cells. The middle third of the table comprises what PCT would probably call configurations, primarily occurring in different areas of the occipital cortex. In an earlier section of this chapter, it was shown how the first nine types of visual function might be related to perceptual examples from the campfire scene. These are shown in the rightmost column of Table 7.3. To simply get to that configuration stage of PCT designations merely allows perceptions such as the following to be

TABLE 7.3 Summary of Visual Functions and Alignment with PCT Perceptual Classes. Comparison of visual input functions, particularly ones described by Glezer (1995), and their approximate equivalents according to a partial PCT classification of perceptions proposed by Powers (1973). Potential neurophysiological mechanisms, approximate locales for the perceptual functions within the nervous system, examples of perceptual control, and instances from the campfire illustration are suggested. See the body of the chapter for details as to neural mechanisms. Construction of visual functions

Mechanism (Structural Reliability)

Example of control (Enacted Reliability)

Campfire instances

Retinal photoreceptors

Pupillary reflex, Squinting

Faintness (of original fire)

Location

Visual Examples of PCT Intensities Brightness distribution

Graded potentials, Hyperpolarized membranes, Integrated responses

Visual Examples of PCT Sensations Visual reliability (signal-to-noise sensitivity)

Difference of Gaussian RFs, Excitatory center & inhibitory surround, Disinhibition signal

1st Retinal triad

Foveal attention

Clear versus indistinct

3

Chromatic detection

Opponent process of cone responses

Retina

Color differentiation

Orange versus yellow

4

Relative brightness (construction of contrast)

Ratio of reflectances (object vs. background)

Retina

More saccades to the borders between patches

Patches

5

Adjusted sensitivity despite changing illumination

Light/dark adaptation (automatic gain control), Pigment breakdown, Synaptic vesicle depletion

Retina

Reposition relative to glare of sun, Stay in dark prior to starwatching

Twilight

6

Onset/Offset (local change in intensity)

On-center & Off-center ganglion cells

2nd Retinal triad

Orienting response

Glow

7

Motion

Fast-Off ganglion cells, Large nonlinear receptive fields of thalamic Y-cells

Retina, Thalamus

Attending to motion

Spark

181

2

How the brain gets a roaring campfire Chapter | 7

1

Continued

Construction of visual functions 8

Light/dark flickering

Mechanism (Structural Reliability)

Location

Example of control (Enacted Reliability)

Campfire instances

Linear on- & off-zones of thalamic X-cells

Thalamus

Watch sun glint on water

Wisp of smoke

Visual Examples of PCT Configurations 9

Orientation

Series of thalamic on-signals, Neighboring inhibition

Thalamocortical projections

Tilt head, Rotate paper

Ascending

10

Low spatial frequency

Gabor function (sinusoid by Gaussian), High threshold

Striate & higher cortical areas

Detecting large relatively uniform area

Size

11

Spatial frequency cycles by orientation

Lattice of Gabor filters

Striate cortex

Gradient discrimination

Shading

12

Contrast of spatial frequencies

Sine wave parameters, Complex RFs, Overlapping phase offsets, Inhibitory periphery of RF

Striate cortex

Study grain of wood, Look through a grill, Manipulate tongs of a fork

Demarcation

13

Depth perception (binocular parallax)

Local phase disparity

Striate & higher cortical areas

Shift fixation depth (eye vergence)

Obstruction

14

Figure versus background

Collating comparable transformations

Striate & higher cortical areas

Differential attention

Fire versus wood

182 SECTION | B Models of brain and behavior

TABLE 7.3 Summary of Visual Functions and Alignment with PCT Perceptual Classes. Comparison of visual input functions, particularly ones described by Glezer (1995), and their approximate equivalents according to a partial PCT classification of perceptions proposed by Powers (1973). Potential neurophysiological mechanisms, approximate locales for the perceptual functions within the nervous system, examples of perceptual control, and instances from the campfire illustration are suggested. See the body of the chapter for details as to neural mechanisms.dcont’d

15

Texture

Difference in autocorrelation

Peristriate cortex

Image differentiation

Wood bark & grain

16

Borders between textures (anomalous contours)

Gradient derivatives, Chains of nonlinear striate inputs

Peristriate & Prestriate cortices

Contour tracking

Shape

Visual Examples of PCT Transitions Large size & fast speed

Larger RFs, Low spatial frequencies, Rate of striate change

Peristriate & Prestriate cortices

Intentional saccades

Sudden flare-up

18

Rate of change (speed & direction of motion)

Ratio of band-pass transient response to low-pass sustained response

Middle temporal (MT) cortex

Pursuit tracking

Smoldering fire versus roaring

19

Motion parallax

Large RFs, Tectal inputs, Pursuit eyemovement signal

Middle temporal cortex

Tracking motion in 3D space

3D relative movements of flames

Visual Examples of PCT Events 20

High spatial frequencies

Filters at harmonic intervals

Striate & higher cortical areas

Outlining crisp details

Points & angles

21

Fundamental & odd harmonics

Modules of harmonic channels

Inferotemporal cortex (ITC)

Mach bands, Edge discrimination

Stark step-like shadows

22

Relative spatial frequencies

Compiled sinusoids, Module superposition, Standardization by harmonic components

Inferotemporal cortex

Surveying scenes blended by scale, orientation, & contrast

Composite changes among the flames

183

Continued

How the brain gets a roaring campfire Chapter | 7

17

Construction of visual functions

Mechanism (Structural Reliability)

Example of control (Enacted Reliability)

Campfire instances

Inferotemporal cortex

Object recognition

“The Campfire”

Posterior parietal cortex (PPC)

Eye-hand coordination

Wood or pot placement

Location

Visual Examples of Form Invariancea 23

Harmonic composition

Successively matching tuned discrimination parameters

Visual Examples of PCT Relationships 24

Spatial relationship

Blended coordinates, Basis functions (sigmoid by Gaussian target)

a Form invariance is not a specific name of one of Powers’ (1973, 1990) levels of perception. It is proposed here as a more complex rendering of the PCT notion of “configuration,” based on Glezer’s (1995) notion that full object recognition takes place via the overall harmonic composition constructed in the ITC region of the brain.

184 SECTION | B Models of brain and behavior

TABLE 7.3 Summary of Visual Functions and Alignment with PCT Perceptual Classes. Comparison of visual input functions, particularly ones described by Glezer (1995), and their approximate equivalents according to a partial PCT classification of perceptions proposed by Powers (1973). Potential neurophysiological mechanisms, approximate locales for the perceptual functions within the nervous system, examples of perceptual control, and instances from the campfire illustration are suggested. See the body of the chapter for details as to neural mechanisms.dcont’d

How the brain gets a roaring campfire Chapter | 7

185

detected: faintness, clarity, color, patches, twilight, glowing, a spark, a wisp of smoke, an ascending direction. This does not cover much of the full experience of the fire. With Table 7.3, it can be seen how additional aspects of the fire may get constructed. For instance, size starts to emerge from the low spatial frequencies of the striate cortex, while it likely takes a lattice of spatial frequency filters to construct shading. The contrast between frequencies is simply the perception of demarcation, and the phase and vergence disparities creating a sense of depth become a perception of obstruction, with one patch in front of another. To actually distinguish fire versus wood requires quite a sophisticated neural compilation of constructed figure versus background signals. So according to the summarization of neural functions in Table 7.3, it takes some fourteen layers of transformation for a fire to have enough perceptual components to be a discernible configuration. However, this is still a fairly static image, since rates of change in the perceptions have yet to be constructed. Continuing with Table 7.3, the actual grain of the wood or bark likely emerges from complicated textural algorithms of autocorrelation differences. And it takes derivatives signaling gradient changes and the chaining of neural inputs along the borders between textures to actually define a perception’s shape. With an actual fire, that shape is constantly changing, where rapid changes from a sudden flare-up rely on the low spatial frequency shifts within large receptive fields. The degree of change itself requires a ratio measure between transient band-pass and sustained low-pass neural signals. Such a neural sensor is needed to notice the rate of oscillation of the flames, for instance whether it is merely smoldering or actually roaring. And perceiving the relative movements between flames at different distances from the eye means adding a directional eye-movement signal to create the sense of motion parallax. Only when all the above perceptual substrates have been constructed does it become possible to outline the crisp points and angles of the shifting flames, using high spatial frequency filters. Even greater heightening of the edges and shadows between different objects in the campground scene comes from modules that collate odd harmonics together. The superposition of relative spatial frequencies (i.e., harmonics) in the inferotemporal cortex is what allows the multitude of orientations and contrasts to be perceived as composite changes among the flames. And the overall harmonic composition that emerges, tuned by successive matching of discrimination parameters among competing modules, becomes the full-fledged object recognized as “The Campfire.” The last row in Table 7.3, together with the discussion in the previous subsection, gives some hints as to further perceptual functions conceptually higher in the cortex. Spatial relationships among the objects discerned in the ITC get laid out in relation to the body’s position, by means of the blended coordinates of basis functions in the posterior parietal cortex. This includes a perception of

186 SECTION | B Models of brain and behavior

wood placement in the campfire, in preparation for proper positioning of a pot of water. The Gaussian component would help to make subsequent eye or hand movements targeted, in terms of direction relative to the body. Any relationships specified by spatial prepositions, such as the pot being placed within the flames, may be represented in the PPC by the sigmoid component in those modules, signifying the degree of co-variation between the pot and the heat. And the very orchestrating of this with that, in setting up those spatial arrangements, may be promoted by binary step functions from a next-higher category level, that is itself carrying out programs of the right sequences for achieving a hot cup of flavored coffee for the campers. Table 7.3 is an expansion and elaboration of what was foreshadowed above in Table 7.2. As such, it condenses an enormous amount of research and modeling on the workings of an adult visual system. Each successive structural layer is built up out of previous capacities already constructed, creating types of perception that Bateson (1979) might call differences that make a difference.16 The neurophysiological evidence presented in the previous chapter as well as this chapter has implications for how to arrange the classifications proposed for a PCT hierarchy of perceptions. Indeed, it appears that the order should be modified slightly from Powers’ (1973, 1979, 1990) formulations. Table 7.4 presents such a modification, with key differences highlighted in bold. This is a revision and expansion of Table 7.1 presented earlier. Examples of each perceptual level are offered in Table 7.4, along with allusions to the respective evidence primarily from Glezer (1995) or from the Hierarchical Temporal Memory model (George, 2008; Hawkins & Blakeslee, 2004), as discussed in these respective chapters. One difference from Powers’ formulation amounts to reintroducing Sequences as a fifth level of the hierarchy, and distinguishing it as a separate level from Events, one level higher. This essentially moves the Sequence level back to its original placement within Powers’ (1973) writings. The reasoning is as follows. While a PCT event is meant to convey a unitary impression with respect to time, it is still a constructed perception. And the elements of its construction are lower level perceptions, usually collections of transitions often requiring a proper sequence for their enactment. The lower-level distinctions must first be constructed, before a higher level can collapse them into a new form of perceiving. Thus, ‘sequence’ is a necessary fifth-order substrate, preceding the emergence of a sixth-order ‘event.’ The evidence admittedly derives more from the HTM model of the previous chapter, than from Glezer’s discussion in the present chapter.

16. An extensive and integrated summary of the mechanisms highlighted in Table 7.3 is presented in the online version of this chapter.

How the brain gets a roaring campfire Chapter | 7

187

The other proposed difference in Table 7.4 is inserting a new seventh level, to be called Form Invariance, with a more elaborated mechanism than the figure-ground distinctions of the Configuration level. Glezer (1995) presents a sophisticated discussion of object recognition, involving fast-operating low spatial frequencies to give a crude initial sense of an object, followed by slower template-matching to learned combinations of high spatial frequency harmonic components. This latter change in the table leads to renumbering the higher levels of the PCT hierarchy, with their total becoming twelve. As emphasized above, these PCT names are simply summary designations that help to conceptually organize what the brain may be doing. Each class may well represent similar perceptions that in fact are generated by quite different neurophysiological mechanisms. Such a list remains a helpful reminder as to the broad range of perceptual transformations constructed by the brain. Because this chapter deals with negative feedback control loops according to a PCT schema, looking at the perceptual side of the loop will always be somewhat incomplete. No arrangement for perceptual control can fully be tested without the output side of the loop sending a feedback connection out through the environment.17 Yin (2014a, 2014b) demonstrates a nuanced understanding of the range of lower level perceptions requiring control to carry out any movement-related goal. According to his view, the basal ganglia form a system that is specialized for transition control. He recognizes that most higher-level systems basically need to enact their forms of control by getting the body to do something in the environment. The basal ganglia are essentially at the head of this final common motor output pathway for the brain. Achieving movement depends on an ascending series of relevant perceptions, which have been summarized by Yin (2014a, 2014b). They include muscle tension (as sensed by the Golgi tendon receptor), muscle length (as sensed by muscle spindles), joint angle (using stretch receptors and other specialized sensors in the joints), balance (using vestibular sensors), and a general sense of body orientation in relation to gravity (using proprioceptive sensors). These contribute to higher level perceptions such as body configuration or posture (as perceived by widely distributed joint angle sensors, together with sensors for effort), head orientation and steering (using visual and auditory sensors for distal stimuli), foveation for the eye and parafoveal angular deviation (using angular maps in the tectum), all of which can be considered forms of position control. The above listing in effect constitutes a specialized motor hierarchy of controlled perceptions, for humans and many other animals. The idea of a “final common path for actions” (Yin, 2014b, p. 19) is seemingly derived from the notion that there are many more degrees of 17. The online version of this chapter includes an extensive survey of what may be PCT output functions, involving gated connections through the basal ganglia, serving as an initial funnel for a common motor output pathway for the brain.

Order (from periphery)

PCT name

Evidence From Glezer

First

Intensities

Graded potentials,

Luminance,

Second

Sensations

Difference of Gaussian RFs, Reflectance ratios, On/Off retinal & thalamic cells,

Detectable signal versus noise, Relative brightness, Sense of motion, Flickering,

Third

Configurations

Spatial frequency by orientation (Gabor filters), Comparable figure/ground transformations, Phase disparity, Texture autocorrelations,

Co-occurring inputs (cortical layer 4),

Figure versus ground, Contrast grating, Depth, Texture,

Fourth

Transitions

Rate of striate change, Band-pass versus lowpass ratio, Pursuit eye signals,

Temporal input patterns (cortical layer 2),

Movement, Speed, Rate of change, Motion parallax,

Fifth

Sequences

Serial order of lateral patterns,

Serial ordering,

Evidence From HTM

Example of perception

188 SECTION | B Models of brain and behavior

TABLE 7.4 Modified Arrangement of PCT Hierarchical Levels of Perception. Starting with Powers’ (1973, 1979, 1990) PCT categories of perception, as listed in Table 7.1 above, certain modifications are proposed, listed here in boldface type. These suggestions are based on neurophysiological evidence supplied by Glezer’s (1995) research into the visual system and by that supplied in support of the Hierarchical Temporal Memory model (HTM, George, 2008; Hawkins & Blakeslee, 2004). There are two key differences proposed here. The level of Sequences has been moved back down to its original placement as a fifth order of perception, and distinguished from the type of perception called Events. A new seventh level has been introduced, called Form Invariance, to capture Glezer’s discussion of full object recognition being based on the harmonic composition of an object in terms of its spatial frequencies.

Events

High spatial frequencies, Harmonic channels, Superimposed sinusoids,

Timing signals from thalamus, Timed sequence,

Edges, Crisp details, Scalable scenes, Time segmentation,

Seventh

Form Invariance

Compiled harmonic profiles,

Spatial objects,

Eighth

Relationships

Sigmoid functions, Blended frames of coordinates, Extra- & intra-personal space,

Spatial relations, Co-variation,

Ninth

Categories

Step function w/ hysteresis,

Members of a class, Conceptual borders,

Tenth

Programs

Checking features along a decision tree,

Contingency networks,

Eleventh

Principles

Heuristic guides, Fuzzy logic,

Twelfth

System Concepts

Integrated resonating systems,

How the brain gets a roaring campfire Chapter | 7

Sixth

189

190 SECTION | B Models of brain and behavior

freedom available for perception than there are for muscle forces exerted on the environment. Nonetheless, if motor output is sequenced or timed appropriately, a wider range of perceptual results can be achieved. And as a lower level perception is stabilized to a sufficient degree, it can become a reliable component of a higher level perception as it becomes stabilized. It appears that the basal ganglia are a key set of structures for achieving the right timing and transitioning of the limited motor-related degrees of freedom.

So how does the brain get that campfire? This chapter has endeavored to show how visual perceptions may get constructed within the brain, at least ones relevant for perceiving a roaring campfire. This is a start toward specifying a range of perceptual input functions, whose values could then be controlled via the principles of Perceptual Control Theory. And in fact, many features of neurophysiological research may be given new significance when considered from this standpoint. The functional map of a PCT control loop, arranged in a nested hierarchy of implementing and contextualizing connections, is a robust structure for surveying neuroscience literature. While admittedly sparse in its organizational design, it nonetheless offers surprising insights into what specific neural structures may be doing in the brain. Indeed, some of the PCT proof-of-principle simulations demonstrate that cascaded perceptual control by means of such a hierarchy is more than just a metaphor. Actual working generative models can be built that way, and so neural tissue could certainly be operating by similar principles. Moreover, when the effects of negative feedback action are measured at the point of perceptual input back into the loop, it is possible to have quite convoluted output paths that still achieve their intended results. For such loops to work properly, the sign of the corrective action is the key requirement. After that, there are some relative timing requirements, such as lower-level loops operating with a faster time constant than higher-level loops. If the overall effect is to reduce the mismatch between a perception and its corresponding reference standard, then negative feedback can produce substantive benefits despite output signals having to traverse a complicated journey down through many layers of control. There is also a complicated journey on the sensory input side of these loops. Following Glezer’s (1995) mapping of the visual system shows that it takes multiple layers of perceptual transformations to set the stage for even the mid-level categories of a PCT formulation. One intriguing finding in the “This Way In” section of this chapter was the importance of perceptual standardization at several points in the visual system. For instance, in the retina sensory input is recoded into relative degree of contrast, rather than being subject to every absolute change in illumination. As it ascends to the thalamus and cortex, it is reconfigured into spatial frequencies, so as not to be dependent on

How the brain gets a roaring campfire Chapter | 7

191

a retinotopic representation. A further standardization occurs in moving toward the temporal cortex, by recoding the input in terms of relative spatial frequencies, i.e., harmonics. Glezer (1995) summarizes a rich literature pertaining to spatial frequencies and makes a compelling case for their role in the brain. In the previous section of this chapter, the beginnings of sensorimotor relations were explored in the parietal cortex, with coordinate systems that captured both extra- and intra-personal spatial orientations. It included a discussion of sigmoid functions, possibly integrating step-function reference signals from a higher ‘category’ level of perception, and conveying them in a sustained way for behavioral ‘events’ at a lower level. The role of movement in effecting changes in perceptual input was also noted, where it seems the basal ganglia stand at the head of a broad motor hierarchy, controlling transitions and timing for position control and other motor-related perceptions below. An important initial impression emerging from these chapters on “How the Brain Gets a Roaring Campfire” is that the functional template of PCT components has a very diverse means of enactment in the nervous system. The brain is an exceedingly complicated set of structures. And yet, overlaying a set of neurophysiological maps with the engineering blueprint proposed by Perceptual Control Theory offers a promising way to explore the brain’s many twists and tangles. So let us return to that initial quiet campground. It is a summer evening, on a warm vacation night. The inferotemporal cortex (ITC) has recognized a campfire, or what is left of one, by the harmonic profile of the semi-bright patches of embers standing out from a darker background. Receptive fields (RFs) in the retina have already started to delineate the dwindling contrasts of the campfire’s distribution of luminance. The on-zones and off-zones of RFs in the lateral geniculate nucleus (LGN) of the thalamus have registered the on/off glowing quality of those embers, whose changes in intensity were first detected by the retinal triad of bipolar, amacrine, and ganglion cells. The embers’ pulsating visual contrasts have been filtered in the complex RFs of the striate cortex, and their shifting orange colors, grainy textures, and flowing borders have been extracted by the algorithms of the peristriate and prestriate cortices. The mostly horizontal orientations of the charred but still glowing wood have been encoded within the vertical functional columns of the neocortex. All of these have become in effect spectral filters or coefficients, segmenting the campfire scene into spatial frequency channels as it is fed on to the ITC. The gradients of shimmering light from the undersides of each piece of firewood constitute radiant bands of different spatial frequencies, which cohere as a whole by virtue of their harmonic components. At the level of the ITC, the existing campfire is recognized in terms of its overall harmonic composition.

192 SECTION | B Models of brain and behavior

Then a mismatch occurs! Orienting movements e with the basal ganglia inhibiting competing non-postural movements, while disinhibiting head and eyes e have caught a foveal visual distinction within nearby campsites. Neighboring campfires, despite being a distance away and falling within a narrow segment of the visual field, seem much brighter than the dying fire in this campsite. Even a low-spatial-frequency first harmonic is enough to detect that! Those blurry patches of distant light are quite vigorous, which is to say, ‘fast-off’ cells in the retina and thalamus are busy indeed. By comparison, this dwindling fire has not enough retinal brightness, not enough thalamic contrast, not enough striate vertical orientation to its spatial frequencies, and certainly not enough rapid motion nor higher harmonics and step functions in the temporal cortex to indicate a burst of stark quivering flames. And so, reference standards are generated from memories of more substantial blazes, and sent from the cortical cells in layer 6 to thalamic reticular and relay areas below, to summon more brightness, and movement, and contrast, and all the other attendant aspects of a soon-to-be roaring campfire. But that is not the whole story. This string of perceptual mismatches along the ventral occipito-temporal visual stream, called the “what” pathway, becomes part of a more profound mismatch. A set of higher level perceptions notices that this is not the right kind of fire to get a pot of water boiling, to achieve the perceptual result of a touch of liqueur-infused coffee before bedtime. And so, a goal is created for the posterior parietal cortex (PPC) to change the spatial relationships within the campfire scene, so that a different sequence of perceptual results emerges. A program is generated, with sensorimotor decision points at each node, to do whatever is needed to create the desired roaring campfire. Because large chunks of firewood would not readily catch from the remaining embers, smaller branches, with higher visual and tactile spatial frequencies that the ITC would recognize as kindling, must be used. Then the PPC must generate the goal of decreasing the spatial distance between the kindling and the embers, getting positional relationships right by translating the foveal retinal coordinates of the eye into the body coordinates of the arm by means of a set of basis functions. The eye’s Gaussian steering target must be combined with the arm’s sigmoidal enactment of distance away from the body. In this way, the external position of the kindling relative to the embers can be smoothly controlled, via changing arm positions and changing grip configurations. Such rate-of-change transitions get funneled through the basal ganglia striatum, at the head of a final common motor pathway. Once the kindling-placement node has been passed, then the finding-largerfirewood node comes into play. The program here is a way of arranging perceptions in the proper sequence, conditional on the right perceptual results being obtained at each step. The dorsal occipito-parietal stream, termed the “where” pathway in the brain, must locate the requisite firewood, not only as

How the brain gets a roaring campfire Chapter | 7

193

to retinal position but also whether it is within reaching distance of the body. Head and eye orientation perform a steering function for the brain, as does the vergence depth perception of both eyes together. Thus, both size disparity and binocular disparity, analyzed in the striate and prestriate cortices, become indicators of displacement away from the body. Each of these channels would already have distinguished a figure in contrast to its background via their ratio of reflectances, first constructed in the retina and elaborated at higher levels, despite a color similarity between the aged firewood and the dusty gray ground. Texture discrimination in the peristriate cortex would be a key distinguishing feature for the firewood. Coordination of eye and hand for reaching, lifting, and gathering, including eventual joint angle transitions, would be handled by the sequence of opponent-process outputs from the substantia nigra pars reticulata (SN-reticulata) in the basal ganglia, as they project to the superior colliculus and other motor areas downstream. Once that sequence of steps is underway, the program goal of finding the proper firewood then gives way to the behavioral events of stoking and placing the wood to best effect, all at the proper rates of change according to middle temporal cortex specifications. Relations described by spatial prepositions are essentially the job of the PPC. So then, the wood must be on top of the previous kindling and within the growing flames, but without the hand itself getting too close (or at least not for too long) to the heat of the fire. Notice the series of relational perceptions that must get controlled here, whether spatial (i.e., on top of, within, not too close) or temporal (i.e., not too long). These are all aspects that the parietal and middle temporal cortices must track. What is more, not all relationships are the spatial ones of the visual cortex. Some are somatosensory, with respect to the heat of the flames. Others are logical or causal, with presumably a prefrontal component e e.g., will the flames be sufficient, given a specific placement of the pot, to bring the water to a boil? These are all matching tasks as well. Left out of this discussion are the implementing layers below the basal ganglia, with coordinated sequences and transitions of changing joint angles for the shoulder, elbow, and hand, complete with their associated accelerations and muscle dynamics, to bring about the proper placement of said firewood and water. Any given step along the PPC program could involve alteration or delay, depending on disturbances and the perceptual state of affairs already achieved. For instance, is there a stable perch for the blackened pot from the mess-kit, without later settling of the wood tipping it precariously? Have the flames gotten large enough and fast enough to crackle around the sides and bottom of the pot? Is there steam arising yet, or the sizzling sound of water near the top edge, as the water starts to boil? Can the campers safely get to it, without a face full of smoke getting in their eyes, nose, or lungs? These are all points of comparison, where perceptual states are explicitly assessed according to corresponding reference standards. Negative feedback arrangements then provide corrective action to get those perceptions to their preferred values.

194 SECTION | B Models of brain and behavior

In the above list of queries, notice the variety of sensory modalities involved in assessing these conditions: tactile, visual, auditory, olfactory, proprioceptive. That is quite a convergence of inputs to decision-making areas of the brain. Notice, too, the degree of reasoning, memory, and executive involvement in sorting these contingencies e considerations that are far beyond the scope of these chapters. It is worth noting that the outcome of any of these decision nodes could take the required actions in quite unexpected directions. Nonetheless, the ultimate intentions would be achieved: the campfire would blaze up, the coffee flavored with liqueur would be consumed, and the campers would get their delightful evening. This is the central insight behind all of control theory, recognized in an early formulation by William James (1890/1950) over a century ago: “(W)ith intelligent agents, altering the conditions changes the activity displayed, but not the end reached” (Vol. I, p. 8). In other words, the end is determined from within and will come about, even though the means may vary. A further important insight arising from this analysis, and this is one emphasized by Perceptual Control Theory, is that such ends can be structured in a hierarchical arrangement of super-ordinate goals and subordinate underlying goals. The sub-goals comprise the means of implementation of the overarching goals. So then, the firewood is placed, so that the campfire flares up, so that the water may get boiled, so that coffee may be consumed, so that the campers may enjoy a delightful evening, so that they are drawn closer together in the process. Partitioning tasks this way gives the brain a certain modular precision for correcting intervening errors. It also demonstrates significant flexibility, when reassessing what degree of attainment is preferred. Goals can be modified at will, or alternate routes sought out for achieving them. It should be clear from the foregoing analysis that goals and sub-goals must be measured perceptually as to whether they are being achieved. What other choice is there? And indeed, any given goal is not about the implementing behavior per se, but about the perceptual effect of that behavior. It is perceptual results that ultimately matter, not the varying means of getting there. There is a clean and intuitive parsimony to Perceptual Control Theory, with its modular equations for inter-nested control loops. Nonetheless, the brain’s process of specifying control loop functions is far from simple, as these chapters demonstrate. There are myriad methods for the nervous system to construct perceptual input functions. Output structures in the thalamus and basal ganglia are somewhat standardized, and yet they have their own complexity. To obtain a roaring campfire involves quite a wondrous interplay of perceptions from a wide gamut of the brain’s capabilities. Perhaps the greatest wonder is that the brain knows how to enjoy campfires at all.

How the brain gets a roaring campfire Chapter | 7

195

Acknowledgments I wish to thank Bruce Nevin, Dominique Berule, and Warren Mansell for very careful readings and perceptive comments on earlier versions of this chapter.

References Andersen, R. A., Essick, G. K., & Siegel, R. M. (October 25, 1985). Encoding of spatial location by posterior parietal neurons. Science, 230(4724), 456e458. Anzai, A., Ohzawa, I., & Freeman, R. D. (1999). Neural mechanisms for encoding binocular disparity: receptive field position versus phase. Journal of Neurophysiology, 82(2), 874e890. Barlow, H. B. (1953). Summation and inhibition in the frog’s retina. Journal of Physiology, 119(1), 69e88. Bateson, G. B. (1979). Mind and nature: a necessary unity. New York: E.P. Dutton. Bateson, G., & Bateson, M. C. (1987). Angels fear: towards an epistemology of the sacred. New York: Bantam Books. Blakemore, C., & Campbell, F. W. (1969). On the existence of neurones in the human visual system selectively sensitive to the orientation and size of retinal images. The Journal of Physiology, 203(1), 237e260. Blakemore, C., Nachmias, J., & Sutton, P. (October 1970). The perceived spatial frequency shift: evidence for frequency-selective neurones in the human brain. The Journal of Physiology, 210(3), 727e750. Campbell, F. W., & Robson, J. G. (August 1968). Application of Fourier analysis to the visibility of gratings. The Journal of Physiology, 197(3), 551e566. Cascio, C. J., & Sathlan, K. (July 15, 2001). Temporal cues contribute to tactile perception of roughness. Journal of Neuroscience, 21(14), 5289e5296. Available at: http://www.jneurosci. org/content/21/14/5289.full Accessed: 3/12/2014. Chen, E. Y., Marre, O., Fisher, C., Schwartz, G., Levy, J., da Silveira, R. A., & Berry, M. J., II (January 2, 2013). Alert response to motion onset in the retina. Journal of Neuroscience, 33(1), 120e132. Churchland, P. S., & Sejnowski, T. J. (1992). The computational brain. Cambridge, MA: The MIT Press. Cools, A. R. (1985). Brain and behavior: hierarchy of feedback systems and control of input. In P. P. G. Bateson, & P. H. Klopfer (Eds.), Perspectives in ethology: volume 6 mechanisms (pp. 109e166). New York: Plenum Press. Cudeiro, J., & Sillito, A. M. (2006). Looking back: corticothalamic feedback and early visual processing. Trends in Neurosciences: Neural substrates of cognition, 26(6), 298e306. De Valois, R. L., & De Valois, K. K. (1988). Spatial vision, issue 14 (oxford psychology series). New York: Oxford University Press. Dewey, J. (1896/1998). The reflex arc concept in psychology. In J. F. Brennan (Ed.), Readings in the history and systems of psychology (second ed., pp. 147e156). Upper Saddle River, NJ: Prentice Hall. Dowling, J. (2007). Retina. Scholarpedia, 2(12), 3487. Fogel, I., & Sagi, D. (1989). Gabor filters as texture discriminator. Biological Cybernetics, 61, 103e113. Foster, K. H., Gaska, J. P., Nagler, M., & Pollen, D. A. (1985). Spatial and temporal frequency selectivity of neurones in visual cortical areas V1 and V2 of the macaque monkey. Journal of Physiology, 365, 331e363.

196 SECTION | B Models of brain and behavior George, D. (2008). How the brain might work: a hierarchical and temporal model for learning and recognition. Doctoral dissertation, Stanford University, 2008. Available at: http://alpha.tmit. bme.hu/speech/docs/education/02_DileepThesis.pdf Accessed: 1/19/2014. Glezer, V. (1995). Vision and mind: modeling mental functions. Mahwah, NJ: Lawrence Erlbaum Associates. Gonzalez, F., & Perez, R. (June 1998). Neural mechanisms underlying stereoscopic vision. Progress in Neurobiology, 55(3), 191e224. Hawken, M. J., Shapley, R., & Grosof, D. (1996). Temporal-frequency selectivity in monkey visual cortex. Visual Neuroscience, 13(3), 477e492. Hawkins, J., & Blakeslee, S. (2004). On intelligence. New York: Henry Holt and Company. Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. Journal of Physiology, 160, 106e154. Hubel, D. H., & Wiesel, T. N. (1974). Sequence regularity and geometry of orientation columns in the monkey striate cortex. Journal of Comparative Neurology, 158, 267e294. Inoue, T., & Imoto, K. (2006). Feedforward inhibitory connections from multiple thalamic cells to multiple regular-spiking cells in layer 4 of the somatosensory cortex. Journal of Neurophysiology, 96(4), 1746e1754. James, W. (1890/1950). The principles of psychology (vol. I). New York: Dover Publications. Jarsky, T., Cembrowski, M., Logan, S. M., Kath, W. L., Riecke, H., Demb, J. B., & Singer, J. H. (July 27 , 2011). A synaptic mechanism for retinal adaptation to luminance and contrast. Journal of Neuroscience, 33(30), 11003e11015. Jin, J., Wang, Y., Swadlow, H. A., & Alonso, J. M. (2011). Population receptive fields of ON and OFF thalamic inputs to an orientation column in visual cortex. Nature Neuroscience, 14, 232e238. Klein, S. A. (1992). Channels: bandwidth, channel independence, detection vs. discrimination. In B. Blum (Ed.), Channels in the visual nervous system: neurophysiology, psychophysics and models (pp. 11e27). London: Freund Publishing House. Kreiman, G. (2008). Biological object recognition. Scholarpedia, 3(6), 2667. Kuffler, S. W. (January 1953). Discharge patterns and functional organization of mammalian retina. Journal of Neurophysiology, 16(1), 37e68. Kulikowski, J. J., & Tolhurst, D. J. (1973). Psychophysical evidence for sustained and transient detectors in human vision. Journal of Physiology, 232, 149e162. Law, C. C., & Cooper, L. N. (1994). Formation of receptive fields in realistic visual environments according to the Bienenstock, Cooper, and Munro (BCM) theory. Proceedings of the National Academy of Sciences of the United States of America, 91(16), 7797e7801. Lee, A. B., Blais, B., Shouval, H. Z., & Cooper, L. N. (2000). Statistics of lateral geniculate nucleus (LGN) activity determine the segregation of ON/OFF subfields for simple cells in visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 97(23), 12875e12879. Lehar, S.(undated). An intuitive explanation of Fourier theory. [On-line.] Available at: http://cnsalumni.bu.edu/wslehar/fourier/fourier.html Accessed: 2/24/2014. Llina´s, R. R., Lesnik, E., & Urbano, F. J. (2002). Temporal binding via cortical coincidence detection of specific and nonspecific thalamocortical inputs: a voltage-dependent dye-imaging study in mouse brain slices. Proceedings of the National Academy of Sciences of the United States of America, 99(1), 449e454. Available at: http://www.pnas.org/content/99/1/449.full Accessed: 2/17/2014.

How the brain gets a roaring campfire Chapter | 7

197

Loftus, G. R., & Harley, E. M. (2004). How different spatial-frequency components contribute to visual information acquisition. Journal of Experimental Psychology: Human Perception and Performance, 30(1), 104e118. Maunsell, J. H. R., & Van Essen, D. C. (May 1983). Functional properties of neurons in middle temporal visual area of the macaque monkey. II. Binocular interactions and sensitivity to binocular disparity. Journal of Neurophysiology, 49(5), 1148e1167. Marken, R. S. (1983). “Mind reading”: a look at changing intentions. Psychological Reports, 53, 267e270. Mishkin, M., Ungerleider, L. G., & Macko, K. A. (October 1983). Object vision and spatial vision: two cortical pathways. Trends in Neurosciences, 414e417. Nawrot, M., & Joyce, L. (2006). The pursuit theory of motion parallax. Vision Research, 46, 4709e4725. Palanca, B. J. A., & DeAngelis, G. C. (August 20 , 2003). Macaque middle temporal neurons signal depth in the absence of motion. Journal of Neuroscience, 23(20), 7647e7658. Parker, A. J., & Cumming, B. G. (2001). Cortical mechanisms of binocular stereoscopic vision. In C. Casanova, & M. Ptito (Eds.), Progress in Brain Research (vol. 134, pp. 1e12). Perrone, J. A. (2004). A visual motion sensor based on the properties of the V1 and MT neurons. Vision Research, 44, 1733e1755. Perrone, J. A. (2005). Economy of scale: a motion sensor with variable speed tuning. Journal of Vision, 5, 28e33. Perrone, J. A. (November 15 , 2006). A single mechanism can explain the speed tuning properties of MT and V1 complex neurons. Journal of Neuroscience, 26(46), 11987e11991. Perrone, J. A., & Thiele, A. (2002). A model of speed tuning in MT neurons. Vision Research, 42, 1035e1051. Plainis, S., Parry, N. R. A., Panorgian, A., Sapountzis, P., & Murray, I. J. (2009). Summation characteristics of the detection of compound gratings. Vision Research, 49, 2056e2066. Poggio, G. E. (May-June). Mechanisms of stereopsis in monkey visual cortex. Cerebral Cortex, 5(3), 193e204. Pouget, A., & Sejnowski, T. J. (1995). Spatial representations in the parietal cortex may use basis functions. In G. Tesauro, D. S. Touretsky, & T. K. Leen (Eds.), Advances in neural information processing systems 7 (pp. 157e164). Cambridge, MA: MIT Press. Pouget, A., & Sejnowski, T. J. (1997). A new view of hemineglect based on the response properties of parietal neurones. Philosophical Transactions of the Royal Society B: Biological Sciences, 352, 1449e1459. Powers, W. T. (1973). Behavior: the control of perception. Chicago: Aldine Publishing. Powers, W. T. (1978). Quantitative analysis of purposive systems: some spadework at the foundations of scientific psychology. Psychological Review, 85, 417e435. Powers, W. T. (1979). A cybernetic model for research in human development. In M. N. Ozer (Ed.), A cybernetic approach to the assessment of children: toward a more humane use of human beings (pp. 11e66). Boulder, CO: Westview Press. Powers, W. T. (1990). A hierarchy of control. In R. J. Robertson, & W. T. Powers (Eds.), Introduction to modern psychology: the control-theory view (pp. 59e82). Gravel Switch, KY: The Control Systems Group. Powers, W. T. (2008). Living control systems III: the fact of control. Bloomfield, NJ: Benchmark Publications.

198 SECTION | B Models of brain and behavior Powers, W. T., Clark, R. K., & McFarland, R. L. (1960a). A general feedback theory of human behavior: Part 1. Perceptual & Motor Skills, 11, 71e88. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960b). A general feedback theory of human behavior: Part 2. Perceptual & Motor Skills, 11, 309e323. Shapley, R., & Enroth-Cugell, C. (1984). Visual adaptation and retinal gain controls. In N. Osbourne, & G. Chader (Eds.), Progress in retinal research (vol. 3, pp. 263e346). Oxford: Pergamon. Shapley, R., Kaplan, E., & Purpura, K. (1993). Contrast sensitivity and light adaptation in photoreceptors or in the retinal network. In R. M. Shapley, & D. M.-K. Lam (Eds.), Contrast sensitivity (vol. 5, pp. 103e116). Cambridge, MA: MIT Press. Sompolinsky, H., & Shapley, R. (1997). New perspectives on the mechanisms for orientation selectivity. Current Opinion in Neurobiology, 7, 514e522. Ungerleider, L. G., & Mishkin, M. (1982). Two cortical visual systems. In D. J. Ingle, M. A. Goodale, & R. J. W. Mansfield (Eds.), Analysis of visual behavior (pp. 549e586). Cambridge, MA: MIT Press. Ungerleider, L. G., & Passoa, L. (2008). What and where pathways. Scholarpedia, 3(11), 5342. Van Essen, D. C., Anderson, C., & Felleman, D. J. (January 24, 1992). Information processing in the primate visual system: an integrated systems perspective. Science, 255(5043), 419e423. Watson, A. B., & Ahumada, A. J., Jr. (February 1985). Model of human visual-motion sensing. Journal of the Optical Society of America, 2(2), 322e343. Watson, N. V., & Breedlove, S. M. (2012). A step further 7.4: most V1 cells are tuned to particular spatial frequencies. In On companion website for The mind’s machine: foundations of brain and behavior. Sunderland, MA: Sinauer Associates. Available at: http://www.mindsmachine. com/asf07.04.html Accessed: 3/23/2014. Wikipedia contributors. (March 30, 2016). Sigmoid function. In Wikipedia, the free encyclopedia. Retrieved May 2, 2016, from https://en.wikipedia.org/w/index.php?title¼Sigmoid_ function&oldid¼731468779. Xing, J., & Andersen, R. A. (2000). Models of the posterior parietal cortex which perform multimodal integration and represent space in several coordinate frames. Journal of Cognitive Neuroscience, 12(4), 601e614. Yin, H. H. (2014a). Action, time and the basal ganglia. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1637). Article ID 20120473. Yin, H. H. (2014b). How basal ganglia outputs generate behavior. Advances In Neuroscience, 28. Article ID 768313. Young, R., & Illingworth, J. (1998). Towards a control model of object recognition. Journal of Perceptual Control Theory, 1. Available at: http://citeseerx.ist.psu.edu/viewdoc/download? doi¼10.1.1.46.2858&rep¼rep1&type¼pdf Accessed: 1/27/2014.

Chapter 8

The phylogeny, ontogeny, causation and function of regression periods explained by reorganizations of the hierarchy of perceptual control systems Frans X. Plooij1 International Research-institute on Infant Studies (IRIS), Arnhem, the Netherlands

Introduction The aim of this book is twofold: first, to publish a post-hoc ‘Festschrift’ for Bill Powers, the (grand)father of Perceptual Control Theory (PCT), and, second, to communicate the many scientific advances based on PCT to a wide, multidisciplinary audience by contributions of authors who are wellacquainted with some conventional branch of the life sciences as well as with PCT d who understand why the older theories were persuasive and even useful, and also what has to be changed about them as the realities of perceptual control are to be introduced into them. My branch of the life sciences is behavioral biology, also called ethology. In the 1970s I studied the behavioral development of free-living chimpanzee infants in the Gombe National Park, Tanzania, East-Africa, and, together with my late wife Hetty van de Rijt-Plooij, the growing independence, conflict and

1. Acknowledgments: My thanks to Lomax Boyd, Jonathan Lombard, Evan MacLean, Warren Mansell, John Richer, Alex Weiss and two reviewers for reading earlier drafts of this chapter and providing useful comments and suggestions for improvement. This research was financially supported by the ‘Netherlands Foundation for the Advancement of Tropical Research’ (WOTRO grant no. W84-66), the ‘Dr. J.L. Dobberke Stichting voor Vergelijkende Psychologie’ at Amsterdam, the ‘Stichting Kinderpostzegels Nederland’ (project no. PH 89/74), and the ‘Fonds Doctor Catharine van Tussenbroek’ in the Netherlands. The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00008-3 Copyright © 2020 Elsevier Inc. All rights reserved.

199

200 SECTION | C Collective control and communication

learning in mother-infant relations of both Gombe chimpanzees and humans. In this chapter, an overview of this body of work is presented that has never been published before. The discoveries made in this body of work emerged from the data. It is an example of induction as opposed to deduction. No theory or hypothesis was guiding the data collection. In the tradition of behavioral biology everyday life of free-living chimpanzees was observed directly in a natural setting, taking care of good observational control and making sure the sample size of the monthly observations was large enough to guarantee reliable, quantitative, behavioral measures. First, I studied the behavioral development of free-living chimpanzee infants (Plooij, 1984). To give an idea of what kind of data were collected, a brief description follows of the observation methods used. It takes years to get familiar with an animal’s behavior (Fagen & Goldman, 1977). In the Gombe Research Center, familiarity with the behavior of chimpanzees was already present and handed over from one researcher to the next. Therefore, my pilot study (to get familiar with the chimpanzees and their environment) only lasted six months, during which time I spent 407 h in the field. First, it took me three weeks before I was able to recognize all the chimps individually. Second, I needed four weeks to follow chimpanzees from all age/sex classes on foot. “It is necessary intellectually to soak in the environmental complex of the animal to be studied until we have a facility with it which keeps us, as it were, one move ahead” (Schneirla, 1950). And third, during the remainder of the pilot study, I followed mother-infant pairs and dictated a running commentary (see Hutt & Hutt, 1970) into a portable cassette-tape recorder. From these running commentaries grew a list of environmental events considered important and behavior categories that were observed repeatedly (Appendix A in Plooij, 1984). The environmental events included all situations in which an infant overtly reacted to: the distance to other chimpanzees, the presence of other species, vegetation noise (a possible danger!), weather variations, and location. Three criteria were used for the definitions of behavior categories. First, the age-dependency criterion: before the age of 5 months it was near impossible to observe discrete behavior patterns that were stereotyped enough to allow definition and quantification; and only after 12 months was it possible to define behavior categories according to consequence (such as ‘nestbuilding’). Second, mutually exclusive behavior categories were not used for the whole chimpanzee. The categories were divided into 11 groups, one per body part. Intra-group categories are mutually exclusive, inter-group categories may combine. The continuous stream of behavior was described in terms of a sequence of such combinations. During the main study I followed the mother-infant pairs monthly for as long as was necessary to get 300 min of good observation, which usually took 1e2 days. The behavior of the baby or infant was recorded continuously together with the behavior of the individual it was interacting with.

The phylogeny, ontogeny, causation and function Chapter | 8

201

The main question that dominated the analysis of the data was whether it was possible to find out what type of organization was underlying the developing behavior. In the 20th century there were two opposing lines of thought about the development of behavior (Kortlandt, 1955): differentiation versus ascending development. Differentiation starts with a perfectly integrated unity that expands progressively. Within this unity parts individuate and become more or less discrete. For example, in my own study of newborn human babies (Plooij, 1978) we observed many examples of this total unity, i.e. the movements of one arm mirror the movements of the other or a blink of the eyes is mirrored in a movement of the feet. Other examples are the works of Coghill (1929) on the embryology of behavior in the newt Amblystoma; Humphrey (1969) on the ‘total pattern reflex’ in the human fetus involving the organism as a whole, and Condon (1979) on self-synchrony in human newborns. In contrast to differentiation, “ascending development proceeds through the emergence of isolated units followed by an ascending integration into a hierarchy” (Plooij, 1984, p. 2). Discrete types of behavior emerge, such as walking, flying, twig-quivering or munching fish in cormorants (Kortlandt, 1955), and after a while these become integrated into functional sequences. Good examples are the works of Kortlandt (1955), Kruijt (1964), and Hailman (1967) concerning different bird species; Rosenblatt’s (1976) review of the studies on altricial, non-primate mammals; Baerends-van Roon and Baerends (1979) concerning the domestic cat; and Chevalier-Skolnikoff (1974) on stumptail macaques. Neither differentiation nor ascending development covers all observed behaviors however. Therefore, it has been concluded that ascending development replaces differentiation during ontogeny. According to Dawkins (1976) this is only logical if one realizes that the unity of the differentiated parts would be lost if there would not emerge a hierarchically superior system integrating them. Traditionally, in behavioral biology at the time, discrete behavior patterns were used to find such systems in adult organisms (Baerends, 1956, 1976; Fentress, 1983; Hinde, 1953).2 However, such discrete behavior patterns do not exist in chimpanzee infants younger than 5 months, so the search for hierarchically superior systems seemed to have ended in an impasse.

2. This is logical if one realizes that ethology’s presumption at the time was the same as the presumptions of the other behavioral sciences, namely that it is behavior that is controlled, not perception. The behavior was considered to be controlled by a hierarchical system of Innate Releasing Mechanisms (IRMs) on the one hand and a hierarchical system of motor centers on the other hand. Each motor center controls a configurational pattern of muscle contraction (Tinbergen, 1951/1974). (pp. 103e104).

202 SECTION | C Collective control and communication

In my efforts to get out of this impasse, I strolled away from the mainstream of ethology and discovered the work of the Estonian biologist Jakob von Uexku¨ll (1933). He focused much more on perception instead of behavior and he came up with the concept of the ‘Umwelt’. The Umwelt is, roughly speaking, the world as it is experienced by a particular organism. Different species all have a different Umwelt because their respective nervous systems have evolved to seek out and respond only to those aspects of the environment that are relevant.3 Over evolutionary time, of course, any persistent long-term change in environmental circumstances can result in selection and adaptation to the new circumstances.4 The creature’s Umwelt will change accordingly. Adriaan Kortlandt was a young student and an early pioneer in the field of ethology (Kortlandt, 1940a; 1940b, 1955) in the 1930s and was influenced by the ideas of Von Uexku¨ll, as he told me himself when I studied in his lab. Kortlandt had observed free-living cormorants. He describes “hierarchically organized appetites” (Kortlandt, 1955) (pp. 171e172) where an appetite is defined as “either the performance of a specific activity or the presence of a specific external object or situation (consummatory situation) which causes the ending of a variable sequence or series of activities leading to this particular situation.” Because the hierarchically organized appetites are ending activities, Dawkins described them as “hierarchically nested stopping rules” (Dawkins, 1976). This does not only apply to situations outside the body of the organism, but also to situations inside the body: movement-regulation feedback through proprioception and perception of results outside the body had been shown to be important in primates (Marsden, Merton, Morton, & Adam, 1978; Polit & Bizzi, 1979), mice (Fentress, 1976), and some bird species (Baerends, 1956, 1970; Bastock, Morris, & Moynihan, 1953). Set values play a crucial role in such feedback (Baerends, 1976); all efforts to restore those values stops as soon as the perception of the situation conforms to the set-values. Kortlandt describes how the appetites and the concomitant consummatory situations are arranged in levels that successively emerge and integrate during ontogeny (Kortlandt, 1955). So, it seemed that these appetites and concomitant consummatory situations could be useful in finding the hierarchically superior systems underlying behavioral development. However, these appetites and concomitant consummatory situations appeared not easy to find, if one wants to avoid the trap of using descriptions of behavior for explaining learning and development (Bickhard, 1992). Furthermore, Golani (1976) had shown that chaos seemed

3. This is in line with findings concerning the selective nature of sensory receptors that act as filters to certain features of the environment (Kolb & Wishaw, 1990). As a consequence, biologists, trained to focus on differences between species, found that some species would learn a specific task when related species would not (Hinde & Stevenson-Hinde, 1973). 4. Recently, for the first time the evolution of ‘prepared learning’ was demonstrated experimentally in a population of Drosophila over 40 generations (Dunlap & Stephens, 2014).

The phylogeny, ontogeny, causation and function Chapter | 8

203

present in the observed behavior as long as one did not know what was controlled. As soon as one did know, order appeared in the same overt behavior. Hofer (1978) was talking about ‘hidden, regulatory processes.’ When my thinking had developed thus far, my colleague A.R. Cools pointed out Powers (1973, 1978, 2005) work to me and suggested that Powers’ hierarchical perceptual control theory might be very useful as a lead to detect the appetites and concomitant consummatory situations at different hierarchical perceptual levels. In addition, three procedures were applied in this search. First, the test for the controlled variable (TCV) by studying the reactions to disturbances (Powers, 1973, 2005). Second, the speed of control systems (the higher the level in the hierarchy, the slower the system according to Powers (1973, 2005) (p. 116). And third, rigidity versus variability of behavior. The third procedure needs some explanation. For his cormorants, Kortlandt (1955) showed that normal variability of behavior occurs mainly at one level below the highest level of the hierarchy. He concluded that this “zone of variability in behavior ascends in the same degree as does the progress in maturation.” (pp. 190e192). From Bingham’s (1928) descriptions of ontogeny in the chimpanzee, Kortlandt suggested that the same rigidity versus variability procedure may apply to behavioral development in this ape. Using Powers’ hierarchical perceptual control theory and these three procedures, I found major changes in the behavioral development of free-living chimpanzee infants that could well be explained in terms of one level after the other being superposed onto the already existing hierarchy (Plooij, 1984). Thereafter, my late wife and I analyzed the data further while focusing on the growing independence, conflict and learning in mother-infant relations. The major changes in the behavioral development and the underlying hierarchical organization that I found appeared to be associated with regression periods (van de Rijt-Plooij & Plooij, 1987). The growing independence of the infant appeared not to be a gradual process, but to occur in leaps. With each ‘leap’ in independence comes a sharp decrease in mother-infant body contact. Immediately before each leap a regression period occurred. Regression was expressed by a temporary shift back to mainly staying closer to mother and by a temporary increase in the amount of ventro-ventral contact. A central role can be assigned to age-related regression periods in the quest for the processes underlying the ontogeny of the hierarchical levels of perceptual control systems. “On the one hand, there is the link with brain changes. On the other hand, each regression period signals the start of a period of developmental progress and the emergence of new skills, task performances, and behaviors. The age-related regression periods stand out as lighthouses to direct the study of developmental change.” (Plooij, 2003).

204 SECTION | C Collective control and communication

They will be described following the “Four Why’s” of behavioral biology (Tinbergen, 1963) concerning evolution, development, causation, and function.5

Evolution The regression periods were discovered in 1973 in free-living chimpanzees in the Gombe National Park, Tanzania, East-Africa (van de Rijt-Plooij & Plooij, 1987). Each regression period was followed by a period of mother-infant conflict over body contact together with the sharp decrease in mother-infant body contact.6 An example of such mother-infant conflict can be seen in Fig. 8.1A and B. Evolutionarily speaking, the phenomenon of regression periods in early development turned out to be very old, indeed, and was reported by Horwich in 1974 for twelve monkey species and two non-primate mammals (Horwich, 1974). In those species, regression was expressed as peaks in nipple contact. These peaks are age-linked and occur at similar times in development, if a correction is made for the speed of development of each species. According to Horwich, the peaks are related to emotional states of insecurity. The next question was whether this phenomenon of regression periods survived during the evolution of our own species Homo sapiens. If it is present in non-primate mammals, monkeys and apes, one would expect it to be present in us as well, because we are so closely related. In 1989, Horwich had published promising results on all four ape species (three orangutans, two gorillas, one chimpanzee and four humans) that showed periods of a recurrence of more time spent nursing and in contact with the mother (Horwich, 1989). However, he sampled monthly, and weekly sampling is required to capture the phenomenon completely.

Development In a study in the Netherlands on human babies, ten age-linked regression periods were found in the first 20 months, covering the sensorimotor stage

5. In 1963 Tinbergen characterized the discipline Ethology as ‘the biological study of behavior’. ‘Behavior’ refers to the fact that Ethology has the inductive description of observable phenomena as a starting point. ‘Biological’ refers to a method of study or the biological method. The latter includes the general scientific method and, in addition, four questions concerning major problems that are studied throughout biology: evolution, development, causation, and function (or survival value). Tinbergen insisted that Ethology gives “equal attention to each of them and to their integration” (Tinbergen, 1963). 6. Trivers argued that, according to parental investment theory, parent and offspring are expected to disagree over the duration and the amount of parental investment that should be given. Interestingly, the periods of mother-infant conflict in chimps are triggered by the regression periods where the infant is demanding extra care (Trivers, 1974).

The phylogeny, ontogeny, causation and function Chapter | 8

205

FIG. 8.1 Mother Passion bite-gnawing, pushing, pulling, pinching, and slapping her baby Prof to get him off the belly and off the nipple (A). Here she failed and resumed walking (B). Line drawings Copyright © 2018 by David Bygott.

(van de Rijt-Plooij & Plooij, 1992). These are periods during which the infant was more difficult than usual according to the mother. In such a period the baby is crying more than usual, clinging to the mother more than usual and more cranky or grumpy, than usual. We described these periods as being characterized by the three C’s: Crying, Clingy and Cranky. Direct observations of mother-infant interactions in the homes confirmed the maternal reports. Just to give one example and a feel for the phenomenon, Fig. 8.2 presents the percentage of direct observation time one baby spent in body contact with the mother by age. While the percentage goes down over time, one can clearly see the temporary peaks in body contact superposed onto the downward trend. These temporary peaks represent the regression periods in which the baby is more clingy than usual. The effects of important sources of ‘noise’ on direct observation measures of regression periods were described using case studies of four infants’ adaptations to special parental conditions (Plooij & van de Rijt-Plooij, 2003). Three independent research groups have replicated the above findings on age-linked regression periods (Lindahl, Heimann, & Ullstadius, 2003; Sadurni & Rostan, 2002, 2003; Woolmore & Richer, 2003). Just like in the chimps, conflict periods followed the regression periods (van de Rijt-Plooij & Plooij, 1993). These data are based on weekly questionnaires combined with in-depth interviews. At first, mothers were worried that something was wrong with their baby during a regression period. Sometimes they would even pay a visit to the pediatrician, only to find out that nothing was wrong. Then their worry would change into annoyance. In the first

206 SECTION | C Collective control and communication BODY CONTACT Infant E 100

90

% of Observation Time

80

70

60

50

40

30

20

10

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65

AGE (weeks)

FIG. 8.2 Temporary peaks in body contact superposed on a downward trend. From Plooij, F. X., & van de Rijt-Plooij, H. (2003). The effects of sources of "noise" on direct observation measures of regression periods: Case studies of four infants’ adaptations to special parental conditions. In M. Heimann (Ed.), Regression periods in human infancy (pp. 57e80). Mahwah, NJ: Erlbaum. Copyright © 2003 by Lawrence Erlbaum Associates, Inc. Reprinted by kind permission of Lawrence Erlbaum Associates, Inc.

few months, mothers would not act on their feelings of annoyance. But at later ages, especially the second half of the first year, they would. This was called ‘promoting progress’, because the mothers sensed that their baby was able to do more and they would demand more of their baby. At this age, they would still use mild strategies by diverting the attention of their baby. And the baby would go along with its mother. But soon, and especially during the second year of life, the babies would not go along with their mothers anymore and straightforward ‘clashes’ would result. Around 18 months, all the mothers reported clashes to occur. Comparing what babies could understand before and after a regression period, we found that they had made a leap and were able to perceive a new order of phenomena. In terms of PCT (Powers, 1973, 2005), we assumed that with each regression period a new type of perception had emerged, and a new perceptual world had opened up for the baby to explore. Following that assumption, we should expect the baby to develop new skills. In collaboration

The phylogeny, ontogeny, causation and function Chapter | 8

207

with my colleagues in Spain, we tested the hypothesis that following each regression period, a cluster of new skills should emerge (Sadurni, Burriel, & Plooij, 2010). We asked the mothers of the babies each week what new skills they had observed that week and the results are shown schematically in Fig. 8.3. The blue line stands for the occurrence of new skills and the red line represents the regression periods and it is clear that the peaks in new skills follow the regression periods. Our original findings having been replicated by independent research teams and validated by direct observation, we moved on to search for convergent evidence of a different kind. The line of argument was as follows. The relation between regression (or disorganization) on the one hand and progression (or reorganization) on the other hand has been considered for almost a century by scientists from various backgrounds (Bever, 1982; Kortlandt, 1955; Kozulin, 1990; Mahler, Pine, & Bergman, 1975; McGraw, 1945/1974; Mounoud, 1976; Peterfreund, 1971; Schore, 1997; Scott, 1986; Smotherman & Robinson, 1990; Thelen, 1989; Werner, 1948). The ethologist Kortlandt even invented the term ‘reprogression.’ Being disorganized, the whole organism is ‘off balance.’ So, the disorganization should not only show in the behavior within the phenomenon of regression periods, but in other aspects as well. The progress in the discipline psychoneuroimmunology over the last few decades has brought to light the complex interactions between behavior, the CNS, the endocrine system, and the immune system (Ader, Felten, & Cohen, 2001). If the organism is disorganized, this should also show in the immune system and the health of the organism, among other things. We approached this question by studying the distribution of illnesses over early age and predicted we should find peaks in illness around the ages at which we had found the regression periods. And so we did (Plooij, van de Rijt-Plooij, van der Stelt, van Es, & Helmers, 2003). Consequently, we did a similar study on peaks of death in infants that died of Sudden Infant Death Syndrome (SIDS) and found small peaks superimposed on the one large peak reported until then in the literature (Plooij, van de Rijt-Plooij, & Helmers, 2003). McKenna, (1990a, b); McKenna and Mosko (1990) has suggested a connection between the one large peak in SIDS and the shift from reflexive to speech breathing with its neurological control system errors. In similar vein there might be more sudden changes in the brain that underlie the other, superimposed peaks in SIDS. These suggestions that sudden brain changes might underlie the regression periods and the peaks in illnesses and SIDS bring us to the third of the Four Why’s: the question concerning immediate causation.

Causation Trevarthen and Aitken (2003) conducted an extensive literature review on the pre- and post-natal development of the central nervous system. There is clear

1° PRC

2° PRC

3° PRC

4° PRC

6° PRC

5° PRC

8° PRC

7° PRC

6,00 5,00 4,00

3,00 2,00

1,00 0,00 0

5

10

15

20

25

30

35

40

45

50

55

60

Estimation of the distribution of each Regression Period according to our data Estimation of the distribution of each Transition Periods according to our data FIG. 8.3 Each regression period is followed by a cluster of new skills. From Sadurni, M., Burriel, M. P., & Plooij, F. X. (2010). The temporal relation between regression and transition periods in early infancy. Spanish Journal of Psychology, 13(1), 112e126. Reprinted by kind permission of the publisher.

208 SECTION | C Collective control and communication

7,00

The phylogeny, ontogeny, causation and function Chapter | 8

209

evidence for sudden, gene-regulated, age-linked brain changes. For instance, sudden increases in the skull circumference were found in the first few months shortly before or at the beginning of regression periods. Sudden age-linked brain changes occur shortly before or at the beginning of most regression periods. Knowing this, one might approach the question of immediate causation of regression periods from another angle, taking purpose into account experimentally by testing for controlled variables (TCV; Marken, 2013), based on PCT. If a living organism is able to control a certain perception, so the theory goes, any deviation from the reference value of that perception should be met with resistance, counteracting the deviation from the reference value caused by the disturbance. If the organism is not able to perceive that type of perception, the disturbance is met with indifference. It is simply not perceived. This is called the test for the controlled variable or quantity (Powers, 1973, 2005). Also in developmental studies this test can play an important role, as we will see shortly. The PCT model of the sensorimotor stage specifies what type of perception is emerging at the beginning of or shortly before each regression period (Plooij, 2003). This is shown in Fig. 8.4. The types of perception that are supposed to emerge are the perception of:

FIG. 8.4 The Hierarchical Perceptual Control Theory model of the sensorimotor stage. From Plooij, F. X. (2003). The trilogy of mind. In M. Heimann (Ed.), Regression periods in human infancy (pp. 185e205). Mahway, NJ: Erlbaum. Copyright © 2003 by Lawrence Erlbaum Associates, Inc. Reprinted by kind permission of Lawrence Erlbaum Associates, Inc.

210 SECTION | C Collective control and communication

Configurations, Smooth transitions, Events, Relationships, Categories, Sequences, Programs, Principles, and System concepts. We conducted some preliminary studies testing this PCT-model with a variant of the TCV. The emergence of the perception of Events and the perception of Sequences were chosen to be tested first for two reasons. First, by combining ethological and experimental, neurophysiological techniques, Aldridge and coworkers (Aldridge, Berridge, Herman, & Zimmer, 1993) have shown that the neostriatum is involved in the perception and control of sequences of behavior and determines the syntax of grooming in rats. Lesions in the neostriatum affect only the serial order of behavior but not the behavioral elements or events. This may show that the perception and control of events is ruled by another hierarchical level than the perception and control of sequences. Second, Diamond and coworkers (Bell, Wolfe, & Adkins, 2007; Diamond & Goldman-Rakic, 1989; Diamond, Werker, & Lalonde, 1994) have shown for human infants that the perception of, or memory for, temporal order is a potential that underlies the development of the whole list of skills, behaviors or task performances. For each type of perception, a battery of some 20 to 30 tasks was developed that were based on that type of perception. Two of my students (Ten Horn & Paro, 1995) then followed individual babies weekly for four months, two months before and two months after the age of onset of a regression period and the supposed emergence of a particular type of perception. Every week we presented the battery of task items to the babies. The prediction was that a baby should not perform well with the battery of tasks before this age of onset, but thereafter should quickly master one after the other. Fig. 8.5 concerns the perception and control of events. In this graph the cumulative number of task items that was finished successfully is plotted against age (in weeks). Weeks 14e17 concern the regressive period as reported

FIG. 8.5 The successfully finished, cumulative number of task items concerning ‘Events’ over age (in weeks). From Ten Horn, J., & Paro, I. (1995). Mentale ontwikkeling bij baby’s: Try-out taakbatterijen. Heymans Bulletin Psychologische Instituten. Scriptie. Ontwikkelingspsychologie. Rijksuniversiteit Groningen. Groningen. Reprinted with kind permission of the authors.

The phylogeny, ontogeny, causation and function Chapter | 8

211

FIG. 8.6 The successfully finished, cumulative number of task items concerning ‘Sequences’ over age (in weeks). From Ten Horn, J., & Paro, I. (1995). Mentale ontwikkeling bij baby’s: Tryout taakbatterijen. Heymans Bulletin Psychologische Instituten. Scriptie. Ontwikkelingspsychologie. Rijksuniversiteit Groningen. Groningen. Reprinted with kind permission of the authors.

by the mothers. The individual graphs of four babies are depicted. Before the regression period, the babies did not master a single item, and thereafter the graphs clearly show improvement. Fig. 8.6 concerns four other babies and the perception of sequences. Weeks 40e44 concern the regressive period as reported by the mothers. A similar picture is shown.

Function Regression periods are difficult on the parents who would sometimes rather do without them. However, they do have a function. For instance, Marten de Vries (1984) studied the survival of Masai children in East Africa with and without a ‘difficult temperament’ in times of famine. A greater number of difficult children survived because they were better able to elicit care from their mothers. Another example concerns data that support the hypothesis that difficult infants activate special family resources, which stimulates intellectual development over the years. Using three temperamentally different subgroups from a large birth cohort, Maziade and coworkers (Maziade, Coˆte´, Boutin, Bernier, & Thivierge, 1987) undertook a longitudinal study of the association between temperament measured in children at 4 and 8 months and IQ assessed at 4.7 years. The data suggested a strong effect of extreme temperament traits on IQ development in middle and upper socioeconomic classes and in families with superior functioning in terms of communication. The temperamentally difficult group unexpectedly displayed higher IQ’s, and the well-replicated effect of socioeconomic status on IQ development was observed mainly in this

212 SECTION | C Collective control and communication

group. According to Belsky and Pluess (2013) difficult babies are more open to experience and do better if parents can cope, but worse if they can’t. We ourselves developed a parental support and education program called ‘Leaping Hurdles’ for a group of single mothers who were at risk of abusing their infants (van de Rijt-Plooij, van der Stelt, & Plooij, 1996). In it we made parents aware of the regression periods, showed them that their babies could not help being difficult and how they could comfort their babies in these difficult periods and facilitate the new type of perception and consequent learning. In an evaluation study we compared an experimental group receiving the program ‘Leaping Hurdles’ with a control group following a comparison program. The effects of the program Leaping Hurdles on the parents were as follows: the parental judgment of the temperament of their baby changed in the sense that it was based on different information, specifically, on the behavior of the infant instead of their own, parental rules and restrictions. Their judgment of the development of their baby was based more on information concerning the mental development instead of the motor development. Finally, the program gave parents a greater sense of control. The effects of the program Leaping Hurdles on the infants were as follows: First, the infants scored much higher on the mental Bayley scales. Second, the program removed gender differences in developmental test scores, the girls did not get lower scores on the Bayley. Third, the program children were socially more accepting and open toward strangers and not fearful and reserved as the control children were. No differences were found in the type of attachment, although this may have been an artifact of the Strange Situation scoring protocol failing to pick up differences. Finally, the program had positive effects on the health of the infants, specifically the girls. So, it is likely that regression periods have the function of activating family resources and thus promoting intellectual and social development as well as physical health (at least as long as the demands of the infant do not exceed parent’s capacity to meet them, a capacity which can be enhanced by appropriate knowledge and support). The improved physical health would promote survival directly, whilst the improved intellectual and social development are likely to do so in the longer term.

Discussion Summarizing our explanation of the findings, the following picture emerged. At gene-regulated ages, a new type of perception emerges intrinsically and is superposed onto the already existing hierarchy of levels of perception, resulting in a disruption in the behavioral organization. Consequently, the baby is stressed and gets closer to the parent, and the next regression period has started. Regression facilitates a trial-and-error process that is aimed at

The phylogeny, ontogeny, causation and function Chapter | 8

213

reducing the stress. This process is called reorganization. However, the baby has very little control over the outer world and needs his carer to reduce the stress and restore homeostasis. The resulting, more intensive caretaking period and parent-infant interaction can involve parent-infant conflict which pushes the baby toward more independence and exploration of the new perceptual world within the boundaries of its maturational abilities. A new type of learning results in a new set of skills apt for controlling the new perceptual world. There are individual differences in the age at which babies master a particular skill, depending on the personal preferences of the baby and the prevailing circumstances of the (social) environment. When the next level of perception emerges, the whole process starts all over again. The findings bring up a number of questions or statements that will be dealt with in the remainder of this chapter.

The field of ethology is ready for a paradigm shift toward PCT Without being exhaustive, the following few examples concerning some major branches of research in ethology suggest the field of ethology is ready for a paradigm shift toward PCT. Cools (1985) tested the PCT model experimentally through chemical stimulation of the brain of adult rats, cats, and monkeys and argued that it is the perception that is controlled through behavior and a hierarchy of feedback systems in the nervous system. This is quite the opposite of classical ethology’s presumption that it is behavior that is controlled, not perception. The behavior was considered to be controlled by a hierarchical system of Innate Releasing Mechanisms (IRMs) on the one hand and a hierarchical system of motor centers on the other hand. Each motor center controls a configurational pattern of muscle contraction. The IRM was supposed to remove a block preventing continuous discharge of a motor center (Tinbergen, 1951/1974) (pp. 103e104). Since Cools’ publication, not much has changed in ethology. If one looks at the box diagrams of the ‘hunger system’ (Fig. 2) or the ‘dustbathing system’ (Fig. 3) in a paper in 2005 reviewing the trends since Tinbergen (Hogan & Bolhuis, 2005), there is still a hierarchy of perceptual mechanisms on the one hand and a hierarchy of motor mechanisms on the other hand, and there is still a linear causation from stimuli to behavior. The paradigm shift toward PCT was then already long overdue. The latter notion is supported by the work of Pellis and coworkers (Pellis, Gray, & Cade, 2009) on crickets. These authors posit that a PCT view of behavior is applicable to the kinds of actions often labeled as ‘Fixed Action Patterns’ (FAPs) or ‘Modal Action Patterns’ (MAPs), that are core concepts of classical ethology. They have shown that such action patterns are not that fixed after all and that the variability serves to maintain some constancy or invariance. “Variability in individual movements by different body parts can be accounted for as compensatory actions that are enacted to preserve the

214 SECTION | C Collective control and communication

invariant features.” Examples of such invariant features are “some fixed relationship between body parts, or of some body part to an environmental cue (animate or inanimate), or to some dynamic aspect of the pattern performed (e.g., displacement, velocity, etc.).” Pellis and coworkers give examples of such invariances for many other species. They explain such invariances in PCT terms: “there is a reference signal that sets the value that the system maintains as constant, and variable behavioral output is what the system does to protect against deviation from the reference signal (Powers, 1973).” PCT may explain the fascinating pattern in the foraging and grouping behavior of social insects and of primates and in the co-ordination in swarms of insects, of fish, and of birds in a small flock by assuming that each individual simply controls its own perceptions according to a reference value or target value. Most astonishing are the complex maneuvers in flocks of tens of thousands of European starlings over the sleeping site at night before they settle in the trees. And yet, explanation of these complex phenomena does not need the assumption that the perception of the complete flock or any leadership or complex cognition is needed. Hemelrijk and coworkers (Hemelrijk, van Zuidam, & Hildenbrandt, 2015; Hildenbrandt, Carere, & Hemelrijk, 2010) have shown that the flocking maneuvers of starlings result from local interactions only. These are the following: “Individuals avoid those that are too nearby, they align with those at intermediate distance, and are attracted to those further away.” In addition, they interact with only six to seven of their nearest neighbors, they remain above their sleeping site at a certain height, and they follow the simplified aerodynamics of flight. Unfortunately, Hemelrijk and coworkers claim that these complex phenomena can be understood with the help of ‘complexity science’ and models of self-organization. But what is complexity science, really? Phelan (2001) has tried to answer this question and concluded that “While it has been relatively simple to show high-level resemblances between the emergent properties of computer models and realworld phenomena, it has proven extremely difficult to calibrate these models to produce correlations or confirmable regularities of real-world systems.” PCT may provide such models, as shown by McPhail and coworkers (McPhail, 2000; McPhail, Powers, & Tucker, 1992; Tucker, Schweingruber, & McPhail, 1999). Their models simulated collective action, arcs and rings in temporary gatherings of humans. While Hildenbrandt et al. (2010) (p. 1350) admitted that “social co-ordination depends on internal motivations,” they built their model in terms of (social) forces. In contrast, Bourbon (1995) suggested that the actions of animals in social groups such as flocks, schools, or swarms “might simply become coordinated with those of their immediate neighbors when each of them controls its own perceptions (p. 165).” In the simulations of McPhail and coworkers each individual simply controls its own perceptions according to a reference value or target value. In PCT, purposive behavior is a core issue. In classical ethology it was not seen as appropriate to “point to the goal, end, or purpose of behavior, or of any

The phylogeny, ontogeny, causation and function Chapter | 8

215

life process.” (Tinbergen, 1951/1974) (p. 4). Barrett (p. 100) states that PCT is an attractive theory because it helps us to work out what perceptions of the world an animal is trying to keep stable (Barrett, 2011). To give one more example, the field of birdsong learning would be appropriate to apply PCT. The diagrams researchers in this field draw are very much like the PCT diagrams; the feedback loop from moment to moment contains the comparator function comparing the song model (reference signal in PCT) with the perception (perceptual signal in PCT) of the young bird’s own emerging song via auditory feedback. If there is a mismatch between the bird’s current song and the song model, an error-signal is produced that results in distinct events of vocal change in the song production (output signal in PCT). This in turn changes the perception of the bird’s own song via auditory feedback. Some changes in the song production only take hours or days, others may take weeks. In the end, the song motif fully crystallizes and the bird rarely changes its song motif. Without mentioning PCT, Adret (2004) is searching for the reference signal (the template) when he says: “Moreover, songs of earlydeafened birds were highly degraded regardless of the amount of prior auditory exposure to conspecific models, suggesting that (1) central motor programs are not sufficient for fully normal song development and (2) selfgenerated auditory feedback is essential for the conversion of memorized songs into produced songs.” And Deregnaucourt et al. (2004) are searching for the error signal. In a recent review article (Bolhuis, Okanoya, & Scharff, 2010) this overall picture is confirmed: “during birdsong . learning . vocal motor output must be monitored continually through auditory feedback and if errors are detected the output should be adjusted.” Furthermore, recent findings are reviewed of how and where this happens. Three interconnected neural networks are mentioned. First, secondary auditory regions have to do with song perception and the recognition of tutor song. Second, the song motor pathway (SMP) has to do with song production and part of song learning. And third, the anterior forebrain pathway (AFP) is crucial for sensorimotor learning and adult song plasticity. The latter two neural networks are called the ‘song system’. In the AFP, a comparator function (in PCT terms) is postulated to be present that compares auditory feedback with the song model and sends an error signal to the motor system, if a difference is found between the desired outcome and the actual performance. The AFP originates in the nucleus HVC that has a functional position at the top of a sensorimotor hierarchy for song, and auditory-vocal ‘mirror neurons’ have been shown to be present in this nucleus (the so-called HVCx cells) (Mooney, 2014). This author concluded that “. the singing-related activity of HVCx cells can encode syntactic information about song in a hierarchical fashion, spanning from the identity of individual syllables to the number of repeated syllables and the nature of inter-syllable transitions.” As for the observation related activity of HVCx cells, “the auditory properties of mirror neurons in HVC are well suited to extract features that are important to song perception at multiple levels of acoustical

216 SECTION | C Collective control and communication

complexity,” including categorical perception of learned vocalizations (Mooney, 2014). There is clear potential for the perceptual hierarchy in PCT to be applied to these findings, potentially providing a more robust model.

PCT informs human developmental studies beyond the sensorimotor stage One may wonder what regressions might be occurring beyond 80 weeks of age - the time span of sensimotor development - given the huge amount of development that still happens thereafter, especially in terms of language and social interaction. Potentially, regression periods will be found at later ages to mark later steps in development. Although the buildup of the hierarchy ends when human development is far from complete at the end of the sensorimotor stage, the usefulness of the PCT model does not end there. As Barton (2012) clarified, the “key aspect of human cognition is. the adaptation of sensorimotor brain mechanisms to serve new roles in reason and language, while retaining their original function as well.” (p. 2098) Because of extended connectivity between the neocortex and the cerebellum, essentially the same kinds of computation appear to underlie sensorimotor and more ‘cognitive’ control processes including speech (p. 2101). Barton (2012) gives examples of computation and control processes at hierarchically different levels concerning events, spatial relationships, sequences and programs. More recently, Barton and Venditti (2014) have even shown that humans and other apes deviated significantly from the general evolutionary trend for neocortex and cerebellum to change in tandem. Humans and other apes have significantly larger cerebella relative to neocortex size than other anthropoid primates. This suggests that the current, almost exclusive emphasis on the neocortex and the forebrain as the locus of advanced cognitive functions may be exaggerated. Instead the cerebellum may play a key role in human cognitive evolution. Recently, Verduzco-Flores and O’Reilly presented a cerebellar architecture “allowing the cerebellum to perform corrections at various levels of a hierarchical organization spanning from individual muscle contractions to complex cognitive operations” (Verduzco-Flores & O’Reilly, 2015). In human adults there is growing evidence that the cerebellum is not limited to sensorimotor control (Manto et al., 2012), but plays important cognitive roles as well (Koziol et al., 2014; Stoodley, 2012; Timmann, Richter, Schoch, & Frings, 2006), including social cognition (Van Overwalle, Baetens, Marie¨n, & Vandekerckhove, 2014), emotion (Schmahmann, 2010), language (Argyropoulos, 2015; Highnam & Bleile, 2011), and even music and timing (E, Chen, Ho, & Desmond, 2014). In addition, there is evidence that “the cerebellum takes an early role in processing external sensory and internally generated information to influence neocortical circuit refinement during developmental sensitive periods” and thus influences cognitive development (Wang, Kloth, & Badura, 2014).

The phylogeny, ontogeny, causation and function Chapter | 8

217

The latter authors propose the ‘developmental diaschisis hypothesis’ that states that “cerebellar dysfunction may disrupt the maturation of distant neocortical circuits.” The above account indicates that the perceptual levels within the PCT hierarchy may form more complex functional processes where the same kinds of computation that have developed during the sensorimotor stage are used time and again to “serve new roles in reason and language, while retaining their original function as well,” (Barton, 2012) (p. 2098). The latter option is in line with the notion of embodied modes of thought (Barrett, 2011) and with the identification of “the origins of narrative in the innate sensorimotor intelligence of a hypermobile human body” (Delafield-Butt & Trevarthen, 2015). The latter authors “trace the ontogenesis of narrative form from its earliest expression in movement.” In light of this it is interesting to note that Homo sapiens is the only species that has language and the only species that has a life history with a childhood (Locke & Bogin, 2006). Childhood is the interval between infancy and the juvenile period. A great deal of language learning occurs during childhood. Bruce Nevin elaborates on a PCT view of language in Chapter 12.

PCT and understanding the evolution of human cognition PCT fulfills the need for a theory that incorporates the following elements: a complex interplay between biological evolution of new brainstructures enabling more complex hierarchical, domain general information processing and perceptual control on the one hand, and the discovery during the lifetime of an individual (through reorganization of the hierarchy) of new ways of producing cultural products on the other hand. This need was expressed by the new view about the evolved human mind (Heyes, 2012): “This new thinking about the evolution of human cognition (i) takes a longer historical perspective and therefore a more comparative approach, (ii) highlights the importance of co-evolution and cultural evolution in generating gradual, incremental change and (iii) suggests that humans are endowed with uniquely powerful, domain-general cognitive-developmental mechanisms, rather than with cognitive modules7.” Many of the articles in the ‘New Thinking’ theme issue (Heyes, 2012) suggest that humans are born with extraordinarily powerful cognitive-developmental mechanisms. These mechanisms are domain-generaldthey use a common set of computations to

7. The type of evolutionary psychology called the ‘Santa Barbara school’ or ‘high church evolutionary psychology’ (Cosmides & Tooby, 1987; Tooby & Cosmides, 2005) suggested that, in contrast to our primate relatives, we have a range of distinctive, special-purpose cognitive gadgets or modules, each responsible for thinking about a particular kind of technical or social problem that confronted our Stone Age ancestors. Experience was assumed to play a limited role in the development of these modules.

218 SECTION | C Collective control and communication

process information from a broad range of technical and social domainsd and they use experience, especially sociocultural experience, to forge new, more domain-specific cognitive-developmental mechanisms of the kinds that control tool-making, mentalizing, planning and imitating the actions of others. The genetically inherited cognitive-developmental mechanisms use computational processes that are also present in other animals, but they are uniquely powerful in their range, capacity and flexibility. The work in these articles shows a complex interplay between biological evolution of new brainstructures that enable more complex hierarchical information processing on the one hand, and the discovery of new ways of producing cultural products on the other hand. The various authors of that work express the need for a theory that accounts for both. PCT provides a useful functional model of these uniquely powerful, domain-general cognitive-developmental mechanisms suggested by the ‘new thinking’. And reorganization in PCT may explain co-evolution and cultural evolution in generating gradual, incremental change (Cloak, 2014; Powers, 2009/2014). According to Powers (personal communication) this is possible because the perceptual control hierarchy is the product of two reorganization systems, each operating by trial and error. The first, evolution, has produced variants over a very large number of generations of autonomous control systems, of which only some succeeded in bringing progeny to reproductive maturity. In consequence, a neonate begins life with many biologically inherited control systems in place and others emerging in the developmental process. The second reorganization system, an important form of learning, makes slight adjustments to synapse weights, amplification (gain), etc. around the loop in perceptual input functions, error output functions, and reference input functions until intrinsic error is decreasing (Powers, 2008).

Summary of the discussion The field of ethology is ready for a paradigm shift toward PCT. This is illustrated with a few examples concerning some major branches of research in ethology such as experimental neuro-ethology, fixed action patterns or modal action patterns, the fascinating pattern in the foraging and grouping behavior and in the co-ordination in swarms, purposive behavior, and birdsong learning. Although the build up of the hierarchy ends when human development is far from complete at the end of the sensorimotor stage, the usefulness of the PCT model does not end there. The perceptual levels within the PCT hierarchy may form more complex functional processes where the same kinds of computation that have developed during the sensorimotor stage are used time and again to serve new roles in reason and language, while retaining their original function as well. The cerebellum plays a much more important role in cognitive development than previously thought and the current, almost

The phylogeny, ontogeny, causation and function Chapter | 8

219

exclusive emphasis on the neocortex and the forebrain as the locus of advanced cognitive functions may be exaggerated. PCT fulfills the need for a theory that incorporates the following elements: a complex interplay between biological evolution of new brain structures enabling more complex hierarchical, domain general information processing and perceptual control on the one hand, and the discovery during the lifetime of an individual (through reorganization of the hierarchy) of new ways of producing cultural products on the other hand.

References Ader, R., Felten, D. L., & Cohen, N. (Eds.). (2001). Psychoneuroimmunology, 3rd ed., Vol. 2. New York: Academic Press. Adret, P. (2004). In search of the song template. Annals of the New York Academy of Sciences, 1016, 303e324. Aldridge, J. W., Berridge, K., Herman, M., & Zimmer, L. (1993). Neuronal coding of serial order: Syntax of grooming in the neostriatum. Psychological Science, 4, 391e395. Argyropoulos, G. P. D. (2015). The cerebellum, internal models and prediction in ‘non-motor’ aspects of language: A critical review. Brain and Language. https://doi.org/10.1016/ j.bandl.2015.08.003. Baerends, G. P. (1956). Aufbau des tierischen Verhaltens. In J. G. Helmcke (Ed.), Handbuch der Zoologie. Berlin: Walter de Gruyter. Baerends, G. P. (1970). A model of the functional organization of incubation behaviour. Behaviour.Supplement, (17), 263e312. Baerends, G. P. (1976). The functional organization of behaviour. Animal Behaviour, 24, 726e738. Baerends-van Roon, J. M., & Baerends, G. P. (1979). The morphogenesis of the behaviour of the domestic cat. Amsterdam: North Holland Publishing Co. Barrett, L. (2011). Beyond the brain. How body and environment shape animal and human minds. Princeton and Oxford: Princeton University Press. Barton, R. A. (2012). Embodied cognitive evolution and the cerebellum. Philosophical Transactions of the Royal Society of London B Biological Sciences, 367(1599), 2097e2107. https:// doi.org/10.1098/rstb.2012.0112. Barton, R. A., & Venditti, C. (2014). Rapid evolution of the cerebellum in humans and other great apes. Current Biology, 24(20), 2440e2444. https://doi.org/10.1016/j.cub.2014.08.056. Bastock, M., Morris, D., & Moynihan, M. (1953). Some comments on conflict and thwarting in animals. Behaviour, 6, 56e84. Bell, M. A., Wolfe, C. D., & Adkins, D. R. (2007). Frontal lobe development during infancy and childhood: Contributions of brain electrical activity, temperament, and language to individual differences in working memory and inhibitory control. In D. Coch, K. W. Fischer, & G. Dawson (Eds.), Human behavior, learning, and the developing brain. Typical development (chapter 9). New York: The Guilford Press. Belsky, J., & Pluess, M. (2013). Genetic moderation of early child-care effects on social functioning across childhood: A developmental analysis. Child Development, 84(4), 1209e1225. https://doi.org/10.1111/cdev.12058. Bever, T. G. (Ed.). (1982). Regressions in mental development: Basic phenomena and theories. Hillsdale, N.J.: Lawrence Erlbaum Ass.

220 SECTION | C Collective control and communication Bickhard, M. (1992). How does the environment affect the person? In L. T. Winegar, & J. Valsiner (Eds.), Metatheory and theory: Vol. 1. Children’s development within social context (pp. 63e92). Hillsdale, New Jersey: Erlbaum. Bingham, H. C. (1928). Sex development in apes. Comparative Psychology Monographs, 5(23), 1e165. Bolhuis, J. J., Okanoya, K., & Scharff, C. (2010). Twitter evolution: Converging mechanisms in birdsong and human speech. Nature Reviews Neuroscience, 11(11), 747e759. https://doi.org/ 10.1038/nrn2931. Bourbon, W. T. (1995). Perceptual control theory. In H. L. Roitblat, & J.-A. Meyer (Eds.), Comparative approaches to cognitive science (pp. 151e172). Cambridge: MIT Press. Chevalier-Skolnikoff, S. (1974). The ontogeny of communication in the Stumptail Macaque (Macaca arctoides). Contributions to Primatology, 2, 1e174. Cloak, F. T. (March 3, 2014). A natural science of culture; or, a neurological model of the meme and of meme replication. Version 3.1. Retrieved from http://www.tedcloak.com/Model_of_ the_Meme.php. Coghill, G. E. (1929). Anatomy and the problem of behaviour. London: Cambridge University Press. Condon, W. S. (1979). Neonatal entrainment and enculturation. In M. Bullowa (Ed.), Before speech: The beginning of interpersonal communication (pp. 131e148). Cambridge: Cambridge University Press. Cools, A. R. (1985). Brain and behavior: Hierarchy of feedback systems and control of input. In P. P. G. Bateson, & P. H. Klopfer (Eds.), Perspectives in ethology (Vol. 6, pp. 109e168). New York: Plenum. Cosmides, L., & Tooby, J. (1987). From evolution to behavior: Evolutionary psychology as the missing link. In J. Dupre´ (Ed.), The latest on the best: Essays on evolution and optimality (pp. 276e306). Cambridge, MA, US: The MIT Press. Dawkins, R. (1976). Hierarchical organization: A candidate principle for ethology. In P. P. G. Bateson (Ed.), Growing points in ethology. Cambridge, England: Cambridge University Press. Delafield-Butt, J. T., & Trevarthen, C. (2015). The ontogenesis of narrative: From moving to meaning. Frontiers in Psychology, 6, 1157. https://doi.org/10.3389/fpsyg.2015.01157. Deregnaucourt, S., Mitra, P. P., Feher, O., Maul, K. K., Lints, T. J., & Tchernichovski, O. (2004). Song development: In search of the error-signal. Annals of the New York Academy of Sciences, 1016, 364e376. Diamond, A., & Goldman-Rakic, P. S. (1989). Comparison of human infants and rhesus monkeys on piaget’s AB task: Evidence for dependence on dorsolateral prefrontal cortex. Experimental Brain Research, 74(1), 24. Diamond, A., Werker, J., & Lalonde, C. (1994). Toward understanding commonalities in the development of object search, detour navigation, categorization, and speech perception. In C. Dawson, & K. Fischer (Eds.), Human behavior and the developing brain (pp. 380e426). New York: Guilford. Dunlap, A. S., & Stephens, D. W. (2014). Experimental evolution of prepared learning. Proceedings of the National Academy of Sciences, 111(32), 11750e11755. https://doi.org/ 10.1073/pnas.1404176111. E, K. H., Chen, S. H. A., Ho, M. H. R., & Desmond, J. E. (2014). A meta-analysis of cerebellar contributions to higher cognition from PET and fMRI studies. Human Brain Mapping, 35(2), 593e615. https://doi.org/10.1002/hbm.22194.

The phylogeny, ontogeny, causation and function Chapter | 8

221

Fagen, R. M., & Goldman, R. N. (1977). Behavioural catalogue analysis methods. Animal Behaviour, 25, 261e274. https://doi.org/10.1016/0003-3472(77)90001-X. Fentress, J. C. (1976). Behavioral networks and the simpler systems approach. In J. C. Fentress (Ed.), Simpler networks and behavior. Sunderland, MA: Sinauer Associates, Inc. Fentress, J. C. (1983). Ethological models of hierarchy and patterning of species specific behavior. In E. Satinoff, & P. Teitelbaum (Eds.), Handbook of behavioral neurobiology (Vol. 6). New York: Plenum. Golani, I. (1976). Homeostatic motor processes in mammalian interactions: A choreography of display. In P. P. G. Bateson (Ed.), Perspectives in ethology (Vol. 2). New York: Plenum. Hailman, J. P. (1967). The ontogeny of an instinct: The pecking response in chicks of the laughing gull (Larus atricilla L.) and related species. Behaviour. Supplement, (15), IIIe159. Hemelrijk, C., van Zuidam, L., & Hildenbrandt, H. (2015). What underlies waves of agitation in starling flocks. Behavioral Ecology and Sociobiology, 69(5), 755e764. https://doi.org/ 10.1007/s00265-015-1891-3. Heyes, C. (2012). Introduction. New thinking: The evolution of human cognition. Philosophical Transactions of the Royal Society of London B Biological Sciences, 367(1599), 2091e2096. Highnam, C. L., & Bleile, K. M. (2011). Language in the cerebellum. American Journal of SpeechLanguage Pathology/American Speech-Language-Hearing Association, 20(4), 337e347. https://doi.org/10.1044/1058-0360(2011/10-0096. Hildenbrandt, H., Carere, C., & Hemelrijk, C. K. (2010). Self-organized aerial displays of thousands of starlings: A model. Behavioral Ecology, 21(6), 1349e1359. https://doi.org/10.1093/ beheco/arq149. Hinde, R. A. (1953). Appetitive behaviour, consummatory act, and the hierarchical organisation of behaviour- with special reference to the great tit (Parus major 1). Behaviour, 5(1), 189e224. https://doi.org/10.1163/156853953X00113. Hinde, R. A., & Stevenson-Hinde, J. (Eds.). (1973). Constraints on learning. Limitations and predispositions. London-New York: Academic Press. Hofer, M. A. (1978). Hidden regulatory processes in early social relationships. In P. P. G. Bateson, & P. H. Klopfer (Eds.), Social behavior: Vol. 3. Perspectives in ethology. New York: Plenum. Hogan, J. A., & Bolhuis, J. J. (2005). The development of behaviour: Trends since Tinbergen (1963). Animal Biology, 55(4), 371e398. Horwich, R. H. (1974). Regressive periods in primate behavioral development with reference to other mammals. Primates, 15(2e3), 141e149. Horwich, R. H. (1989). Cyclic development of contact behavior in apes and humans. Primates, 30(2), 269e279. https://doi.org/10.1007/BF02381314. Humphrey, T. (1969). Postnatal repetition of prenatal activity sequences with some suggestions of their neuro-anatomical basis. In R. J. Robinson (Ed.), Brain and early behaviour (pp. 43e84). London: Academic Press. Hutt, S. J., & Hutt, C. (1970). Direct observation and measurement of behavior. Springfield, IL: Charles C. Thomas. Kolb, B., & Wishaw, I. Q. (1990). Fundamentals of human neuropsychology. New York: Freeman. ¨ bersicht der angeborenen Verhaltungsweisen des mittel- euroKortlandt, A. (1940a). Eine U pa¨ischen Kormorans (Phalacrocorax carbo sinensis [Shaw & Nodd.]), ihre Funktion, ontogenetische Entwicklung und phylogenetische Herkunft. Archives Neerlandaises de Zoologie, 4, 401e442. Kortlandt, A. (1940b). Wechselwirkung zwischen Instinkten. Archives Neerlandaises de Zoologie, 4, 443e520.

222 SECTION | C Collective control and communication Kortlandt, A. (1955). Aspects and prospects of the concept of instinct (vicissitudes of the hierarchy theory). Archives Neerlandaises de Zoologie, 11, 155e284. Koziol, L. F., Budding, D., Andreasen, N., D’Arrigo, S., Bulgheroni, S., Imamizu, H., et al. (2014). Consensus paper: The cerebellum’s role in movement and cognition. The Cerebellum, 13(1), 151e177. https://doi.org/10.1007/s12311-013-0511-x. Kozulin, A. (1990). The concept of regression and Vygotskian developmental theory. Developmental Review, 10, 218e238. Kruijt, J. P. (1964). Ontogeny of social behaviour in Burmese red junglefowl (Gallus Gallus spadiceus) bonnaterre. Behaviour. Supplement, (12), Ie201. Lindahl, L., Heimann, M., & Ullstadius, E. (2003). Occurrence of regressive periods in the normal development of Swedish infants. In M. Heimann (Ed.), Regression periods in human infancy (pp. 41e55). Mahwah, NJ: Erlbaum. Locke, J. L., & Bogin, B. (2006). language and life history: A new perspective on the development and evolution of human language. Behavioral and Brain Sciences, 29(3), 259e280. https:// doi.org/10.1017/S0140525X0600906X. Mahler, M. S., Pine, F., & Bergman, A. (1975). The psychological birth of the human infant: Symbiosis and individuation. New York: Basic Books. Manto, M., Bower, J. M., Conforto, A. B., Delgado-Garcı´a, J. M., da Guarda, S. N. F., Gerwig, M., et al. (2012). Consensus paper: Roles of the cerebellum in motor controldthe diversity of ideas on cerebellar involvement in movement. The Cerebellum, 11(2), 457e487. https:// doi.org/10.1007/s12311-011-0331-9. Marken, R. S. (2013). Taking purpose into account in experimental psychology: Testing for controlled variables. Psychological Reports, 112(1), 184e201. Marsden, C. D., Merton, P. A., Morton, H. B., & Adam, J. E. R. (1978). The role of afferent feedback in the regulation of movement. In D. J. Chivers (Ed.), Behaviour: Vol. 1. Recent advances in primatology. London: Academic Press. Maziade, M., Coˆte´, R., Boutin, P., Bernier, H., & Thivierge, J. (1987). Temperament and intellectual development: A longitudinal study from infancy to four years. American Journal of Psychiatry, 144(2), 144e150. McGraw, M. B. (1945/1974). The neuromuscular maturation of the human infant. New York: Hafner Press. McKenna, J. (1990a). Evolution and sudden infant death syndrome (sids), Part I: Infant responsivity to parental contact. Human Nature, 1, 145e177. McKenna, J. (1990b). Evolution and the sudden infant death syndrome (SIDS), Part II: Why human infants? Human Nature, 1, 179e206. McKenna, J., & Mosko, S. (1990). Evolution and the sudden infant death syndrome (sids) Part III: Infant arousal and parent-infant co-sleeping. Human Nature, 1, 291e330. McPhail, C. (2000). Collective action and perception control theory. In D. L. Miller (Ed.), Introduction to collective behavior and collective action (2nd ed., pp. 461e465). Prospect Heights IL: Waveland. McPhail, C., Powers, W., & Tucker, C. (1992). Simulating individual and collective action in temporary gatherings. Social Science Computer Review, 10, 1e28. Mooney, R. (2014). Auditory-vocal mirroring in songbirds. Philosophical Transactions of the Royal Society of London Series B Biological Sciences, 369(1644), 20130179. https://doi.org/ 10.1098/rstb.2013.0179. Mounoud, P. (1976). The development of systems of representation and treatment in the child. In B. Inhelder, & H. Chipman (Eds.), Piaget and his school. A reader in developmental psychology (pp. 166e185). Berlin: Springer Verlag.

The phylogeny, ontogeny, causation and function Chapter | 8

223

Pellis, S. M., Gray, D., & Cade, W. H. (2009). The judder of the cricket: The variance underlying the invariance in behavior. International Journal of Comparative Psychology, 22, 188e205. Peterfreund, E. (1971). Information, systems, and psychoanalysis. An evolutionary biological approach to psychoanalytic theory. New York: International Universities Press. Phelan, S. E. (2001). What is complexity science, really? Emergence: Complexity and Organization, 3(1), 120e136. https://doi.org/10.1207/S15327000EM0301_08. Plooij, F. X. (1978). De ontwikkeling van preverbale communicatie in de moeder- kind interaktie methodologische aspekten. In F. J. Mo¨nks, & P. G. Heymans (Eds.), Communicatie en interaktie bij het jonge kind (pp. 45e64). Nijmegen, Netherlands: Dekker & van de Vegt. Plooij, F. X. (1984). The behavioral development of free-living chimpanzee babies and infants. Norwood, N.J.: Ablex. Plooij, F. X. (2003). The trilogy of mind. In M. Heimann (Ed.), Regression periods in human infancy (pp. 185e205). Mahway, NJ: Erlbaum. Plooij, F. X., & van de Rijt-Plooij, H. (2003). The effects of sources of "noise" on direct observation measures of regression periods: Case studies of four infants’ adaptations to special parental conditions. In M. Heimann (Ed.), Regression periods in human infancy (pp. 57e80). Mahwah, NJ: Erlbaum. Plooij, F. X., van de Rijt-Plooij, H., & Helmers, R. (2003). Multimodal distribution of SIDS and regression periods. In M. Heimann (Ed.), Regression periods in human infancy (pp. 97e106). Mahwah, NJ: Erlbaum. Plooij, F. X., van de Rijt-Plooij, H. H. C., van der Stelt, J. M., van Es, B., & Helmers, R. (2003). Illness-peaks during infancy and regression periods. In M. Heimann (Ed.), Regression periods in human infancy (pp. 81e95). Mahwah, NJ: Erlbaum. Polit, A., & Bizzi, E. (1979). Characteristics of motor programs underlying arm movements in monkeys. Journal of Neurophysiology, 42, 183e194. Powers, W. T. (1973). Feedback: Beyond behaviorism. Science, 179(4071), 351e356. Powers, W. T. (1973, 2005). Behavior: The control of perception (2nd ed.). Bloomfield, NJ: Benchmark Publications. Powers, W. T. (1978). Quantitative analysis of purposive systems: Some spadework at the foundations of scientific psychology. Psychological Review, 85(5), 417e435. https://doi.org/ 10.1037/0033-295X.85.5.417. Powers, W. T. (2008). Living control systems III: The fact of control. Benchmark Publications. Powers, W. T. (2009/2014). From reorganization to evolution and back. In D. Forssell (Ed.), Perceptual control theory. Science & applications. A book of readings (4 ed.). Hayward CA: Living Control Systems Publishing. van de Rijt-Plooij, H., & Plooij, F. X. (1987). Growing independence, conflict and learning in mother- infant relations in free-ranging chimpanzees. Behaviour, 101, 1e86. van de Rijt-Plooij, H., & Plooij, F. X. (1992). Infantile regressions: Disorganization and the onset of transition periods. Journal of Reproductive and Infant Psychology, 10, 129e149. van de Rijt-Plooij, H., & Plooij, F. X. (1993). Distinct periods of mother-infant conflict in normal development: Sources of progress and germs of pathology. Journal of Child Psychology and Psychiatry, 34, 229e245. van de Rijt-Plooij, H., van der Stelt, J., & Plooij, F. X. (1996). Hordenlopen. Een preventieve oudercursus voor de eerste anderhalf jaar. Lisse: Swets & Zeitlinger. Rosenblatt, J. S. (1976). Stages in the early behavioral development of altricial young of selected species of non-primate mammals. In P. P. G. Bateson, & R. A. Hinde (Eds.), Growing points in ethology. Cambridge: Cambridge University Press.

224 SECTION | C Collective control and communication Sadurni, M., Burriel, M. P., & Plooij, F. X. (2010). The temporal relation between regression and transition periods in early infancy. Spanish Journal of Psychology, 13(1), 112e126. Sadurni, M., & Rostan, C. (2002). Regression periods in infancy: A case study from catalonia. Spanish Journal of Psychology, 5(1), 36. Sadurni, M., & Rostan, C. (2003). Reflections on regression periods in the development of Catalan infants. In M. Heimann (Ed.), Regression periods in human infancy (pp. 7e22). Mahwah, NJ: Erlbaum. Schmahmann, J. D. (2010). The role of the cerebellum in cognition and emotion: Personal reflections since 1982 on the dysmetria of thought hypothesis, and its historical evolution from theory to therapy. Neuropsychology Review, 20(3), 236e260. https://doi.org/10.1007/s11065010-9142-x. Schneirla, T. C. (1950). The relationship between observation and experimentation in the field study of behavior. Annals of the New York Academy of Sciences, 51(6), 1022e1044. https:// doi.org/10.1111/j.1749-6632.1950.tb27331.x. Schore, A. (1997). Early organization of the nonlinear right brain and development of a predisposition to psychiatric disorders. Development and Psychopathology, 9, 595e631. Scott, J. P. (1986). Critical periods in organizational processes. In F. Falkner, & J. M. Tanner (Eds.), Human growth (2 ed., Vol. 1, pp. 181e196). New York: Plenum. Smotherman, W., & Robinson, S. (1990). The prenatal origins of behavioral organization. Psychological Science, 1, 97e106. Stoodley, C. J. (2012). The cerebellum and cognition: Evidence from functional imaging studies. The Cerebellum, 11(2), 352e365. https://doi.org/10.1007/s12311-011-0260-7. Ten Horn, J., & Paro, I. (1995). Mentale ontwikkeling bij baby’s: Try-out taakbatterijen. Heymans Bulletin Psychologische Instituten. Scriptie. Groningen: Ontwikkelingspsychologie. Rijksuniversiteit Groningen. Thelen, E. (1989). Self-organization in developmental processes: Can systems approaches work? In M. R. Gunnar, & E. Thelen (Eds.), Systems and development (pp. 77e117). Hillsdale, NJ: Erlbaum. Timmann, D., Richter, S., Schoch, B., & Frings, M. (2006). Cerebellum and cognition: A review of the literature. Aktuelle Neurologie, 33, 70e80. Tinbergen, N. (1951/1974). The study of instinct (2nd ed.). New York and Oxford: Oxford University Press. Tinbergen, N. (1963). On aims and methods of ethology. Zeitschrift fu¨r Tierpsychologie, 20, 410e433. Tooby, J., & Cosmides, L. (2005). Conceptual foundations of evolutionary psychology. In D. M. Buss (Ed.), The handbook of evolutionary psychology (p. 5). Hoboken, NJ, US: John Wiley & Sons. Trevarthen, C., & Aitken, K. (2003). Regulation of brain development and age-related changes in infants’ motives: The developmental function of regressive periods. In M. Heimann (Ed.), Regression periods in human infancy (pp. 107e184). Mahwah, NJ: Erlbaum. Trivers, R. (1974). Parent-offspring conflict. American Zoologist, 14, 249e264. Tucker, C. W., Schweingruber, D., & McPhail, C. (1999). Simulating arcs and rings in gatherings. International Journal of Human-Computer Studies, 50, 581e588. von Uexku¨ll, J. (1933). Streifzu¨ge durch die Umwelt von Tieren und Menschen. Frankfurt am Main: Fischer. Van Overwalle, F., Baetens, K., Marie¨n, P., & Vandekerckhove, M. (2014). Social cognition and the cerebellum: A meta-analysis of over 350 fMRI studies. NeuroImage, 86, 554e572. https:// doi.org/10.1016/j.neuroimage.2013.09.033.

The phylogeny, ontogeny, causation and function Chapter | 8

225

Verduzco-Flores, S. O., & O’Reilly, R. C. (2015). How the credit assignment problems in motor control could be solved after the cerebellum predicts increases in error. Frontiers in Computational Neuroscience, 9, 39. https://doi.org/10.3389/fncom.2015.00039. NA. de Vries, M. W. (1984). Temperament and infant mortality among Masai of East Africa. American Journal of Psychiatry, 141, 1189e1194. Wang, S. S. H., Kloth, A. D., & Badura, A. (2014). The cerebellum, sensitive periods, and autism. Neuron, 83(3), 518e532. https://doi.org/10.1016/j.neuron.2014.07.016. Werner, H. (1948). Comparative psychology of mental development. Chicago: Follett. Woolmore, A., & Richer, J. (2003). Detecting infant regression periods: Weak signals in a noisy environment. In M. Heimann (Ed.), Regression periods in human infancy (pp. 23e39). Mahway, NJ: Erlbaum.

Chapter 9

Social structure and control: perceptual control theory and the science of sociology Kent McClelland Department of Sociology, Grinnell College, Grinnell, IA, United States

Like psychology, sociology is a relatively young discipline. Although both disciplines can trace their lineage to earlier thinkers, these two ways of studying human behavior first came into universities as academic disciplines between 1880 and 1920, about the same time as other social sciences, such as anthropology, economics, and political science. Like these other social sciences sociology took its scientific paradigm from a set of research procedures modeled on the scientific laws of 19th century physics. In my own field of sociology, researchers sought to cast their observations of human social behavior in the form of scientific laws based on the empirical relationships of independent, intervening, and dependent variables. This causal modeling approach, although it proved useful for describing broad social patterns, left many questions about the social world unanswered. What is the origin of the empirical regularities sociologists have observed, and how do they develop? What maintains these social patterns, and what are the forces of social change? And when deviance from these norms is observed, how do we account for it? More profoundly, exactly what are these social structures that sociologists have endeavored to describe? In the 20th century, physicists turned to relativity theory and quantum mechanics, moving away from the earlier scientific-law perspective, but these hard-science developments had little impact on the research methods of sociology and other social sciences. Similarly, when engineers developed systems theory in the middle of the 20th century, some prominent social scientists became enthused for a time about cybernetics as a way of studying human behavior, but they eventually lost interest. So my training as a research sociologist in the latter part of the 20th century was focused almost entirely on the statistical intricacies of causal modeling.

The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00009-5 Copyright © 2020 Elsevier Inc. All rights reserved.

229

230 SECTION | C Collective control and communication

While I was in graduate school, it seemed to me that causal models provided the key to sociology as a science, but once I had done this type of research for several years, I was not so sure. When I finally encountered PCT, I was intrigued. Just as PCT had opened an entirely new perspective on psychology, I saw in this theory the promise of a new way of thinking about sociology. Over the last three decades, I have explored control-systems modeling as a way of understanding social interactions, and in this chapter I describe how the PCT perspective has led me to some novel ideas about what social structure is and how it works. A longer version of this chapter is offered as a digital supplement to this handbook, and in that version I go into much greater detail about the sociological implications of my work. Here, I lay out the basics of PCTinformed approach I have developed for understanding social life, which I think can shed new light on some of sociology’s central questions.

Overview of my argument The standard PCT model of a perceptual control system (see Preface in this volume) connects events occurring in a person’s brain and neural circuits with events occurring in the person’s physical environment. Control loops thus have two parts. The upper segment of the loop shown in PCT diagrams pertains to things happening internally in the body, where neural circuits convert sensory impulses into perceptions, compare these perceptions to their references, and transmit perceptual errors as commands for muscle movements. The lower segment of the control loop includes things happening in the external environment, where feedback functions link physical actions to sensory inputs. Researchers interested in applying PCT to psychology have usually focused on the upper half of the control loop, seeking to ascertain the perceptions a person controls and the references for those perceptions. I argue in this chapter that researchers interested in applying PCT to sociology and other sciences of human interaction need to focus more of their attention on the lower half of the control loop, on events in the physical environments that people share, which provide the nexus for their social interactions. My argument begins from the central premise of PCT that people control their own perceptions. To control their perceptions, they must continually engage in modifying the physical environments in which they live. The actions that people take to control their perceptions, if successful, counteract the forces in the environment that disturb those perceptions. Thus, people act to bring conditions in their environments, as they perceive them, into line with their own references and keep them there. In the course of everyday life, people constantly manipulate, arrange, and rearrange physical objects in their immediate environments, including and especially their own physical bodies. The essential point for sociology as a science is that these physical actions to control perceptions have physical effects that other people sharing the same environment can perceive.

Social structure and control Chapter | 9

231

The preferred research method of PCT, the Test for the Controlled Variable (TCV) (see Chapters 2, 3, and 16 in this volume), depends on the fact that the physical effects of perceptual control are perceptible to others. In the “coin game,” an example Powers often used to describe the TCV, one person observes another taking action to arrange a set of coins. The first person then disturbs this arrangement in some way, and watches to see if the second takes any physical action to rearrange the coins to match some reference condition known to the second person, but not yet known to the first. By disturbing the arrangement of the coins in a variety of ways and noting which disturbances are countered and which are not, the first person can usually guess the reference condition that the second must hold for the preferred arrangement of the coins. More generally, when people take physical action to arrange some aspect of the environment and then continue to act to maintain the stability of that arrangement in the face of disturbances, observers can reasonably infer that the people are imposing their own reference conditions on the aspect of the environment that appears to be stabilized. Many of the modifications people make in their physical environments stay in place well beyond the time that the person is actively controlling those perceptions. Although some of the physical effects of controlling perceptions are evanescent, disappearing almost immediately in the ordinary flux of disturbances, other physical effects of human action can last for months or even years, depending on the physical characteristics of the materials that have been modified. The arrangement of a bunch of coins on a tabletop is easily disturbed, for instance, but a carving made from a piece of wood is more resistant to disturbances, and a sculpture carved from a block of marble is even more so. Archeologists have dug up stone and metal artifacts made by ancient peoples that have survived thousands of years. Thus, the pockets of environmental stability created when people modify the physical environment to control their perceptions may long outlast the efforts of the individual controllers. Indeed, the point of modifying a relatively durable part of the environment in the first place is ordinarily to facilitate the long-run control of the perception that the maker of the modification had in mind. A good tool can make a worker’s job easier, for example, and that benefit can last through years of work. The environments in which we carry on our daily activities are packed with durable physical stabilities that have been expressly created to facilitate the control of certain perceptions. Unless we’re out hiking in the wilderness, nearly every object in our immediate environment is likely to be something created by human actions, which thus bears the imprint of someone’s previous attempt to control perceptions. The relatively permanent parts of our living and working environments, like buildings, furniture, clothing, vehicles, roads, and all kinds of infrastructure have been designed and manufactured not only with the control of certain perceptions in mind but also with the intention of making the control of those perceptions easier for users of

232 SECTION | C Collective control and communication

these objects into the indefinite future. Thus, our social lives take place in humanly modified environments that have been intentionally shaped to facilitate certain kinds of perceptual control. If social scientists want to infer the perceptions that people are seeking to control and their references for those perceptions, the place to start, I will argue, is by analyzing the socially constructed environments they have created. Careful analysis of physical stabilities in people’s living and working environments thus provides the empirical foundation for a new kind of social science. While physical objects and the part that they play in the construction of social life will be the initial focus of my argument, my view of the physical stabilities that make social life possible is actually much broader. Three additional ideas will complicate the picture I offer in this chapter. First, as Powers argued in putting forth his theory of the hierarchical organization of perception, people control perceptions at many different levels of complexity, from the low-level perceptions that enable us to navigate our physical environments, to the middle-level perceptions needed for rational decision making, to the highest levels of perceptions such as principles of behavior and social identities. I will argue in this chapter that modification of the physical environment is essential for controlling perceptions at all of these levels. Obviously, to navigate our physical environments, we must move objects around, including our own physical bodies. But when it comes to controlling more abstract, higher-level perceptions, our only recourse is to the same method of using our bodies to make physical changes in our environments. We control abstract perceptions by manipulating physical objects and other physical phenomena that we regard as symbolic of the abstract perceptions we seek to control. For instance, we often control abstract perceptions by using language. We speakdmake noises with our throats and mouthsdand writeduse our hands and specially designed tools to put marks on paper or electronic screens. We then offer those physical phenomena as symbols of the perceptions and references for perceptions that we want to communicate to others. Many other kinds of physical objects and bodily action can also serve to symbolize more abstract perceptions. In short, our only method for controlling all kinds of perceptions, from the most concrete to the most abstract, is by acting to modify our physical environments. The second important idea in this chapter is that my analysis applies to recurrent patterns of human activity, as well as to humanly modified physical objects. Like durable physical objects, recurrent patterns of action are physical phenomena that other people can employ to control their own perceptions. I have mentioned languages as physical phenomena that enable people to control many different kinds of perceptions. Languages are highly repetitive patterns of action, and when one person makes use of these recurrent patterns, other people can take it as signifying perceptions the speaker is controlling and thus can make inferences about the speaker’s

Social structure and control Chapter | 9

233

intentions. Social scientists have long been aware of the importance of language and communication in social life, so I’m not really saying anything new here, but my point is that the collective use of language creates environmental stabilities that people can then rely on for keeping their own perceptions in control. More broadly, all kinds of human behavior patterns contribute to the structuring of social life, and my analysis will have some new things to say about how such broad patterns are created, maintained, and changed. Finally, the third important idea underpinning my analysis in this chapter is collective control. When multiple people control the same or closely similar perceptions, their collective actions result in more stability in shared environments than any single person’s actions could produce. As I’ve argued previously (McClelland, 1994, 1996, 2004, 2006, 2014), social power arises from collective control, and powerful leaders exercise their power by specifying reference conditions for multiple followers to use for controlling their perceptions. Furthermore, even when collective control leads to conflicts, due to the different reference conditions participants bring to the interaction, such conflicts often produce relatively stable outcomes, although not necessarily the outcomes that any of the participants would prefer. Collective control, I will argue, is the bedrock of social stability, and social constructed environments result from the collective control of perceptions at many different perceptual levels simultaneously, thus combining stabilized physical objects and phenomena with predictable patterns of human actions to create the social and cultural structures that envelop our everyday lives. Because no person can see directly what is transpiring in the mind of another, the physical effects of other people’s actions provide our only means of understanding other people’s intentions. No direct brain-to-brain communication is possible. Thus, social interaction takes place entirely in the physical environment (including, of course, our physical bodies). My goal this chapter is to turn “control theory glasses” (Marken, 2002) on that physical matrix of interaction to show how the social stabilities that govern our interactions have come to be, how they are maintained, and why they change. I will begin by developing some new technical vocabulary for talking about the feedback portion of perceptual control loops, the part that passes through the physical environment. I then turn to an examination of collective control, showing how people can collectively control multiple levels of perceptions simultaneously. The next section of the chapter describes four kinds of collective control that are found in every social structure, and I offer a chart of the typical anatomy of a social structure. The final section of this chapter discusses the dynamics of social structures: the ways they form, are maintained, and change. As a whole, this chapter argues that sociologists and other social scientists can take a more scientific approach to their disciplines by adopting the powerful tools of controltheory analysis.

234 SECTION | C Collective control and communication

A control-theory analysis tool kit Atenfels: physical components of feedback loops Perceptual control theory offers a detailed picture of the hierarchical structure of control loops inside the brain and nervous system of the human body, but PCT theorists have generally had less to say about the segments of control loops that pass through the physical environment. The diagram of a “general model of a feedback control system and its local environment” offered in Behavior: The Control of Perception shows an arrow labeled “physical laws” linking the “proximal results of muscle tensions” to “remote physical phenomena,” and another arrow from “cause of disturbance” to the physical phenomena. A third arrow labeled “physical laws” links these physical phenomena to “proximal physical stimuli” (Powers, 1973, p. 61). The diagram thus depicts a chain of physical causation bringing feedback from muscle movements to sensory organs via physical phenomena that are subject to other disturbances. In his later diagrams of control systems, Powers sometimes described this causal chain as a “feedback function” (2008). For talking about the psychology of individual humans, as Powers did, vague terms like remote physical phenomena work fine, but for understanding how social interactions take place in the physical environment, we need terminology with a little more precision. One solution is to describe the chains of physical cause and effect as feedback paths between muscle output and sensory input. These feedback paths, however, may include many disparate elements, and to focus more precisely on the various physical phenomena involved in feedback paths, I make use in this chapter of a newly coined term, “atenfel.”1 The word atenfel refers to actual or potential components of environmental feedback paths, including states of physical objects or patterns of other organisms’ actions that can be used by an organism to keep its perceptions in control. Fig. 9.1 presents a diagram of an elementary control system from the middle of the perceptual hierarchy, and it also shows how this new terminology for elements of feedback paths applies to control loops. In Fig. 9.1, the label “Atenfels” is attached to the two arrows that represent the elementary control unit’s feedback paths through the external environment. The first arrow goes from muscles to a “Complex Environmental Variable” (CEV). The CEV is another label for “remote physical phenomena” in the sense that the CEV designates the entire set of physical variables that a person must manipulate in some way to keep a perception in control. The CEV is also the target of the physical forces that tend to disturb a controlled perception. The CEV thus includes all the physical phenomena corresponding to a given perception, and

1. “Atenfel” is a technical term invented by Martin Taylor and me. Taylor also uses the term in his chapter in this volume.

Social structure and control Chapter | 9

235

FIG. 9.1 An elementary control unit (ECU) in a neural hierarchy. Adapted from Fig. 10.2 in Taylor’s chapter, this volume.

we describe it as complex because any upper-level perception to which a CEV corresponds represents a complex combination of lower-level perceptions. A second arrow in Fig. 9.1 labeled as an “Atenfel” connects the CEV to the sense organs. These atenfels include things like light waves or sound waves that transmit physical energy from the CEV to the sense receptors of person’s body. Atenfels of this second type correspond to the “proximal physical stimuli” in the original diagram of a control system originally offered by Powers (1973, p. 61). To get a sense of what I mean by the word atenfel, consider an everyday example: a person pouring water into a glass. Assume that the person pouring a

236 SECTION | C Collective control and communication

glass of water seeks to control the perception that the glass is filling up, but not to the point that it spills. Thus, the complex environmental variable in question is the level of water in the glass. The reference for this perception takes the form of a goal: to fill the glass just full enough. What are the atenfels in the feedback path the person uses to control this perception? The feedback path starts with the person’s own hands and arms. From the point of view of the person’s nervous system, hands and arms are part of the external environment, and the muscles of the hands and arms provide an “interface” to the environment. The person’s muscles move the hands and arms, and hence the hands and arms serve as atenfels in the feedback loop. Suppose the person is grasping a pitcher of water in one hand and an empty glass in the other. The pitcher and glass also serve as atenfels in the chain of environmental causes and effects needed to control the perception that the glass is filling up with water. And the water that the person pours from the pitcher to the glass is another atenfel necessary for controlling this perception. The elements of the feedback path from the person to the CEVdthe level of water in the glassdthus include arms, hands, pitcher, water, and glass. The feedback path, however, does not stop there. The chains of physical causes and effects needed to complete the feedback paths from person to CEV and back to the person include the transfer of various kinds of physical energy from the water and glass to the person’s various sense organs. These transfers of energy provide the sensory feedback necessary for the ECU’s in the person’s neural hierarchy to construct the perception of the glass filling up with water. One type of sensory feedback is provided by light waves bouncing off the glass of water and striking the person’s eyes, allowing control systems in the person’s neural hierarchy to form a visual perception of the water level rising up the side of the glass. Another feedback path involves the sound waves that convey the rising pitch of the water’s gurgle as the glass fills up. Yet another channel of sensory feedback comes from the constantly changing weights of the pitcher and the glass in the person’s two hands, as water transfers from the one to the other. The changing muscle tensions required to keep the positions of the pitcher and the glass relatively steady, as well as the tactile sensations of the pitcher and glass gripped in the hands and fingers, provide haptic feedback that becomes part of the complex perception of pouring the water. Thus these various forms of physical energydlight waves, sound waves, and changing amounts of weight and tactile pressuredcomplete multiple feedback paths from the water glass to the sensory organs of the body. All of these forms of energy may be considered atenfels, since they all contribute to the person’s multimodal control of a perception that turns out to be surprisingly complex: the changing level of water in a glass.

Matching atenfels to perceptions The physical objects that most obviously fit the definition of atenfels are those ordinarily thought of as tools, implements, or utensilsdlike pitchers

Social structure and control Chapter | 9

237

and water glasses. Such objects are designed with a particular purpose in view or in other words as a means of controlling some particular perception. To think of arms and hands as atenfels may be more difficult, although anyone who has temporarily or permanently lost the use of limbs knows that a great many perceptions are much more difficult to control without the use of these appendages. To think of light waves or sound waves as atenfels may seem even more of a stretch. But imagine the difficulty of controlling the perception of filling a water glass in an environment of complete darkness and with some extremely loud noise blocking out the gurgle of water in the glass. Thus, the physical media of energy transfer from the CEV to the sense organs also serve as essential links in feedback paths. And without sensory access to the CEV that is the object of control, the control of a perception based on that variable is impossible. If we want to find out the perceptions a person is controlling, a first step is to take note of the atenfels the person is using. Because everyday manufactured objects are designed with ostensible purposes in view, one can easily match these objects with the perceptions people ordinarily control by using them. Coffee mugs, for instance, may often be used as atenfels in feedback paths for controlling the perception of having a cup of coffee. Pieces of paper and pens or pencils, together with hard surfaces to write on, regularly serve as atenfels for controlling the perception of jotting down a note. In many cases, then, an observer can make a good guess about the perception someone is trying to control simply by noting the objects the person is using. Matching physical objects to the perceptions they help to control can be a little tricky, however, because there isn’t necessarily a one-to-one relationship between physical objects and controllable perceptions. Individuals have the freedom to use physical objects in various ways for controlling their perceptions or to use a variety of physical objects in the control of any given perception. Hammers, for instance, are useful for driving nails, but a hammer can serve as an atenfel for many other kinds of tasks, like propping open a window, or scratching one’s back, or even as a weapon for defending oneself. For driving a nail, a hammer may usually be the preferred tool, but a rock, stick of wood, or metal pipe might also work.2 Thus, objects do not determine their own purposes. Rather, it is the humans who use objects who determine the purposes for them, and an object may be used as an atenfel for controlling many different perceptions. The atenfels associated with an object include all the potential purposes, even totally symbolic ones, to which a particular aspect of the physical environment could be put by a given individual or group of individuals with their particular perceptual hierarchies of controllable perceptions. Note also that a particular object may be used by one person as an

2. I’ve borrowed the hammer example from a 2012 post on CSGnet by Bill Powers (2012.11.07. 12,915 MST).

238 SECTION | C Collective control and communication

atenfel for controlling one perception and by another person as an atenfel for an entirely different perception, depending on the different purposes to which they put the object. Just as different objects may be used as atenfels for controlling a particular perception, any given object may provide an atenfel for multiple perceptions, depending on the ingenuity of the person using the object. Some objects are in fact designed with multiple purposes in mind. A Swiss Army Knife, with its assorted blades and other tools, is perhaps the archetype of such objects, but smart phones and tablet computers, with their built-in cameras and plethora of apps, also fit the description. The control of a particular perception may require the use of several objects together, sometimes in complex combinations. Examples of these combination atenfels include offices, factories, stores, restaurants, playgrounds, parks, and furnished homes.

Atenfels and the facilitation of feedback paths In many cases, to control a particular perception a person must have access to a specific kind of atenfel or atenfels. To drink a glass of water requires both some water and a glass. Enjoying the taste of a favorite food depends on having a portion of the food available to eat. To drive a car, one needs a car and the car keys. In other cases, one might control a perception in other ways, but the use of an appropriate atenfel can greatly facilitate the control of the perception. Consider the difference between hand tools used in carpentry and power tools. One can use a handsaw to cut a piece of wood or a hand drill to drill a hole in it, but a power saw or power drill will do these jobs more quickly and with less effort. Power tools can also extend the range of the perceptions possible for an individual to control. An object too heavy for an individual to pick up may not present any problem for an individual with a front-load tractor. In PCT terms, the use of a power tool as an atenfel increases the loop gain in the control loop. Loop gain is a measure based on the ratio of the output in a control loop to the perceptual error in the loop, or, in other words, a measure of the amplification of the signal in the loop. In practical terms, loop gain indicates the sensitivity of control loops to error, because the gain corresponds to how quickly and how thoroughly control systems act to reduce differences between their perceptual signals and their reference signals. Thus, control systems with high loop gain control their perceptions more tightly than systems with low loop gain. Tools of all kinds increase the loop gain for controlling perceptions by adding energy to the output or else directing the application of the person’s output energy more precisely, as when the blow of a hammer concentrates the energy of the person’s arm motion on the head of a nail. Thus, atenfels that facilitate control by making physical output more powerful or precise can mediate between the person and the complex environmental variable on the

Social structure and control Chapter | 9

239

output side of the feedback path. The control of perceptions can also be greatly enhanced by atenfels on the input side of the feedback path, the part that links the CEV to the person’s sensory organs. For example, microscopes, telescopes, amplifiers, and similar instruments for intensifying or clarifying sensory input can facilitate control by increasing the feedback gain on the input side. Thus, scientific-measurement instruments and graphical displays of data serve the function of increasing the loop gain, giving scientists greater control of the perceptions involved in making their observations. The availability in a person’s environment of a tool or other atenfel for facilitating control of a given perception does not force the person to control that particular perception, just as the scarcity of appropriate atenfels for controlling a given perception does not absolutely preclude a person from attempting to control it, but for most people in most situations the selection of atenfels available may limit the kinds of perceptions they choose to control. In other words, the availability of atenfels for doing things one way in an environment and the scarcity of atenfels for any other way of doing things will create “paths of least resistance” for controlling a perception. If a tempting snack is sitting on the table, for instance, the chances that a person will have something to eat are higher than if there’s no food in the house. Even with a bare cupboard and empty refrigerator, however, a person can still control the perception of getting something to eat, but only by relocating to a different environment to find the appropriate atenfels. Thus, the atenfels available in a given environment create “feedback paths of least resistance,” which encourage the control of some perceptions and discourage others.

The mirror world Our everyday lives have an “environmental impact,” including of course the impact of our actions on the world of nature, but in other ways, as well. The physical energy we expend in controlling our perceptions nearly always leaves observable traces of some kind on the physical environments in which we act. A careful examination of our living and working environments will reveal the “footprints” of our efforts to control our perceptions, as we constantly reshape these environments by arranging and rearranging the atenfels in our surroundings to facilitate control of our perceptions. The combinations of atenfels kept in place by a group of people can provide social scientists with empirical evidence about the perceptions the group is trying to control. The discipline of archaeology, of course, has always been based on the analysis of physical artifacts left behind by ancient peoples in order to understand the implicit purposes these artifacts once served. Social scientists studying contemporary societies might do well to follow that example. Taylor describes what he calls the “Mirror World” of complex environmental variables that reflect in their arrangement the perceptions people control internally (Taylor, 2019). He gives as an illustration the example of a

240 SECTION | C Collective control and communication

person (a) tensing fingers muscles (b) to strike keys on a computer, (c) to enter words into a file, (d) to write a report for a client, (e) to earn some money. These actions, he notes, form a hierarchy of perceptual goals, and successful control of each perception contributes to control of the next more complex perception in the hierarchy. Someone observing the person sitting at the computer would be able to see the environmental impacts of the person’s perceptual control, in the form of (a) muscle tensions moving fingers, (b) which cause computer keys to move, (c) which cause words to appear on a screen, (d) which result in a readable document, (e) which eventually yields a new bottom line on a bank statement. Each of these changing aspects of the physical environment, according to Taylor, is a complex environmental variable that corresponds directly to a perception controlled by the person working on the report. The hierarchy of CEV’s that have appeared in the physical world is a mirror image of the hierarchy of perceptions the person has controlled internally. To control perceptions people stabilize the complex environmental variables corresponding to these perceptions by opposing any disturbances to them, at least for the time that they actively control those perceptions. As Taylor’s example of the Mirror World shows, the CEV’s stabilized in controlling lower-level perceptionsdthe momentary holding down of a key on a computer keyboard, for instance, or the emergence of a sequence of letters on the screendbecome atenfels for the control of higher-level perceptions, like the writing of a report. The mirror worlds of local environments, in other words, reflect in a physically concrete way the histories of people’s attempts to control their perceptions. Of course, not everything people try to stabilize stays stable. Disturbances keep happening, including the impacts on local environments of other people’s attempts to control perceptions that are different from their own. One last but very important point about atenfels is that this term also applies to the physical actions of people, whenever these actions provide links in the feedback paths that other people use in controlling their own perceptions. In collective control, when people in a common environment make simultaneous use of some selection of the physical atenfels in that environment to control their own individual perceptions, these collectively coordinated actions can, in turn, serve as atenfels for them to control higher-level perceptions, including the “social” perceptions that sociologists have seen as characteristics marks of social structures. To understand this process of turning physical perceptions into social perceptions requires an understanding of the properties of collective control.

Collective control processes In previous articles (McClelland, 1996, 2004, 2006, 2014), I have presented results of my computational-modeling simulations of collective control

Social structure and control Chapter | 9

241

processes in which two or more simulated agents attempt to control their perceptions of an identical variable in a shared environment. As these articles have shown, the outcomes of collective control processes depend crucially on the degree of alignment (that is, similarity) between the reference values that the control agents use to control their perceptions of the environmental variable. Interactions involving collective control are not necessarily cooperative. When multiple participants use identical reference values to control their perceptions of a shared environmental variable, it increases the stability of the variable. The result is the same as if some super-powerful agent were tightly controlling its perception of the variable by applying to the control process an amount of loop gain equal to the sum of loop gains of the all the participating parties. In some sense, the power of a collective control process equals the sum of all the participant’s individual contributions. Conflict occurs, however, if the participants in a collective control process disagree in their choice of reference values for controlling their perceptions of the environmental variable. As Powers often explained (e.g., Powers, 2005: 266), conflict occurs whenever two or more control systems attempt to control their perceptions of the same environmental variable by using different reference values. The visible sign of conflict is that the participants begin working against each other, pulling in opposite directions, as it were, in what looks to outside observers like a runaway process of positive feedback, but which nonetheless results from the interaction of control systems operating by means of negative feedback (McClelland, 2014). As all the participants seek to reduce their own perceptual errors, their actions continually disturb the perceptions of other participants. Conflict is inherently inefficient for the participants, because they cannot control their perceptions as closely as they could in the absence of the actions of the others.

Cooperation and conflict An important but paradoxical finding from my simulations is that collective control processes involving conflict can stabilize an environmental variable as effectively as processes in which the reference values of the participants are perfectly aligned. Even when conflict is occurring, collective control can produce stability that seems as firmly set as if a super-agent were tightly controlling its perception of the variable with a loop gain equal to the sum of the contributions of all the participants involved in the process. The value at which the environmental variable is stabilized when conflict occurs, however, reflects no one’s choice, but is instead a “virtual reference level” (Powers, 2005, 267) that represents a compromise among all the participants’ differing preferences. Conflicts tend to escalate as long as all of the participating agents keep experiencing perceptual errors, or until enough of the participating agents

242 SECTION | C Collective control and communication

reach their maximum outputs. Thus, conflicts between control systems can simultaneously result in both escalation of output and stalemate of outcome, as is often the case in deeply entrenched conflicts between ethnic groups or political parties. In everyday interactions, however, not every conflict escalates into an all-out battle. People manage their conflicts in a variety of ways. In a perceptual-control hierarchy, even if a conflict prevents good control of a perception at one level, many higher-level perceptions can still be kept in reasonable control by working around the lower-level perceptions that are stuck because of the conflict, provided that the higher-level perceptions connect to enough combinations of lower-level perceptions so as not to have to depend on those particular lower-level perceptions for adequate control. In other words, when conflict prevents the use of one means for controlling important higher-level perceptions, another means can sometimes be found, allowing the person to turn attention away from control of the problematic lower-level perception and thus reducing the escalation of the conflict. Another way that people manage everyday conflicts is by reorganizing their higher-level perceptions to completely bypass perceptions stuck in conflict. When control of a given perception is unsuccessful for an extended length of time, the ongoing process of reorganization of neural connections in the brain increases with regard to the elementary control units in question (see the Preface of this volume for more on reorganization). If successful, the trial-and-error process of reorganization yields a new higher-level perception that does not need to draw upon the problematic lower-level perceptions. People may also reorganize their perceptual hierarchies to develop the perceptions necessary to recognize when conflicts are taking place and thus become conflict-avoidant, retreating from conflicts rather than engaging in them. In practical terms, talking to other parties in a conflict offers yet another technique of conflict management and often the most effective. If the differences between parties in their reference values for the contested variables are not too great, negotiations may help to get reference values back in alignment, or at least keep the differences within the bounds of toleration. Negotiations may also help disputants to go “up a level” by finding higherlevel perceptions on which they agree, and the cooperative control of those higher-level perceptions can then allow the flexibility to iron out or avoid lower-level differences. Seeing the conflict from the perspective of the other party may also provide a template for reorganization of new perceptions that makes the conflict unnecessary. Because each person’s perceptual world is unique, some degree of conflict is unavoidable in social interactions. Nevertheless, everyday interactions are more often cooperative than conflictive. The important point to remember here is that collective control can result in stabilization of some features of shared environments whether or not conflict is involved.

Social structure and control Chapter | 9

243

Giant virtual controllers: the social power of numbers The simulations reported in my previous papers (McClelland, 1994, 1996, 2004) have underlined the importance of numbers of participants and alignment of their references in the creation of social power. Collective control tends to stabilize the complex environmental variables linked to the perceptions being controlled, and in general the more individuals who participate in a collective control process, the more firmly fixed that stability. Any single individual who attempts to control his or her own perceptions using reference values at variance with the ultra-stable outcome of a massive collective control process will encounter resistance from the combined control efforts of the great many others participating in that process. When the collective actions of perhaps millions of people have produced stable outcomes, the actions of any single individual can have only a negligible effect. Individuals who seek changes in these conditions are likely to be frustrated by the rigidity of the aspects of the shared physical environment stabilized by this collective control. Ultra-stabilized collective control processes give social power to majority groups in a population, whenever their references are reasonably well aligned, because they are able to shape the commonly shared physical environment to provide atenfels compatible with their own purposes, but not necessarily with the perceptions controlled by minority groups. In these cases, it is as if some “virtual social actor” (McClelland, 2006: 91) with super-human powers were in control, acting according to purposes that reflect the consensus of the dominant group. Taylor (2019) has called these massive collective control processes “giant virtual controllers” (GVC’s), and he describes the environmental stabilities thus created as “collective complex environmental variables” (CCEV’s). I am adopting the same terminology in talking about these phenomena. Giant virtual controllers, I will argue, and the collective complex environmental variables stabilized by them, provide a macro context-a kind of social landscape-that constrains the ways in which smaller-scale social structures can be built.

Human activities as feedback paths Comparing human activities to physical artifacts Sociologists have always regarded recurrent patterns of human actions as a central focus of the discipline’s concerns, while generally paying less attention to physical objects. Looking with control theory glasses (Marken, 2002) at routine human actions, however, we can see some similarities between human actions and physical objects and treat them in much the same way. People’s physical actions, like physical objects, are observable phenomena that can serve as links in feedback paths for the control of perceptions. If, for instance, one person is controlling for seeing another person act in some specific way, then an action on the part of the second person that approximates the first

244 SECTION | C Collective control and communication

person’s references helps the first to keep perceptions in control. When a young child doesn’t misbehave in a social situation, for example, the parent may breathe a sigh of relief. Moreover, the action of one person may stabilize a physical object or an arrangement of physical objects that then allows another person to keep a perception in control, as when the first brings the second a cup of coffee. Thus, physical actions, like objects, can fit the definition of atenfels. People’s own physical actions are ordinarily the first steps in feedback paths for keeping perceptions in control in response to disturbances, but people can also incorporate other people’s actions into the feedback paths for controlling their own perceptions by making use of the physical stabilities those actions create. And, as with physical objects, observers of people’s routine patterns of action can gain clues for identifying controlled perceptions by watching to see which potential disturbances to routine actions are resisted and which are not. One important difference between routine human actions and physical objects is that actions must be performed anew every time they are to be used as atenfels, while many objects can be used over and over. In other words, actions used as atenfels lack the durability of some physical objects. Dances, theater performances, or conversations serve as atenfels while they are happening, but when they are over the only remaining feedback paths are the memories. Making a video recording of a dance or other performance, however, creates a physical object, which can then be used and reused as an atenfel for enjoying the performance. Written transcripts, similarly, can turn conversations into reusable objects. Such recordings of actions gain in permanence but lose in multi-sensory immediacy, as the many perceptual channels of the live experience are reduced to a set of visual images (perhaps with sounds) on a screen or a page. In sum, just as objects may be atenfels, actions too may be atenfels, although objects generally have more permanence than actions.

Collective control and levels of perception Like solo actions, collective human actions can produce stabilities in a shared physical environment that then serve as atenfels for individuals to use in controlling their own perceptions. It’s important to realize that this use of collective human actions as atenfels may occur at many different perceptual levels simultaneously. Take, for instance, a choir as an example of a group of people engaged in collective control. In their performances, members of a choir can collectively control perceptions ranging across all eleven levels of the hierarchy of perceptions proposed by Powers (2011), and their collective control of their own lower-level perceptions produces atenfels for them to control higher-level perceptions at the same time. Singing together, a choir can control perceptions of intensitydby singing louder or softerdsensationsdby intoning single pitches in unisond configurationsdby singing pitches that form a chorddtransitionsdby

Social structure and control Chapter | 9

245

changing from one chord to anotherdeventsdby singing a melody line or a progression of chords with a beginning, middle, and enddrelationshipsdby singing two melody lines in counterpointdsequences and categoriesdby singing a series of songs of a particular genredprogramsdby performing concerts on stagedprinciplesdby extreme care in staying on key or enunciating words in unisondand even system conceptsdby doing all the things a group must due to be recognized as a choir, from having a director and a name to following a schedule of group practices to wearing choir robes at performances to giving concerts or making recordings. When a choir sings a concert, all of these levels of collective control take place concurrently, and the choir’s collective control of their higher-level perceptions builds on their control of lower-level perceptions. At each level, the complex environmental variables made stable by collective control provide atenfels for the control of more complex perceptions at yet higher levels. The choir’s collective control of the configuration perception of singing of chord, for instance, requires that each group of choir membersdsopranos, altos, tenors, and basesdcontrol a different sensation-level perception of intoning a pitch. The physical sound of the chord sung by choir becomes an atenfel for controlling their perception of singing a progression of chordsda transitionlevel perceptiondwhile the sound of the chord progression becomes an atenfel for controlling the event-level perception of having sung a musical phrase, and so forth. Their collective control of an extremely high-level perception, such as the collective identity of belonging to a choir, depends ultimately on the perceptible physical impacts of having collectively controlled a host of CEV’s at lower perceptual levels, all of which have served as atenfels for successively higher levels of control.

Protocols: structural frameworks for dyadic interactions Though many face-to-face interactions between two people take no predictable form, some interactions fall into highly structured routines. In some routine interactions, people behave in parallel, that is, try to control the same kinds of perceptions at the same time, as in some of the examples of collective control that I’ve given above. Many rituals have this interactional form. In other interactions, a structure emerges from the joint performance of a protocol, which occurs when two people do different things, rather than trying to control similar perceptions, each seeking to control perceptions by means of the actions of the other, and thus acting in mutually beneficial ways. Taylor (this volume) argues that the structure of protocol interactions is based on the control of both the perception that led to the initiation of the protocol and control of a set of perceptions of belief about the intentions of the other person. Thus, in collective-control interactions two or more people concurrently control their perceptions of the same CEV’s (sometimes with different references), and in a protocol interaction the partners control their perceptions of

246 SECTION | C Collective control and communication

different CEV’s, with one person’s actions providing atenfels for the other person’s control, and vice versa. Taylor describes four basic types of protocolsdgetting help, giving help, teaching, and learningdand within those basic types he identifies a variety of subtypes, including trading, contracting, question-answer, buy-sell, and barter (Taylor, this volume). These various types of protocols share a common framework: to begin the interaction, one party disturbs a perception controlled by the other party in order to indicate that the first party seeks to control some perception with the assistance of the other. The second party then continues the interaction (or not), by acting so as to reduce the first party’s perceptual error (or not). The interaction ends with first party acknowledging the second party’s actions in a way that cancels the disturbance that initiated the interaction. Protocols are frequently more complicated than this basic outline, involving many back-and-forth and simultaneous moves, as well as attempts by one party or the other to repair any errors of communication that may occur. Taylor and his associates have put forward a “Layered Protocol Theory” along with “General Protocol Grammar” that offers a systematic way of mapping the possible moves involved in protocols (see Taylor’s chapter, this volume). As Taylor explains, different protocols differ in the extent to which they are routinized or ritualized, ranging from protocols that must be performed in the same way every time to those that require a great deal of on-the-spot improvisation. The degree of ritualization may vary, he argues, depending on the cultural context in which the interaction. To sum up, in everyday interactions between people, observable patterns of interaction gain their stability from the performance of complexes of collective control of well-aligned perceptions and protocols involving the collaborative control of a variety of perceptions at different perceptual levels. For a deceptively simple example that illustrates how everyday interactions can include complex combinations of protocols and collective control, consider the custom of a handshake, where one person extends a hand for the other to shake, and vice versa. A handshake is structured by a highly ritualized protocol, the initial stage of which requires that the two partners control symmetric perceptions, with each partner controlling for the other to perceive his or her willingness to shake hands with the other. To continue the performance of the protocol, however, each partner must independently control his or her own low-level perceptions of extending a hand, grasping the hand of the other, and then making an up-and-down motion. Simultaneously, at a higher level of perception, the physical actions of the other party (along with the haptic, kinesthetic, and visual feedback from the participant’s own actions) provide each party with atenfels for perceiving that a handshake ritual is in progressda perception that they collectively control. At some higher level, the handshake ritual may function as one step in a culturally

Social structure and control Chapter | 9

247

prescribed protocol for greeting old acquaintances or being introduced to new ones, in which the participants may play their parts by taking turns to choose from a variety of culturally stereotyped comments, thus controlling different perceptions in the service of collectively controlling a yet higher-level perception of sociability and goodwill between the two of them. In the handshake many culturally relevant messages may pass from one person to the other, such as the assertion of dominance by a strong grip and firm control of the hand motion, an acceptance of submissiveness by letting one’s hand be squeezed, the giving of affection by different subtleties of grip, and so forth. Such messages may in some cases be collectively controlled and distinctive across cultures, while others may be as cross-culturally relevant as the smiles that Darwin associated with all primates, or may be as idiosyncratic as the “secret handshake” that members of a secret society use when they meet other members of the club.3 As the name might suggest, protocols include routines of politeness, such as conventional greetings or farewells, and also the stereotyped exchanges of small talk. More broadly, commercial exchanges of all kinds fit into the category of protocols. Taylor (this volume) argues persuasively that many cultural patterns and the patterns of languages can also be fruitfully analyzed as protocols, and even complex conversations can be analyzed for their protocol structure. In all of these kinds of interactions, the participants control their own perceptions, which results in environmental stabilities that function as atenfels for other participants to control their own perceptions. In many kinds of protocols the participants also collectively control higherlevel perceptions of participating jointly in the protocol. Protocol structures may also apply to interactions between more than two people. Some familiar patterns of group interaction with one-person-to-many exchanges or many-to-many exchanges can be understood as protocols. For example, lectures, in which one person speaks to a group of others who attentively form an audience, and team sports, in which two teams vie symmetrically for a victory that only one can enjoy, fit this general pattern, in that these kinds of interactions combine collectively controlled perceptions with dyadic protocol loops that give the participants complementary “roles,” as Taylor describes it (Taylor, this volume). The example of two teams competing against each other also shows how the intertwining of collective control with protocols need not be entirely free of conflict. The two teams collaborate to control collectively the perception of engaging in the contest, but they use different references to control the perception of who will win.

3. My thanks to Martin Taylor for suggesting the examples in this paragraph (personal communication December 6, 2015).

248 SECTION | C Collective control and communication

Atenfels and human interaction Symbols and meaning from the PCT perspective If we were to observe two people shaking hands, we could describe their physical actions from a PCT point of view as serving to control their own individual perceptions by the performance of a protocol. However, we might also assume in common-sense terms that the handshake “means” something. As I noted above, one possibility is that the two people have met for the first time, and the handshake is one component of a more comprehensive protocol specifying the expected forms of polite behavior on making new acquaintances, so that it might be accompanied by stock phrases like “Pleased to meet you.” Another scenario could be that the people shaking hands are old friends who haven’t seen each other in a while, and the handshake is one indication of their pleasure in seeing each other again. Yet another possibility is that the two are heads of state, and the handshake seals an international agreement, a treaty that will bind the future governmental policies of the two countries. In each of these cases, the handshake could be understood from the PCT perspective as a protocol loop in which the participants control their own perceptions of physical motions and sensations, while these actions simultaneously serve as atenfels for the control of more complex, higher-level perceptions. The handshake in each of the three examples would be embedded in a more complex protocol performed by the two parties: a protocol for making new acquaintances, a protocol for greeting old friends, or a protocol for making treaties between countries. At some higher level of perception, the performance of these more complex protocols would provide the atenfels for the parties to collectively control the perception that friendly relations exist between them, or in the case of the handshake between heads of state, the perception that a deal has been sealed for some cooperative arrangement between the two countries. These higher-level perceptions, with the emotional resonances that accompany them, make the physical actions of the handshake meaningful to the participants themselves, as well as to any onlookers (assuming that the onlookers understand the cultural context of the handshake). Thus, in PCT terms, the meanings of the handshake for the participants and for onlookers include the set of higher-level perceptions to which the physically observable actions contribute, and we can say that the physically observable actions are symbolic of these higher-level perceptions. We see from these examples that the physical actions of a handshake can be a symbol of the quality of the relationship between two people who are shaking hands and, conversely, that their relationship can give meaning to their handshake. But how exactly are such meanings connected to the physical objects and actions that symbolize the higher-level perceptions that I’m describing as meanings? And how can the participants themselves know that a handshake means the same thing, or pretty much the same thing, to both of them? And what would enable an observer to interpret a handshake as

Social structure and control Chapter | 9

249

symbolic of any particular kind of relationship? Such issues of symbolic meaning have not yet been explored in any detail by PCT theorists, nor have they been the focus of PCT-informed research, but we can make some plausible guesses about how these processes may work by considering them in light of Powers’s proposal that control systems in the human brain are arranged hierarchically in interconnected layers.

Associative memory and the organization of the brain According to Powers’ theory of the hierarchical organization of the brain, each elementary control unit (ECU) in the brain sends its perceptual signals to multiple control units in higher layers while receiving reference signals from another perhaps overlapping collection of higher-order ECU’s. Similarly, each ECU connects to multiple ECU’s at lower levels, sending error signals to some (which are then received as reference signals) and receiving perceptual signals from others (refer to Fig. 9.1). Estimates of the number of synaptic connections that can be made from an individual neuron to other neurons range into the thousands or even tens of thousands,4 so it isn’t unreasonable to assume that a typical ECU connects to a very large number of other control circuits. The patterns of these connections, according to PCT, are established in the process of learning from experience, which Powers described as “reorganization” of connections between neural circuits (Powers, 2005). Brain research suggests that children’s brains start out with extremely dense interconnectivity between neurons, and that in child development a rewiring of these connections occurs as some synapses are strengthened, while others are pruned. The aphorism, “neurons that fire together, wire together,” conveys the finding from neurological research that when neurons repeatedly discharge in concert the synapses between them tend to strengthen, while other synapses tend to atrophy and disappear. Thus, we can assume that the interconnections among ECU’s are strengthened by experiences in which collections of them are active at the same time. These multiple connections between ECU’s at different levels of the hierarchy form the basis of associative memory, and associative memory is essential to meaning and symbolism. Consider what happens when an observer sees two people shaking hands. The observer’s shifting retinal images of the moving hands disturb the perceptions controlled by a set of low-level ECU’s in the observer’s brain. Because these ECU’s send their perceptual signals upward, news of these disturbances propagates upward through a ramifying network of higher-order ECU’s. As these changes in the input to higher-order ECU’s combine with streams of input to these circuits from other lower-order

4. See https://en.wikipedia.org/wiki/Neuron#Connectivity (retrieved December 19, 2015).

250 SECTION | C Collective control and communication

ECU’s, the resulting change in some cases will be sufficient to create perceptual errorsda mismatch between these inputs and their current reference signalsdand the higher-order ECU’s will then spring into action, sending outputs in the form of error signals cascading downward to change the reference signals for the latticework of lower-level ECU’s with which they connect. Powers argues that the reference signals coming from above serve as “address signals” that “select from lower-order memory those past values of perceptual signals that are to be recreated in the present time” (Powers, 2005, 219). Thus, disturbances to low-level ECU’s can activate associations in memory between perceptions controlled at different perceptual levels. Fig. 9.2 presents a schematic simplification of how this dynamic process of associative memory may work in the case of attributing meaning to a handshake, and how associative memories can serve as the basis of meaning and symbols. The figure represents a few of the many perceptions that may be controlled by an observer who witnesses people shaking hands and seeks to ascertain what it means. The figure shows connections between ECU’s at three different levels of perception in the observer’s brain. The boxes in the figure stand for the ECU’s and the arrows indicate associative connections between them. In the middle of the bottom row of Fig. 9.2 we see an ECU for the perception of “clasped hands,” which would be a fairly low-level perception in the observer’s neural hierarchy, although no doubt a perception constructed from large number of even lower-level physical perceptions of intensities of light and shade, different textures, and so forth. In the middle row, we see ECU’s for three somewhat higher-level perceptions of types of social interactions that may be indicated by a perception of clasped hands. Do the

FIG. 9.2 Associative connections between ECU’s in the process of symbolization and meaning.

Social structure and control Chapter | 9

251

clasped hands mean that one person has grabbed the other and is trying to pull him or her? Do they indicate a handshake? Or might the two be armwrestling? (Of course, clasped hands might signify many other perceptions, as well, not shown in the figure.) The top row of the figure shows yet higher-level perceptions of some protocols that might provide templates for two people’s interactions. As I suggested above, a handshake might be part of a protocol for meeting new acquaintances, greeting old friends, or enacting a treaty between countries. Of course, the clasping of hands could also symbolize even higher-order perceptions about the quality of the relationship between the two people, extending to other layers of perceptions not shown in the diagram. One protocol that might be relevant, for instance, would be a marriage ceremony in which the two hold hands as they repeat vows of honor and life-long fidelity. An observer familiar with the cultural context of the interaction would probably be able to sort through these many possibilities easily, settling immediately on some reasonable interpretation of the clasped-hands perception. To see how this interpretation of the meaning of the gesture may happen, consider the associative connections between the perceptions depicted in Fig. 9.2. The figure shows the perception of clasped hands sending perceptual signals upward to three of the many possible perceptions that might be associated with clasped hands: the grab-and-pull ECU, the armwrestling ECU, and the ECU for the handshake protocol. These higher-order ECU’s for perceptions probably act as something like logical on-off switches in a computer.5 An ECU for the handshake perception, for instance, could be either activedsending a perceptual signal for “handshake occurring” upward and sending its reference signal downwarddor else inactivedjust idling in the background, as it were, by not sending out any signals itself because the reference signal supplied to it from higher-order systems is effectively zero. Only one of the three mid-level ECU’s shown in Fig. 9.2 would be likely to be active at any given time, because the perceptions of shaking hands, arm wrestling, and grabbing and pulling someone are all incompatible with each other, at least in most imaginable situations.6 What, then, enables the observer’s perceptual hierarchy to settle on just one of these three plausible interpretations of the clasped-hands perception instead of the others? The other low-level perceptions evident to the observer could provide a context for assigning an unambiguous meaning to the

5. In a conference paper, Powers (2003, 13) argued that higher orders of perception appear to be “discrete variables, such as logical variables and category names” that do not take the form of the “weighted summations” familiar from most PCT simulations of hierarchical connections between levels. 6. Taylor has provided a detailed description of how this inhibitory process may work (Taylor, 2019).

252 SECTION | C Collective control and communication

observed perception.7 Since the brain constructs higher-level perceptions from combinations of many lower-level perceptions, a change in the perceptual signal reaching an ECU from just one of those lower-level ECU’s would not ordinarily be enough to switch the ECU from inactive to active mode. But when a large enough number of those lower-level ECU’s start sending up stronger perceptual signals, an ECU could become active and at the same time inhibit the activity of other ECU’s at the same perceptual level that might be incompatible with it. Fig. 9.2 shows the clasped-hands ECU as sending its perceptual signals up to all three of the ECU’s in the middle level of the figure. However, the figure also shows the ECU for “smiling expressions” as actively sending signals to the handshake ECU, but not to the other two perceptions at the middle level. The observer’s perceptual context in this example, then, includes the perception that the two people clasping their hands are also smiling at each other. In addition to the perception of smiling expressions, the observer’s perceptual context might include several other active ECU’s, such as a perception of up-and-down motions of the two hands, as well as perceptions of certain positions of the hands and arms and the angles of the elbows. Any combination of these lower-level ECU’s that was likely to send perceptual signals to the handshake ECU would not be equally likely to send signals to the armwrestling or grabbing-and-pulling ECU’s. Thus, the context of lower-level perceptions would, in effect, be consistent only with the handshake perception, allowing the observer’s perceptual hierarchy to rule out such alternative perceptions as arm-wrestling or grabbing and pulling. Once the context of the two people’s other observable actions has enabled the observer to bring the handshake perception into control, how would the observer go on to attach higher layers of meaning to the interaction? Fig. 9.2 shows the handshake ECU as sending its perceptual signals upward to ECU’s for the occurrence of three possible protocols: the introduction of new acquaintances, greetings between old friends, or the formal announcement of a treaty by heads of state. The observer could attribute a more comprehensive social meaning to the handshake interaction in any of these three ways (and, of course, in several other ways as well). Again, the context of the interaction would be likely to provide lower-level perceptions that are more consistent with one of these possibilities than the others. Fig. 9.2 shows the ECU for the “new acquaintance protocol” as receiving perceptual signals from the handshake ECU but also from a lower-level ECU for the phrase, “How do you do?” This ritual greeting between new acquaintances is rarely used in other contextsdcertainly not between old friends or between heads of state just concluding a difficult negotiation. Thus, a perception of the sound of this phrase, along with a variety of other clues

7. Nevin (this volume) emphasizes the importance of context in how we assign meaning to words.

Social structure and control Chapter | 9

253

from the observed postures and expressions of the two people shaking hands, and also perhaps from the observed actions of a third person performing the introductions, would allow the observer’s perceptual hierarchy to activate the ECU for the “new acquaintance protocol” and control for it while leaving ECU’s for the other possibilities inactive. Fig. 9.2 also shows downward connections from the new-acquaintance ECU to the handshake ECU and then on down from the handshake ECU to the clasped-hands ECU. A handshake is of course an expected part of the protocol for greeting new acquaintances, and clasped hands are expected to occur in a handshake. So each of the higher-level ECU’s, when actively controlled, will send error signals directly to the next lower-level ECU. In other words, once the interaction has been interpreted as meaning that the two people are just meeting each other, the observer’s perceptual hierarchy will be “looking for” a handshake to take place by sending error signals that contribute to the reference signals for the lower-level ECU’s controlling the clasped-hands perception, as well as other perceptions that contribute to the perception of a handshake. What in other circumstances might be regarded as a feeble excuse for a handshakedsay, limp wrists, fingers barely touching, and the merest hint of an up-and-down motiondmight well be perceived as an adequate handshake for the purpose by an observer who is controlling for the perception that the two people must be shaking hands because they have just been introduced. Thus, we see how the imputation of meaning and significance can travel both directions in a perceptual hierarchy. Physical events that evoke lowerlevel perceptions can provide atenfels for the control of the higher-level perceptions that are symbolized by these lower-level perceptions. At the same time, the ECU’s for these higher-level controlled perceptions, along with others that are activated in associative memory, can send error signals downward, which then may become reference signals for the same lower-level perceptions that have contributed to the higher-level perceptions of meaning. I have offered this example from the point of view of an observer of two people shaking hands, but similar processes of signification and meaning must take place within the perceptual hierarchies of the people who are engaged in the act. As noted earlier, the felt qualities of the handshake, whether distantly feeble or warmly firm, may send messages upward in hierarchies of the people themselves about their social status vis-a`-vis each other or their interest in the other person. Conversely, the participants’ high-level perceptions of the quality of their own relationship, combined with their perceptions of physical feedback received as the handshake takes place, may result in high-level error signals that become the reference signals for real-time corrections in low-level perceptions of hand pressures or arm movements, so that their perceptions of the physical motions of the handshake are controlled to be consistent with their perceptions of the quality of their social relationship (or perhaps the hoped-for relationship, if they’ve just been introduced).

254 SECTION | C Collective control and communication

One could elaborate these examples almost indefinitely, describing how layers of symbolic meaning can be understood as connections in associative memory between layers of perceptions at different levels. The main point I want to make, though, is that the PCT model of the architecture and dynamic functioning of the brain has the potential to take the field of sociology well beyond the commonplace observation that people use symbols in their social interactions by attributing meaning to objects, actions, and events. This theory opens the door to a scientific account of “meaning-making.” Although I have given no more than a sketch of this unimaginably complex process, the basic pattern offered in Fig. 9.2, when repeated on the scale of the billions upon billions of interconnections in the human brain, has the scope needed to account for the many kinds of meaningful connections we see ourselves and others make.

Meaning, symbols, language, and culture Spoken and written languages provide the most common examples in everyday life of perceptions at various levels linked by symbolization and meaning. Languages allow people to share their own perceptions with others, at least approximately so, and the means of this sharing depends on combinations of physical events or objects: arbitrary patterns of sound emitted from people’s mouths and arbitrary shapes written on pages or screens. These lowlevel perceptions, of course, come to be associated in memory with higherlevel perceptions of words that, in turn, are associated with other words and with nonverbal perceptions in many different levels of the brain’s hierarchy, perceptions that provide the meaning of the words and for which the words stand as symbols.8 Taylor (this volume) offers a PCT-based account of how the protocols that constitute languages may have developed, and Nevin (this volume) describes in detail the complex combinations of controlled perceptions needed for making our utterances and written words intelligible to others. Languages, obviously, provide much of what binds social structures together. Without going into the many complications, dealt with elsewhere in this volume, of how languages grow and are stabilized and then contribute to the maintenance of social stability, I want to note that language is yet another example of stabilized physical phenomenadboth the patterns of human actions as we speak and the configurations that we recognize as wordsdthat provide people with atenfels for controlling higher-level perceptions. But how is it that meanings are evidently shared among people or, in other words, that the same symbolic links between perceptions are found in the perceptual hierarchies of different people? The answer, of course, lies in 8. Nevin (this volume) suggests that perceptions of language may form a parallel hierarchy of perceptions at all of the different levels of the perceptual hierarchy for nonverbal perceptions.

Social structure and control Chapter | 9

255

the cultural environments that people share. In later sections of this chapter, I will explain how giant virtual controllers (GVC’s) can be understood to stabilize the cultural aspects of social structures. My argument is that GVC’s responsible for cultural perceptions create massive redundancies in the physical environments that people encounter. Young children hear same arbitrary combinations of sounds, for instance, over and over while in the presence of the same kinds of physical objects and human actions, and the reorganization process fixes associative memories between these sounds and the objects or actions in the brain, which then allow the sounds to be interpreted as words, and the meanings of the words to be associated with the objects or actions so consistently and repeatedly paired with them. In the same way, because of cultural redundancies, physical objects or actions can become commonly understood symbols for high-level perceptions, because those objects or actions are so often found in situations in which people also attend to the high-level perception. Consider, for example, a sports fan’s perception of a favorite team. Powers referred to this perception as an example of a systems concept, and he described such perceptions as almost “ethereal” because all of the more tangible lower-level perceptions that contribute to the perception of a team identitydthe players, coaches, owners, win-loss records, stadiums in which they playdcan change through the years without disturbing the fan’s perception that the team as an entity continues through time (2005: 172). But when the same fight song, the same uniforms, the same nickname, the same logo are seen at every game the team plays, the control of any one of these low-level perceptions soon becomes enough for fans to control in imagination the system-concept perception of the team, and thus these lower-level perceptions come to have their symbolic meaning in the identity of the team. This account, of how people can create the atenfels they need for controlling their higher-level perceptions in a shared physical environment stabilized by collective control and how these perceptions may correspond to the socially shared meanings that people attribute to everyday objects and events, completes the tools needed for constructing a PCT-based analysis of social stability. In the next section, I consider in more detail the way that collective control stabilizes the interaction and structure of social groups.

Collective control networks and social groups My analysis so far has focused on environmental stabilities generated as a byproduct of perceptual control processes. In this section, I turn my attention to the collectivities of people doing the controlling. I will use the label collective control network to designate any collection of people engaged in a collective control process. Collective control networks thus connect people who are using the same sets of atenfels to control their own independent perceptions.

256 SECTION | C Collective control and communication

The people in a collective control network need not be in direct interaction with each other, but if one person’s efforts to control a perception can influence, however minutely, the set of environmental arrangements from which a second person also draws atenfels for controlling a similar perception, I will describe the two people as being part of the same collective control process and thus connected in a network of collective control. An election for a political office provides one example of a collective control process. Voters in different localities typically do not have any direct interaction with each other, and they may differ in their references for the perception of which candidate to vote for, but each individual’s control of that perception has an impact on the overall outcome of the process, and thus, by my definition, all of the voters in a given election form a collective control network for that collective control process. This collective control network also includes, I would argue, all the eligible voters who abstain from voting in the election and whose failure to vote also influences the collective outcome. In discussing the behavior of collectivities of people, we address subject matter long considered central to sociology. Sociologists have always been interested in how social groups work: what holds groups together and splits them apart, how group boundaries are defined and maintained, and how people join and leave groups. Although the kinds of social groups most interesting to sociologists are more complex than a single collective control network, the concept of a collective control network provides one conceptual tool for constructing a theory of social groups from the PCT perspective and, more broadly, a theory of social structures. To join a social group, as I will argue, can be understood as becoming a participant in a many-layered complex of overlapping collective control networks in which the other members of the group also participate.

Scale and stability of collective control networks Collective control networks can range in size from as few as two peopled dancing together, for instance, or having a conversationdto millions of peopledall those participating in a national election, for exampledor even billions of peopledall the people living within the boundaries of a large country. The longevity of a collective control network depends in part on its size. In the smallest-scale collective control networks, which consist of only two people, whenever one person stops controlling his or her perception of the collective complex environmental variable that is the focus of the network, the control collective process ceases. Collective control networks of three or more people, however, do not necessarily disappear when one of the participants exits, because the remaining two can carry the process on. As the membership in a collective control network gets larger, the gain or loss of an individual participant makes less and less difference to the whole. Thus, collective control processes can last longer than any single person’s

Social structure and control Chapter | 9

257

participation in them, and the larger the scale of a collective control network, the smaller the impact of a single person’s entrance or exit. Large-scale collective control networks and the collective control processes they enactdsuch as the use of a spoken language by a language community or belief in a religious doctrine by members of an organized religiondmay outlast many generations of participants, with new recruits joining a collective control network and then eventually exiting, perhaps decades later. Furthermore, the atenfels used in large-scale collective control networks may often take the form of physical objects that can be standardized in their manufacture and marketed around the world, or else they may take the form of routine patterns of action simple enough to be followed by many different people in different locations. Because these kinds of atenfels can be duplicated endlessly, large-scale collective control networks may be widely dispersed in space as well as time. The atenfels to support the spread of a fad like a newly popular dance style, for instance, include recordings or videos of pop songs that are widely distributed together with sets of dance moves that are widely imitated. Everyone who then participates in learning the new dance becomes part of a widespread collective control network. In general, as the scale of collective control networks increases, their stability through time and across space increases as well, although the limited timespans of fads like dance styles show that even widespread collective control networks can rapidly come and go.

Multiple overlapping collective control networks Earlier in the chapter, I described how a choir can simultaneously control collective perceptions at many different levels of the perceptual hierarchy, ranging from low-levels perceptions of intensity, like the loudness or softness of their singing, to high-level perceptions of system concepts, like their collective sense of identity as a choir. The choir is a good example of a collective control network, and in looking at their actions with control theory glasses we see the concurrent activity of many different collective control networks, each network revolving around one of the many perceptions they control together. Not every member of the choir necessarily participates in every collective control network associated with the choir as a group. The sopranos and altos may control the perception of singing one melody, for instance, while tenors and basses sing another, all as part of controlling the their perceptions of performing a song together. The structure that emerges, then, in small groups like choirs consists of diverse collective control networks formed by the same cast of characters working together in various combinations. Some of these collective control networks are temporary, others long lasting, but they all interconnect and densely overlap. Collective control networks relating to lower levels of perception may spring into action but go dormant shortly thereafter, as people jointly make use

258 SECTION | C Collective control and communication

of the atenfels in their shared environment to control their dynamically changing perceptions. Just as the speed of processing of perceptions is faster in lower-level circuits, networks related to lower-level perceptions are likely to materialize and then rapidly vanish, as the members of the network move from using one object to another or one routine action to another as atenfels in their collective feedback paths. In general, the higher the order of perception of a collective control network, the longer it is likely to be active. For example, the collective control network for a choir’s sequence-level perception of singing a song might be active for only a few minutes, while their program-level perception of giving a concert might be active for an hour or more, and the collective control network relating to the system-concept level of belonging to the choir could last for yearsdhowever long the choir continued as a functioning organizationdeven while individual participants in that collective control network came and went. The overt actions that people take in their collective control of such highlevel networks are likely to be intermittent, as these networks are reactivated by their members on occasions on which they perform some function of the group, like a choir’s weekly practice session or a seasonal concert, but then set aside at other times, as members of the group go about the other business of their lives. It is as if their ECU’s for this perception control for a reference of zero during the meantime, but stand ready for members to activate them again at the next opportunity. Choir members would probably say yes, for example, if asked between practice sessions whether they belong to the choir, and at practice each week they may take up the control of their perceptions of participating in the choir pretty much where they left off the week before. Social groups, then, comprise many overlapping collective control networks of different durations, all featuring the same set of individuals as they pursue their joint activities. One might envision a kind of hyper-network made up of multiple layers of nested collective control networks, where the networks pertaining to lower-level, physically concrete perceptions serve as atenfels for the members of the group to use in controlling their intangible higher-level perceptions, including their perceptions of belonging to the group and perceptions of the group as a group. Thus, when choir members hum notes together and sing chords and melodies to perform songs to give concerts to act as members of a choir, they make the choir happen by collectively controlling their perceptions of it. Fig. 9.3 offers a schematic and drastically simplified sketch of a hypernetwork of collective control networks bringing together a group of interacting individuals. The figure shows four individuals (labeled A through D) linked in networks of collective control centered on nine perceptual variables (labeled 1 through 9). The variables fall into three broad levels of perception: variables 1 through 5 (shown as circles in the diagram) depict perceptions of the physical world of objects or actions, corresponding to the six lowest levels

Social structure and control Chapter | 9

259

FIG. 9.3 A hyper-network of collective control networks.

of perception in the eleven-level scheme proposed by Powers (2011); variables 6 through 8 (shown as triangles) depict middle-level perceptions of rational processing, such as the category, sequence, and program perceptions in Powers’s eleven-level scheme; variable 9 (shown as a star) represents a highlevel perception, such as the perception of a principle or system concept. The arrows in Fig. 9.3 represent two-way connections in control loops. Thus, these arrows are not quite the same as the arrows conventionally used in control-system diagrams (such as Fig. 9.1 above), which indicate the one-way flow of signals in the control loops in the brain and the causal chains of feedback-path connections through atenfels in the physical environment. The arrows here are double-headed to indicate the two-way connections of signification and meaning that occur between perceptions at different levels of the hierarchy, both input signals that flow upward and error signals that flow downward. The solid arrows that pass through the physical environment (in the connections to the lowest perceptual levels) indicate, in one direction, the downward flow of error signals in the perceptual hierarchy, creating reference signals for the control systems in muscles that produce physical actions and from there via atenfels in the physical environment to the complex environmental variables associated with the perceptions being controlled. In the other direction, these solid arrows trace feedback paths back from these complex environmental variables to the sense organs, which produce the signals within the individual’s brain that become the perceptual input functions that complete these control loops. The dashed arrows show two-way connections between controlled physical perceptions (1e5) and middle-level perceptions (6e8), as

260 SECTION | C Collective control and communication

well as between middle-level perceptions and the highest-level perception in the diagram (9). The dotted arrows show two-way connections from the midand high-level perceptions (6e9) to other control systems units in the perceptual hierarchies of the individuals controlling those perceptions. Fig. 9.3 represents just one moment in the stream of ongoing interaction of the group. A moment later, another set of collective control networks in the overall hyper-network of the group might be active, and some of those shown might go dormant, particularly at the lowest levels of the hierarchy (the perceptions depicted as circles), as the group members change the physical actions they are using as atenfels and also use different selections of objects in their environment as atenfels. This figure depicts the collectively controlled perceptions 1 to 9 as events occurring in the participants’ shared environment, but the situation is actually a bit more complicated. From a more sophisticated perspective than this diagram can offer, these are all individually constructed perceptions that are experienced internally. The feedback paths for controlling the lowest-level perceptions must pass through atenfels in the physical environment, but the connections between perceptions at different levels allowing the construction of middle- and higher-level perceptions, and even of most lowerlevel perceptions, take place internally as perceptual signals and reference signals flow up and down the perceptual hierarchies of the participants. However, the collective control processes depicted in the diagram affect the shared physical environment of the participants by stabilizing portions of that environment corresponding to the perceptions being controlled. Even if all the participant do not have exactly the same perceptions of what is going on between them, and despite any possible lack of alignment in their references for what should be going on, all their efforts to control their own perceptions will in combination determine a set of effects on the physical environment described earlier as the collective complex environmental variables (CCEV’s) corresponding to these perceptions. What the diagram shows, then, is a kind of convergence of the “mirror worlds” of each of the participants: the CCEV’s located in the physical environment shared by the actors, as they perceive it. Observers of the interactions shown in Fig. 9.3 might notice the physical stabilities created by these collective control processes, but unless their perceptions were similar to those of the persons shown in the diagram they would not necessarily interpret them in the same way. Observers unfamiliar with the cultural milieu in which the interactions are taking place would be like anthropologists observing the rituals of some indigenous group for the first time, able perhaps to see physical objects in use and patterns of interactions between people but clueless as to the symbolic meanings of those objects and patterns for the participants. In other words, these physical phenomena would not necessarily provide an observer who lacked the requisite connections in his or her perceptual hierarchy with the atenfels necessary for constructing the higher-level perceptions that the participants experience.

Social structure and control Chapter | 9

261

Fig. 9.3, then, takes neither the perspective of any single participant nor the perspective of an observer. The figure instead provides a PCT analyst’s view of the hyper-network, depicting both physical occurrences in the environment and perceptual phenomena in individuals’ brains.9 In a more detailed representation of a collective control hyper-network than this figure provides, clear distinctions would need to be made between the differing perceptions constructed from their various individual perspectives by the participants. This more detailed representation of the network from the analyst’s perspective would show the participants’ perceptions as individually distinct but overlapping, even though all of their perceptions draw on the same sets of physical stabilities in their common environment, which serve as the CCEV’s corresponding to the perceptions being controlled. Despite its drastically simplified form, this diagram still looks pretty complexdan indication of the truly formidable complexity of the kinds of social interactions I am attempting to depict. To unpack a little of the complexity, Figs. 9.4e9.7 prune some details away from Fig. 9.3 to reveal sets of subsidiary control networks embedded in the hyper-network depicted in the figure. The hyper-network depicted in Fig. 9.3, as I have just noted, is presented from the point of view of an analyst whose perspective embraces all of the control loops active at a given

FIG. 9.4 One Individual’s perspective on the collective control hyper-network.

9. Taylor (this volume) explains the distinction between the viewpoints of the controller, the observer, and the analyst.

262 SECTION | C Collective control and communication

FIG. 9.5 A middle-level collective control network.

FIG. 9.6 A lower-level collective control network.

moment, but the perspective of any individual participant in the hypernetwork is necessarily more limited. Fig. 9.4 focuses on the hyper-network of Fig. 9.3 as seen from the perspective of a single participant. In Fig. 9.4 only the links to collective control networks in which individual B participates are shown in black, while the other links in the hyper-network, those not involving individual B, are grayed out.

Social structure and control Chapter | 9

263

FIG. 9.7 A protocol loop embedded in the collective control hyper-network.

In Fig. 9.4 we see individual B’s contribution to the collective control of low-level perceptions 2 and 3, and via those perceptions to the control of mid-level perceptions 6 and 7, and ultimately to the control of high-level perception 9. Fig. 9.4 also shows individual B as connected to perception 8, although not directly contributing to the collective control of that perception. This part of the diagram is meant to indicate that individual B acts merely as a passive observer of perception 8, and thus makes no contribution to the control of this perceptual input, except perhaps to keep his or her perceptual input channels, such as eyes or ears, focused on the physical stabilities that provide atenfels for the individual’s construction of perceptions at this level. However, individual B uses the CCEV produced by the other participants’ control of perception 8 as an atenfel for controlling perception 9. Fig. 9.5 prunes away a different set of connections to focus on the collective control networks involving one of the middle-level perceptions. In Fig. 9.5, only the control loops contributing to the collectively controlled middle-level variable 6 are highlighted, while the remaining control loops, which contribute to perceptions 7, 8, and 9, are grayed out. In the example of a choir, a middle-level perception like number 6 might be the sequence of words and melodic phrases, controlled at lower levels of perceptions, which make up the song sung by the choir. Notice in Fig. 9.5 that perception 1, which is controlled individually by person A, does not contribute to the control of perception 6 and is thus grayed out. Fig. 9.6 prunes away additional details to focus on a single lower-level collective control network, the network for perception number 2. The figure grays out all of the control loops for higher-level perceptions, as well as for the

264 SECTION | C Collective control and communication

other lower-level perceptions. Since individual A is not involved in the collective control of perception 2, that individual is also shown in shadow. In the example of a choir, perception 2 could be a configuration perception, like a chord consisting of three tones sung together by three different voices, individuals B, C, and D. More detailed depictions of the overall complex of collective control networks might show the hierarchical connections between individually controlled perceptions at these lower levelsdthe sensation-level perceptions of the three different pitches, for exampledand this collectively controlled perception at the configuration leveldthe chord represented as perception 2. Fig. 9.7 presents one more diagram drawn from the hyper-network of control loops in Fig. 9.3. In order to highlight a two-person protocol loop within the overall pattern, Fig. 9.7 shows individual A controlling lower-level perception 1, with the CEV for perception 1 providing an atenfel for the control of middle-level perception 8 by individual B, while individual B participates in the collective control of perception 2, which similarly makes it possible for individual A to control perception 6. Thus, individual A’s actions help B to control a perception, while B’s actions help A to control a different perception, with neither individual’s actions contributing directly to the middle-level perceptions the other individual’s actions enable them to control. Most collective control hyper-networks probably contain numerous protocol loops like this embedded within them, with the protocol loops serving as building blocks for complexes of collective control networks. In the context of the choir, a protocol loop could occur when one singer’s performance of a musical phrase provides the cue for another singer’s entrance, and vice versa. A choir is a relatively simple organization, one which engages in a limited set of activities, and which would have an almost flat organizational chart, if it even needed one. Many groups and organizations in contemporary societies, like business corporations, have far more complicated layers of organizational hierarchy. Some, like families, also engage in a much wider range of activities and functions. The participants in these more complex social groups have sets of collective control networks that are no doubt far more numerous and varied than in a choir. The next section of this chapter presents a more comprehensive description of some of this complexity, offering generic definitions from a PCT perspective of social structures, culture, and social institutions. The theoretical concepts of atenfels, collective control processes, giant virtual controllers, protocols, symbolization, and collective control networks, all of which I have explored in this first half of the chapter, are the conceptual tools needed for this more complex analysis.

The anatomy of social structures My overall objective in this chapter is to offer a general theory of what social structures are and how they work. With its base in PCT, a general theory of

Social structure and control Chapter | 9

265

human behavior, the theory offered in this chapter is intended to apply to social structures of all kinds, from the most micro to the most macro. A social structure, as I am defining it here, occurs when an identifiable group of people, perhaps shifting over time, participate in a hyper-network of collective control networks that rely on a shared set of atenfels, including physical objects and patterns of human activity. I will refer to the groupings of people who participate in one of these hyper-networks as the members of the social structure. A social structure derives its dynamic stability from giant virtual controllers, which provide members of a social structure with a cultural framework for coordinating their activities. In this section of the chapter, I will attempt to unpack this densely technical definition and demonstrate some of its applications. Social structures to which this theory can apply come in an immensely wide variety of forms. Social groupings held together by relatively stable networks of social relationship and patterns of collective activity range from dyadic relationships, like long-term friendships or marriages; to small groups, like families, gangs, or the choir described in the preceding section of the chapter; to formal organizations, like businesses, clubs, or government agencies; to communities, like small towns, city neighborhoods, or monasteries; to large-scale groups like social movements and political parties; to nations and societies; to global formations like networks of international commerce or the networks made possible by social media; to social and cultural institutions, like languages or patterns of racial or gender relationships. In each case, I submit that we can analyze the structure of the grouping as a hyper-network of collective control networks.

The four main types of collective control networks in social structures Collective control networks are the basic building blocks of social structures. We can picture social structures as comprising collective control networks in interlocking layers or strata, from networks controlling lower-order perceptions at the base of a structure’s hyper-network, through layers controlling progressively higher-order perceptions, up to layers controlling very high-level and abstract perceptions at the top of the hyper-network. Some of these collective control networks, especially high-level networks that provide a cultural context for interaction, involve the participation of all of the members of the social structure. Many of the lower-level networks, however, can form and then dissolve fairly rapidly in the course of everyday interactions among members of a social structure, and in most cases only a subset of the structure’s members will participate in these networks. Starting with the lowest-order collectively controlled perceptions, members of a social structure stabilize complex environmental variables by controlling their concrete perceptions of the shared physical environment in

266 SECTION | C Collective control and communication

which they are interacting. I will describe these layers of control as providing the social structure’s material base. By controlling their perceptions of these physical objects, the structure’s members create feedback paths for themselves and others to use in controlling higher-level perceptions, thus providing the atenfels that make up predictable and familiar living or working environments. The most-often used feedback paths incorporating these atenfels then become feedback paths of least resistance, atenfels for the perceptions that people habitually control when they move about and do things without conscious thought. Since many of the physical objects stabilized by collective control are constructed from durable materials, thus remaining in the same form and location until impinged upon by some outside force or human action, the relative stability of the physical objects making up the common environment for interaction helps to stabilize the social structure over time. The next layers of collectively controlled variables relate to the characteristic patterns of activity, behavior, and communication of the members of the social structure. These layers of control provide the social structure’s behavioral base. As with perceptions of physical objects, participants create atenfels for themselves by controlling these routine actions and rituals in order to bring yet higher-level perceptions into control. The purposes of these routine actions are often evident to outsiders, since the physical actions of the structure’s participants, like the physical objects that provide feedback paths for them, are open to empirical observation. If, however, observers of routine actions lack the cultural knowledge of how things are done in the social structure under observation and how the members of the structure themselves interpret those actions, they may see people taking repetitive actions but have difficulty figuring out the purpose of what they are seeing, that is, understanding the higher-level perceptions symbolized by these observable actions. Yet higher layers of collective-control networks in social structures involve the control of abstract variables such as patterns of relationships, social identity, and belonging among members of the structure, as well as the relationships of people to things in their shared physical environment, like ownership of things or access to resources. These higher-order perceptions include program-level perceptions, such as the stories by which people narrate their personal histories as individuals in relation to other people and the things around them, as well as the principle-level perceptions that define their personalitiesdthe values and other traits seen as characterizing their behaviordand the system-concept perceptions that combine all these highorder perceptions to constitute the social identities of persons and groups. Sociologists have traditionally focused on these perceptions of social relationships and group memberships in attempting to describe social structures, but the perspective I am offering makes clear that these “structural” variables provide only part of the picture. Because variables at

Social structure and control Chapter | 9

267

different perceptual levels are so closely interrelated, any full description of a social structure must take into consideration the material and behavioral bases of social structures, as well. The ultimate layers of collective control networks in social structures control perceptual variables at high levels of abstraction that provide the culture of the social structure: norms and other rules for “correct” behavior; technologies for manipulating the physical environment; systems of knowledge, science, and religious belief; and culturally prescribed values and virtues. The collective control networks for such institutionalized cultural phenomena are giant virtual controllers (GVC’s) that may extend far outside any single social structure, sometimes extending to the society-wide or even global scale. Smaller social structures, like groups, organizations, and communities within a national setting, can borrow these abstract perceptions from the large-scale social structures in which they are nested, and the collective control networks for these cultural perceptions are not only broader in scale than the hyper-networks of single social structures, but also longer lasting, because such GVC’s do not depend on participation by any particular members of any single social structure. The smaller-scale social structures embedded within larger structures, like families, friendship groups, teams, and gangs, can also have collective control networks for their own distinctive cultural perceptions. These “idiocultures” (Fine, 2012) include collective control networks that maintain the group’s history, traditions, and distinctive norms and customs. Thus, the collective control networks for cultural perceptions, like other collective control networks, come in many different sizes. These cultural layers of collective control networks interpenetrate the three lower layers of networks in the sense that cultural perceptions provide references that serve as templates or patterns for the control of perceptions of objects and actions by the lower layers of collective control networks, as well as references for the perceptions of identities and relationships attributed to these objects and actions. If, in a given social structure, we observe that many objects are standardized and that people can use these standardized objects interchangeably, that people’s actions tend toward a predictable uniformity, that people attribute much the same kinds of meaning to the standardized objects and patterns of action around them, that successful communication and coordinated interaction take place routinely, all this is evidence of a stabilized cultural environment, which is maintained by the GVC’s in the highest layers of the social structure. Each individual participant in a social structure inevitably becomes involved in collective control networks of all four main types. First, the individual must come to recognize, navigate, and manipulate the physical objects that form the material base for the structure, the buildings, tools, clothing, vehicles, communications equipment, or whatever. Second, the individual must carry out patterns of action from the social structure’s

268 SECTION | C Collective control and communication

behavioral base in order to interact with others. These patterns often take the form of protocols, which are the basis of most interactions. Third, the individual must maintain a social identity within the structure and learn to recognize the identities of other participants, as well as relations of ownership between people and objects. Fourth, the individual must conform to the cultural norms characteristic of the structure, or else face the difficulties that arise when an individual attempts to resist these GVC’s. The individual must also acquire cultural knowledge about the structure and how it works in order to have smooth interactions with others in the structure. In sum, every social structure comprises collective control networks relating to four general types of perceptions: a material base of perceptions of physical objects and relationships between objects, a behavioral base consisting of patterns of routine human actions, layers of higher-level perceptions of social identities and perceptions of relationships among people and between people and objects, and diverse layers of higher-level cultural perceptions. One cannot make a full description of a social structure without giving attention to all of these kinds of collectively controlled perceptions. I turn next to ways that social structures may be nested or intermeshed with each other.

Embedding and interleaving of social structures The collective control networks of micro-scale social structures, including small groups and dyadic relationships between individualsdsuch as families, teams, or gangs in which the group members all have face-to-face contact with each otherdare often embedding in larger social structures, such as organizations, communities, or even societies, in which most or all of the members of the smaller social structures participate. Meso-scale social structures like organizations and communities also tend to be nested in yet larger structural units, such as societies.10 And the collective control networks of contemporary societies are at least partially embedded in even wider scale collective control networks like defensive alliances between countries, the United Nations, or global trading networks. However, the boundaries of the networks of smallerscale social structures do not always fit neatly within these larger units. Multinational business organizations, for instance, comprise collective control networks with members spread across different countries. In prehistoric times, the members of hunting and gathering groups almost never participated in collective control networks outside the boundaries of the social structures of their local groups. The only exceptions were trading networks set up by members of some groups for the exchange of goods and 10. On these more macro scales, the boundaries of social structures become permeable and uncertain, since people may participate in many of the lower-level collective control networks associated with these structures without perceiving themselves as belonging to the more comprehensive group.

Social structure and control Chapter | 9

269

resources with other groups of people in the area. However, the social structures of hunting and gathering groups usually contained various smaller-scale social structures, such as family groups or clans, men’s groups, or age-grade cohorts, and any given individual would also participate in a handful of such subsidiary structures (Flannery and Marcus, 2012). In contemporary societies, by contrast, individuals usually participate in a wide variety of social structures, including family networks, friendship groups, schools, neighborhoods, work groups, communities, and larger political structures like cities and countries. The list of structures in which members of contemporary societies often participate could be extended almost indefinitely, and many of these structures do not nest neatly in larger ones, as the social structures of contemporary societies overlap in complex ways. Because social structures in contemporary societies do not necessarily overlap, individuals may participate only intermittently in any given structure, perhaps by participating in different structures at different times of the day or in the presence of different collections of people, as when going to work or school. Nevertheless, in moving from participation in one social structure to another, individuals often carry along with them physical and behavioral manifestations of their history of participation in other structures. The physical features of a person’s body, like hairstyles and hair coloring or posture and musculature, or a person’s habitual ways of acting, like patterns of speech or ways of walking and moving one’s limbs, along with styles of dress and personal grooming, may all bear the cultural imprint of the person’s history of social-structure participation. When people enter new social structures they carry along these habits and physical features developed in previous social structures, including such high-level perceptions as funds of knowledge or principles of expected behavior. Job applicants, for example, whose looks or speech patterns betray a social-structural history that differs from the backgrounds of current jobholders may not impress interviewers, despite adequate qualifications in other respects.

A conceptual map of a social Structure’s collective control networks To summarize this discussion of the four types of collective control networks in a typical social structure, Fig. 9.8 presents a “map” of the types of perceptions controlled in the four layers of collective control networks forming the hyper-networks of social structures at the micro level. The figure is meant to show how these layers of collective control networks interrelate. The ovals labeled Material Base, Behavioral Base, Social Identities and Relationships, and Idioculture stand for the four layers of collective control networks, and the box around the ovals marks the boundaries of the social structure as a whole. People who participate in any given social structure participate in all of these types of networks. The ovals interpenetrate to show the interconnections between the various layers of collective control networks.

270 SECTION | C Collective control and communication

FIG. 9.8 Map of the collective control networks of a social structure.

Fig. 9.8 also shows larger gray ovals that depict the collective control networks of the Cultural Perceptions in which the members of the social structure participate, and which help to structure their interactions with other members of their own social structure and other structures. Cultural phenomena take a variety of forms, including norms, laws, and other rules for “correct” behavior; technologies for manipulating the physical environment; systems of knowledge, science, and religious belief; and culturally prescribed values and virtues. Sociologists often apply the label of “social institutions” to the most widespread cultural phenomena, particularly norms and values, although other forms of cultural knowledge are sometimes given that label, and invidious phenomena such as racism and sexism are frequently described as institutionalized inequalities. The labels in smaller type in the ovals indicate some of the many kinds of perceptions controlled by these layers of networks. Fig. 9.8 depicts only a single social structure, but in contemporary societies the members of a structure routinely participate in many social structures at once, each of which is composed of similar layers of collective control networks. Thus, a more comprehensive schematic diagram of all of the social structures in which members of a single social structure participate would

Social structure and control Chapter | 9

271

show these other nested or overlapping social structures. The overall picture would be extremely complex, and even a more complicated figure showing the many overlaps could by no means do justice to the complexity of social structure as a phenomenon. Conceptualizing social structure in this way, however, moves us far beyond the vague metaphors sociologists have often used to conceptualize social structure.11

Three mechanisms of social stability One strength of the PCT perspective is that whatever the scale of the interaction or type of social setting, the same principles of explanation apply. In particular, my theoretical exposition in this chapter relies on three explanatory mechanisms, all of which apply to every form of social structure I have just described. The first is collective control, the second is the stabilization of physical environments for interaction, and the third is the stabilization of cultural environments by the social construction of the shared higher-level perceptions that are ordinarily described as symbols and meanings. These mechanisms for creating social stability are closely related to each other. Collective control stabilizes the physical environment in which the collective control takes place. A stabilized environment facilitates the control of some perceptions while making other perceptions more difficult to control. This restriction in the range of perceptions easily controlled within a stabilized physical environments increases the probability that individuals in that environment will control similar kinds of perceptions and, because of the neural process of reorganization, will develop associative memories that make for similar kinds of meaningful connections between their perceptions. A shared cultural environment emerges from these similarities in meaningful connections between perceptions. These cultural forms, such as language and protocols, are stabilized by giant virtual controllers, which of course are a variation of collective control. Stabilized cultural forms can facilitate the coordination of people’s interactions, which helps to make their collective control activities more cooperative than conflictive. Effectively coordinated collective actions are necessary for maintaining physical environments in stable forms. In short, people stabilize both their physical and their cultural environments by means of collective control, and these stabilized physical and cultural environments allow for the meaningful social interactions by which people maintain stable social structures over time. These three explanatory mechanisms can account for stability in social structures of all types, from the smallest to the largest.

11. For more information on previous sociological conceptions of social structure, see the online version of this chapter.

272 SECTION | C Collective control and communication

Collective control and social structural levels First, collective control is obviously important to social structures at all levels. At the micro level, the organizational structures of small groups and organizations are formed from collective-control hyper-networks, and most interactions within groups and organizations get their structure from the cultural protocols collectively controlled by the participants. At the meso level of social structure, communities of various kinds provide arenas for social interactions that are ordinarily structured by protocols and thus involve collective control. At the macro level of structure, the social institutions and cultural forms that are maintained by giant virtual controllers equate to collective control on a macro scale. Hence, collective control is an essential ingredient of social structural stability at every level.

Stabilization of physical environments The relative stability of our physical living environments is also crucial for maintaining the stability of social structures at every level. We go through our lives moving and manipulating the physical objects in our immediate environments, including our physical bodies, in order to do what we want to do and perceive the things we want to perceive. From getting our bodies out of bed in the morning, to putting clothes on our backs and food in our mouths, to moving our bodies to daily sites of work or other activities, to manipulating the tools necessary for those activities, to poking and swiping at the electronic devices that provide information and entertainment (and which are omnipresent in contemporary urban environments), to making the motions of speech and body that communicate to other people, practically everything we do requires the movement or manipulation of physical objects in the service of controlling our perceptions. It is much easier to control the perceptions we want to control when the physical objects, the atenfels, most useful for the control of those perceptions are conveniently at hand in the immediate environment. The humanly constructed environments, in which people in the developed world spend most of their time, have been designed to provide a convenient selection of the physical objects for controlling the perceptions that people are expected to want to control in those environments. The environments of homes, schools, offices, factories, stores, restaurants, theaters, hotels, vehicles, and public buildings are all furnished with a variety of tools, adornments, furniture, and other physical objects intended to facilitate the control of the perceptions customarily controlled by inhabitants of those surroundings. This equipping of physical environments with atenfels designed for controlling a culturally limited assortment of perceptions does not actually compel anyone to control those particular perceptions in those locations, but the convenient availability of the physical means for controlling the customary perceptions makes it an easy thing to do, the path of least resistance, as I argued earlier.

Social structure and control Chapter | 9

273

I emphasize the physical environment here, because sociologists for the most part have conceptualized social structure entirely in terms of patterns of human behavior, rather than anything to do with the physical environment apart from humans themselves. My argument, however, is that a physical environment designed and stabilized to support a customary assortment of perceptions forms an integral component of every social structure, whether micro or macro. While human behavior is labile, often subject to change, physical objects are much less so. The physical objects that fill our living and working environments ordinarily change only when humans take action to change them. Thus, physical objects are on the whole more stable than human actions, and physical environments that have been intentionally designed to provide atenfels for the control of perceptions that are characteristic of particular social structures serve, in effect, as balance wheels to stabilize the social structures that have produced them. Of course, different physical objects are needed to support the continuity of social structures at different structural levels. At the micro level, interactions among members of many small groups take place in physical environments specially designed for these groups, such as suitably furnished buildings, homes, or clubhouses, and the physical objects that small-group members use in their interactions often include special tools, implements, or clothing. Members of formal organizations are also likely to conduct most of their interactions in purpose-designed buildingsdwhich might be offices, factories, stores, or schoolsdusing specialized tools and communications equipment to facilitate interaction among organization members. Physical objects of another type are also essential to functioning of most organizations: the files or written records of the organization’s rules and transactions. At a more inclusive level of social structure, communities based in physical locations must have buildings of various kinds, along with physical infrastructure such as roads, sewers, and communications equipment. Most cities also have sports stadia and local news outlets like newspapers and TV shows that provide community members with atenfels that support the collective control of high-level perceptions of community identity. Members of communities centered on particular activities need access to physical objects designed for carrying out the activity, such sports equipment, vehicles, characteristic clothing, or special tools. These communities of interest often feature periodic gatherings relating to the activity, such as conventions or contests, and suitable physical locations are necessary for the gatherings. When members of these communities are dispersed, some physical means of communication, such as magazines, newsletters, or websites can provide atenfels for maintaining the members’ perceptions of connection with the rest of the community. At the most macro level of social structure, the giant virtual controllers that produce cultural perceptions usually also involve physical objects such as widely distributed verbal messages, images, and videos, which can

274 SECTION | C Collective control and communication

symbolize for readers and viewers the high-level perceptions that constitute the given cultural features. Supporting the social institution of racism, for instance, are myriad depictions of people and people’s interactions in books, magazines, movies, videos, and other kinds of accounts and images, showing examples of people culturally typed as belonging to different races and of relations between people of different races. Each of these depictions is a physical object, and many of these objects are widely duplicated and dispersed, or else, like images and accounts on the Internet, capable of endless duplication and worldwide dispersion.12 Anyone wanting to eradicate the institution of racism must contend with this formidable collection of physical objects intended to symbolize the perceived differences between people typed as racially different. Many other types of cultural perceptions are buttressed by large collections of physical objects in just this way. Thus, physical objects are integral to social institutions at the macro level, just as specially designed physical objects and physical environments for interaction support social structures at the meso and micro levels of society. At every level of social structure, physical objects and stabilized physical environments contribute to the stability and continuity of the structural arrangements.

Cultural environments and social structural levels Stabilized cultural environments are also integral to the stability of social structures at every level of the social structural landscape. Most basically, for people to gain access to the social power and stability of outcome that can be achieved through collective control, they must be able to control similar, if not entirely identical, perceptions. Widespread collective control of similar perceptions can only occur in stabilized cultural environments, which thus are the key to social coordination. Take, for example, all of the specially designed buildings, tools, communication and transportation infrastructure, and countless other physical objects in the stabilized physical environments that I described above. Such standardization would not be possible without stable and widely shared cultural environments that facilitate the coordinated collective control necessary for their manufacture and distribution. Stabilized cultural 12. In his detailed historical study of the social institution of racism in the United States, Joe R. Feagin (2013) describes the “white racial frame” of white superiority, as he calls it, and argues that this perceptual framework for understanding racial differences dates back to the earliest days of the American colonies and has thoroughly imbued the ways that that most whites think about race every since. Feagin’s account makes clear that this collectively controlled system concept of racial superiority (as I would describe it) connects to a vast hierarchy of lower-level perceptions about the (imagined) superiority of whites to other races and that many of these perceptions depend on the atenfels supplied by physical objects, such as depictions of racial differences in books, magazines, videos, etc., as well as on widespread patterns of people’s customary actions.

Social structure and control Chapter | 9

275

environments also allow for the sharing, or approximate sharing, of the higherlevel perceptionsdthe meaningsdthat people attribute to other people’s actions and to their interactions with them. Social interactions that are meaningful to the participants are fundamental prerequisites for a functioning social structure, whether micro, meso, or macro in scale. Cultural perceptions depend on collectively controlled perceptions, in this case, giant virtual controllers (GVC’s). Many of the networks for cultural perceptions are global in scopedmost obviously, the networks associated with global capitalism that extend into almost every region and country. Among the GVC’s for cultural phenomena that transcend the political boundaries of countries are those associated with use of the Internet, treaties and other international agreements such as international law and human rights regimes, and styles of sport and entertainment. GVC’s for racism, sexism, and similar systems of invidious distinctions between people are also international in their ubiquity. Participants in smaller-scale social structures regularly draw on these and other globally scaled collective control processes in doing the work of their own social structures and managing interpersonal relationships within them. Somewhat smaller-scale collective control networks for cultural perceptions are confined within the boundaries of large social structures like countries, nationalities, ethnic groups, or social classes. For example, researchers using affect control theory, which is based in part on PCT, have shown how people from different countries attach distinctive patterns of emotional resonance to common words like mother, worker, or boss (see Heise, 2007). Additional examples of culturally relevant collective control networks at this less-than-global scope include languages or dialects that are unique to a particular country or region and varying styles of national dress and ethnic food preparation. Such cultural patterns are maintained by virtual controllers that might be described as meso or micro virtual controllers, rather than giant virtual controllers, depending on the scale of the social structure.

High and low culture and layers of perception Although sociologists often appear to be talking about very high-level perceptions (norms and values, for example) when they talk about culture, to understand how culture works we must keep in mind the symbolic connections between these high-level perceptions and the physical, behavioral, and relational perceptions that constitute the lower-level layers of the collectivecontrol networks making up social structures. Every set of high-level cultural perceptions thus corresponds to some set of lower-level perceptions that stabilize atenfels, the physical feedback paths necessary for constructing and controlling the higher-level cultural perceptions. The laws of a country, for instance, are perceptions at the level of principles and programs, but they are linked to concretely physical perceptions of

276 SECTION | C Collective control and communication

words written down in law books or saved digitally for display as web pages on computer screens. The bodies of knowledge comprising academic disciplines center on system concepts and principles, but they are represented concretely in physical objects like books and journal articles and also physically enacted in lessons given by teachers in school classrooms. The concrete perceptions of physical works of art, performances of music, and stories or poems written down in books or magazines are linked symbolically to abstract cultural perceptions of value and emotional meaning associated with the high arts. Anthropologists sometimes use the term material culture to describe the potsherds, arrowheads, and rubble of buildings uncovered at archaeological dig sites, and my PCT analysis applies the term culture much the same way, not only to sets of abstract perceptions but also to the concretely physical feedback paths that link to these perceptions by symbolizing them. Middle-level perceptions can also symbolize higher-level cultural perceptions, and the control of these middle-level perceptions enables people to control their higher-level perceptions successfully. A cultural system of law, for instance, implies the existence of courtroom procedures; roles for the behavior of judge, jury, prosecutor, defense lawyers, and other participants in legal proceedings; and perhaps even prescribed styles of dress for the participants, like the robes worn by judges. A cultural system of technology implies not only the existence of machines but also procedures for building and using the machines and roles for the personnel who operate and maintain them. We can describe these kinds of middle-level perceptions as cultural, in the same way that the high-level concepts they symbolize are cultural. The collective control networks for maintaining these middle-level perceptions can be just as widespread as those for higher-level cultural perceptions, and they often form important components of social institutions. In sum, collective control networks for cultural perceptions are found in every social structure, and many of these cultural networks extend beyond the boundaries of single social structures. By participating in a social structure’s collective control networks, the members of a social structure experience for themselves the control of high-level cultural perceptions associated with the social structure, as well as perceptions signifying the social institutions in which such structures are embedded. For example, participants in legal systems ultimately pursue perceptions of justice; participants in families seek family love and honor, participants in schools, growth in learning by students, and so forth. In contemporary societies, interactions between people are not always confined within a single social structure. Many events and encounters in everyday social life occur at structural boundaries and thus involve interactions between people who have no social structures in common, on the local level, at least. Commercial transactions, for instance, ordinarily take place at the boundaries of the social structures of commercial establishments

Social structure and control Chapter | 9

277

and involve an interaction between a sales person acting on behalf of the social structure and a customer who is not a member of the social structure. Events like parties, meetings, or public performances are put on by members of social structures like families, organizations, or performing companies, but the people taking part in the event may often include attendees or audience members who do not belong to the sponsoring social structure. And, as I noted above, people in urban settings often encounter people from disparate social structures in their daily round, for example while using public conveyances or other services. Even though such encounters and events do not bring together people from the same social structures, the behavior of the participants is typically well structured and often highly predictable because everyone involved also participates in the widespread collective control networks for culturally institutionalized perceptions of how people should behave in these settings. In today’s world, truly intercultural encounters are rare; two random people may not speak the same language, but it’s unlikely that they will have no other collective control networks for cultural perceptions in common. Thus, the macro-scale collective control networks for cultural perceptions serve not only to structure interactions within local social structures, but also to knit local social structures together and to provide a framework of interaction for people from different social structures who nonetheless participate in some of the same collective control networks for cultural perceptions.

The two faces of social structure Summing up this discussion of my three explanatory mechanisms, I have argued that social structure has two faces. From the bottom up, we see social stability as resulting from the work of particular social structures, which are constructed from layers of interlocking collective control networks that create and maintain the necessary lower-level feedback paths for allowing the members of a social structure to control high-level perceptions of group identity. Small-scale social structures interpenetrate in complex ways, as individuals in the course of their daily lives may participate in many different social structures. In contemporary societies, individuals typically participate in combinations of social structures that are unique to the individual, which is an important source of individuality in contemporary life. Social structures are all alike, however, in creating and maintaining physical environments conducive to the purposes of the structure. Looking at social structure from a top-down or macro perspective we see social stability as maintained not only by discrete social structures but also by the action of enduring cultural institutions that can be understood as forming a cultural environment for social structures. Like discrete social structures, these cultural institutions consist of layers of collective control networks, but the collective control networks of cultural institutions are giant virtual controllers,

278 SECTION | C Collective control and communication

more extensive and enduring than the networks of particular social structures. The collective control networks of cultural institutions transcend local social structures by bringing together the members of disparate local structures in the control of cultural perceptionsdways of seeing the world, doing things, making things, and using things that are common to the social structures of a region or country, or even globally. Thus, the control of cultural perceptions knits different social structures together in a cultural environment that fosters many kinds of similarities in the perceptions, the actions, and the physical environments of nearly everyone in a given population. Using these three explanatory principlesdcollective control, the stabilization physical environments, and the stabilization of cultural environmentsdI turn next to a more detailed examination of how social stability is achieved and maintained and also what causes social structures to change.

The dynamics of social structures Social structures are collective achievements. Every individual participant in a social structure contributes to its overall stability and longevity by joining into the structures’ collective control networks. Despite all this collective effort, however, the resulting social structures are never completely stable. At the micro level, the perceptions that people control collectively are constantly shifting, as I described above in regard to the hyper-networks of collective control networks that constitute the social structures of small groups (see Fig. 9.3e9.7), and the GVC’s of macro-level cultural institutions slowly evolve over decades and centuries. Thus, to understand the dynamics of social structures, we need to look both at stability and change. For a social structure to remain stable requires that the structure’s material and behavioral base be maintained and reproduced over time, and also that its environment of cultural institutions, which give it meaning, have stability. This section of the chapter begins with an examination of how the material, behavioral, and cultural patterns of social structures are first stabilized and then maintained. I argue that what makes structures work is in fact the work that people do, and that the activities we ordinarily regard as work contribute directly to the stability of social structures by providing atenfels for control of the higher-level perceptions central to social structures. My first topic, then, is work and its meaning in the context of the theory of social structure offered here. In the second part of my discussion of social structural dynamics, I examine processes of social change. Because social structures, as I have defined them, are composed of hyper-networks of interconnected collective control networks, to understand change at the social structural level we must examine how and why collective-control networks change. Briefly, I argue that changes in a social structure’s hyper-networks can stem either from changes in the structure and operation of the collective control networks that

Social structure and control Chapter | 9

279

make up a structure’s existing components or from the creation of new collective control networks for controlling novel perceptions. In the first case, the membership of existing collective control networks changes over time, as members of networks depart and others join in. These changes in membership can lead to shifts in the consensus reference points for the collective control processes of these networks, and the reference points of individual network members can shift, as well. Such shifts in the reference points for collective control processes may have an impact on the environmental stabilities, the atenfels, supported by these control processes. Another kind of change in a social structure’s hyper-network results from the formation of new collective control networks for previously uncontrolled perceptions. We ordinarily describe this process as innovation, and I will offer a PCT perspective on why innovations occur and spread, as well as on the kinds of people most likely to innovate. Innovations occur, I argue, because the atenfels stabilized by the existing collective control networks of a social structure are not sufficient for some members to control all of the perceptions they would like to control. When people find themselves unable to control their perceptions, frustration ensues, and the process of neural reorganization, which is always going on at some low level, accelerates with respect to the kinds of perceptions they are seeking to control. Such reorganization of perceptual hierarchies, if successful in bringing people’s perceptions back into control, provides them with new ways of perceiving and manipulating their shared environmentsdin other words, innovative ideas, behaviors, and physical objects. When others imitate such innovations, entirely new collective control networks can be formed within the complexes of networks that constitute social structures. And finally, because of the complex interconnections among the collective control networks comprised by a social structure, changes in collective control networks at any level of the perceptual hierarchy may result in compensating shifts of virtual reference points for other collective control networks all up and down the many layers of the structure’s collective control hyper-network. Changes that provoke further changes may include changes in the macro social structures in which local social structures are embedded, changes in other social structures to which the members of the social structure belong and which affect their ability to control perceptions within a structure’s hyper-network, ongoing events or happenings that alter collectively controlled perceptions which had previously been taken as settled facts, the diffusion of innovations that create new atenfels or new ways of looking at the world, and disruptions of the physical environments in which the social structure operates, as can occur in natural and humanly caused disasters. For any and all of these reasons, social structures are subject to continual change, and in the final part of this section I will take a closer look at these sources of dynamism and change.

280 SECTION | C Collective control and communication

Work and social structures Creating and maintaining stable feedback paths The kinds of activities described as work in everyday language are activities that create stable feedback paths in a shared environment for the benefit of other people. The word is also commonly used to refer to the kinds of activities that maintain these feedback paths in place. Thus, work activities produce some kind of environmental stabilization, the creation of some atenfel for use in controlling other perceptions.13 Manual workers create stable feedback paths by manipulating physical objects; they build things, make things, and clean things up. Agricultural workers produce fields of crops and confinements full of animals to be used as food. Transportation workers move truckloads of products from factories to stores, where sales workers make those products available to customers in exchanges with predictably structured protocols. Service workers manipulate and stabilize the immediate physical environments of individuals, including their dwellings and even their physical bodies, as barbers and hairdressers do. Healthcare workers attempt to stabilize the physiological functioning of people’s bodies. Educators strive to turn out classes of graduates with predictable abilities and skills, people who can then be hired to put their skills to work creating various kinds of feedback paths for others. Government workers maintain stability and order for the community in a wide variety of ways, from removing trash to providing and enforcing laws designed to regulate commercial transactions and maintain public order, and thus preventing large disturbances that would make control of other perceptions difficult. The purposes of any given social structure are reflected in the work done by its members, that is, the ways they seek to stabilize some portion of their shared environment. Thus, we can classify social structures by the kind of work their members do: for example, families, ideally at least, stabilize a home environment for family members; schools aim provide stable flows of individuals with the tools to take action in predictable ways; businesses provide people with goodsdobjects that can be used as feedback pathsdand

13. A possible exception to this generalization may occur when the purpose of the work is destructive, as with workers whose job is tearing down buildings or terrorists bent on blowing things up. Even in these cases, however, the evidence of work having been done is provided by perceptible changes in the physical environment, changes that allow for the better control of the perceptions defining the purpose of the work. As I will argue below, changes in the physical environment that promote the control of some perceptions will ordinarily make a range of other perceptions harder to control in that environment. By the same token, the destabilization of portions of the physical environment that have been stabilized to facilitate the control of some perceptions may clear the way for the control of other perceptions, as when workers tear down buildings so that others may be built. How any given change in the physical environment is to be evaluated depends entirely on the perceptions one seeks to control.

Social structure and control Chapter | 9

281

servicesdroutine actions that serve as feedback paths for controlling the perceptions of those who receive the services; and governmental structures are intended to prevent the kinds of disturbances to a shared environment that would make the work of other social structures more difficult. Even workers whose work seems somewhat abstract must produce physically perceptible stabilities, which can then be used as feedback paths for controlling lower-level perceptions essential for control of the higher-level, more abstract perceptions that provide the ostensible objectives of their work. Administrators and business executives create feedback paths by organizing the routine activities of others into predictable and efficient patterns for getting the work of an organization done. Knowledge workers put words on paper or images on electronic screens in order to send symbolic messages to others, thus facilitating their readers’ control of higher-level perceptions. Entertainers offer their performances hoping to attract audiences, who will then use the performances as feedback paths for controlling perceptions of excitement or amusement. In every case, the creation of some perceptible product in the form of stabilized portions of the physical environment or stabilized patterns of human actiondin other words, atenfelsdprovides the empirical evidence that work has been done. These types of stabilities form the material and behavioral bases of social structures, and thus by producing these physical and behavioral stabilities people contribute to the overall stability of the social structures to which they belong. In some kinds of work, people maintain feedback paths rather than creating them. People doing this work take the existence of certain feedback paths as perceptions to be controlled and then seek to protect them against the ongoing effects of disturbances. The janitor cleaning a building, the systems engineer fixing software bugs, the emergency responder driving an ambulance, or the baby’s caretaker changing a diaper, all work to maintain feedback paths for others. Thus, the feedback paths in our shared environment depend on constant human attention and effort to do the work necessary to keep them stable. Without continual work, a humanly structured environment begins to crumble over time, like ghost towns or ancient ruins. The environments that most people live in are filled with feedback paths, both physical objects and routine actions that have been shaped and maintained by human work. An important fact to remember about all of these kinds of work is that while the feedback paths stabilized by a given kind of work can facilitate the control of perceptions compatible with the purposes of a particular social structure, these stabilized feedback paths can at the same time make the control of an array of other perceptions more difficult, or perhaps even impossible, because the atenfels stabilized by the work are incompatible with the control of other perceptions in the same physical space. For example, workers in extractive industries for strip-mining coal or producing oil from shale use heavy machinery to remove whole mountaintops or large tracts of forest, along with all the plants and animals that formerly lived in those

282 SECTION | C Collective control and communication

habitats. While that work can provide the necessary feedback paths for other workers to extract and process the coal or shale oil, this removal of plants and animals and disruption of the physical landscape makes it impossible for people to control their perceptions of outdoor activitiesdlike hunting, fishing, hiking, or loggingdthat had depended on feedback paths no longer available in the disrupted landscapes. In sum, when we understand work activities as stabilizing feedback paths for controlling certain perceptions, we see that these activities can have a downside, as well as an upside, by limiting the opportunities for people to control other kinds of perceptions.

Work and resources The concept developed in this chapter for describing elements of feedback pathsdatenfelsdhas some similarities to the sociological concept of “resources” (Giddens, 1984; Sewell, 1992). These physical and behavioral components of feedback paths can be regarded as resources in the sense that, as relatively invariant phenomena in the physical environment, they make possible the control of certain perceptions. Just as timber is a resource that makes possible the construction of wooden objects, coal is a resource that can be burned for heat, and money is a resource for procuring of all manner of objects and services, the objects and patterned actions classed as atenfels are resources that make it possible for people to reach their goals by allowing the control of the perceptions defining those goals. As we have seen, an essential part of every perceptual control loop is a feedback path through the physical environment (see Fig. 9.1), and in the physical environments in which most people live, almost every feedback path necessarily passes through humanly constructed atenfels that can be described as resources because they provide the physical infrastructure for most human activities.14 Thus, resources, as I have defined the term here, are always the result of the work done by members of the social structure. In doing this work, however, the members must make use of other resources: existing atenfels in support of the perceptions involved in the work activities themselves. In other words, social structures both depend on and produce resources in their work. For example, a logging company is a social structure that exploits timber resources, which are then cut into boards of shapes and sizes that can be sold as resources for people outside this social structure, allowing them to build houses or to make furniture or other wooden objects. To do their work, the 14. Designated wilderness areas might be considered an exception to this generalization, because people hiking through them are unlikely to encounter many humanly created atenfels (beyond those packed in by the people themselves). However, in today’s world an entire wilderness area can be regarded as a complex set of atenfels for experiencing perceptions of being in the wild, since such areas must be set aside for this special use, and then the some members of some social structure or structures must take responsibility for maintaining the boundaries of the areas and policing the activities that go on in them.

Social structure and control Chapter | 9

283

loggers need to use roads into the forest, trucks, chainsaws, and other woodcutting tools, etc.dthe various physical objects that serve as resources for the purposes of logging. The uncut forest can of course also be regarded as a resource, but it takes the application of further resources, including money and the routine work activities of the loggers, in order to transform the forest resource into resources useable by others. A logging company without access to all these tools and other resources would have no purpose, nor would it have any purpose if it failed to produce further resources for people outside the social structure to use. In sum, I have argued in this section that in every social structure certain work activities are carried out, which result in the stabilization of atenfels for use by participants in the structure, as well as by others. Moreover, work activities that produce the atenfels necessary for the control of certain kinds of perceptions inevitably make other perceptions more difficult to control in that environment, even as they facilitate control of the perceptions that the structure is designed to support. I have also argued that the atenfels produced by the work of a social structure can be regarded as resources for controlling perceptions, and that social structures must have a set of resources already in place in order to produce additional resources that link to the objectives of the social structure. Finally, I have argued that these work activities contribute directly to the stability and longevity of the structure itself. I turn in the next part of this section to considering possible disruptions to the stability of social structures, the sources of social change.

Socialization of new members of social structures Redundancy of feedback paths and reorganization of perceptual hierarchies An important source of change in social structures is change in the membership of the collective control networks in the structures’ hyper-networks. Changes in membership occur when new members join in or older ones leave. The reasons for losses of membership seem straightforward enough: people may die, may exit from a collective control network to do something else, or may simply stop participating. The addition of new members to collective control networks, however, is more complex. Unless the new members have already participated in similar collective control networks in other structural settings and thus have developed the necessary control systems for participation in the new structure, they will need to learn how to control the relevant perceptions before they can participate effectively. In other words, they must develop elementary control systems that resemble those of the other people engaged in the collective control process. The socialization of newborn babies, whose brains lack the control systems necessary for participation in social structures at an adult level, clearly illustrates the steps necessary for adding new members to the collective control networks of social structures.

284 SECTION | C Collective control and communication

A baby born into the social structure of a family encounters an environment that has been stabilized in many ways. The mother’s breast or a bottle provides the infant with its first feedback path (literally!) for assuaging hunger. The repetitive actions of the infant’s caretakers, as they rock, diaper, bathe, and clothe the baby, provide stabilized feedback paths for the infant to control perceptions of comfort and warmth. As the baby gets older, caretakers continue to interact with the child in highly repetitive ways, using the same phrases over and over or reading the same storybooks again and again. The child is given physical objectsdtoysdto play with day after day. These repetitive experiences of socially stabilized atenfels, as the child explores its world, provide the environmental invariances necessary for the child to form new control system units in its brain by the process of reorganization, circuits that the child then can begin to use to control higher-level perceptions based on the atenfels and molenfels provided in the local social structure. As reorganization of the perceptual hierarchy enables infants to control perceptions at successively higher levels (Plooij, this volume), control of their lower-level perceptions becomes more and more automatic. Although their first steps are tentative and unsteady, by the time children are two or three years old, most can walk and run with practiced ease. Thus, the control of these lower-level perceptions becomes habitual, in the sense that these control circuits are well tuned to the environmental structures that the children ordinarily encounter, and control of these lower-level perceptions thus does not require any conscious attention, at least as long as the control at that level is successful. By developing habitual control of lower-level perceptions, the child is free to direct attention to the control of higherlevel perceptions that still remain challenging. As Powers argues, conscious attention is a sign that reorganization is taking place (Powers, 2005, 201e203). Thus, children develop habits of perceptual control that come to conform to the feedback paths their social structures have offered them, since the use of these paths provides the surest route to controlling higher-level perceptions that are important to them. Their verbalizations begin to approximate words repetitively used by their caretakers, and their favorite foods are selected from what they have been fed. The feedback paths stabilized by the collective control networks of the social structures into which they have been born become feedback paths of least resistance for them.15 Of course, young children seeking to control their own perceptions sometimes resist using the feedback paths provided by their social structures (and sometimes do so loudly), but their resistance brings counter resistance from their caretakers, as well as from other members of the collective control networks that prescribe

15. Taylor (2019) describes in considerable detail how the process of reorganization can work in language learning.

Social structure and control Chapter | 9

285

and maintain the social structure’s feedback paths, people who depend on the stability of these feedback paths for controlling their own valued perceptions. Thus, children caught in conflict with their elders cannot always control their perceptions to their own satisfaction, because of the pressures they feel to conform to the established ways of doing things. According to PCT, such frustration is accompanied by a faster rate of reorganization, and the reorganization rate slows down again when a child finds new ways of perceiving and acting that bring their perceptions back into control.16 In most cases, this process of perceptual reorganization tends to increase the consistency between an individual’s perceptual hierarchy and those of other people in the individual’s immediate living environment. Giant virtual controllers can produce environmental stabilities that are so long lasting and obdurate that a single individual has virtually no leverage for affecting them. Such entrenched stabilities form the perceptual worlds that individuals subjected to their influence come to regard as normal. These islands of perceptual stability include not only physical landmarks, such as buildings and roads, but also behavioral and cultural consistencies, such as familiar rituals or ways of eating and talking. Thus, people who grow up in the same environment, experiencing the same stabilized environmental features, are likely to develop perceptual hierarchies that resemble each other’s in many respects, if not in all details.

Learning by imitation and play Imitation is an important part of the process that aligns children’s perceptual hierarchies with the feedback paths provided to them by the social structures of their childhood. Imitation occurs when people see others controlling their perceptions successfully and then try to make their perceptions of their own actions match their perceptions of what others are doing. Imitation of course involves the trial and error process of reorganization, but it also involves building new control systems from combinations of controllable perceptions that are already part of person’s hierarchy. Learning another language, for instance, requires careful imitation of the speech of others, and adult learners often have difficulty learning to speak without an accent in a new language, because their perceptual hierarchies lack the lower-level control systems necessary for hearing and producing the sounds of the new language that were absent from the language or languages they learned as children. The sounds they know how to make often get in the way, because correct pronunciations of words in the new language can’t start with the wrong set of 16. The process of reorganization has an element of randomness, and children will sometimes find entirely new ways to satisfy the perceptions that they seek to control. Thus, reorganization can result in innovation, rather than in conformity to patterns being pressed upon them by their caretakers. I discuss innovation later in the chapter.

286 SECTION | C Collective control and communication

sounds. Children’s brains are more plastic, that is, more easily reorganized, and they have less difficulty building through imitation the low-level perceptual circuits needed to hear and produce the subtle distinctions of another language’s sounds. Activities that we describe as play also enter into the process by which children align their perceptions with others in their social structures and thus develop facility in using the available feedback paths to control their own perceptions. From a PCT perspective, play activities allow children to explore the limits of their control. In contrast to work activities, which have the objective of creating or maintaining stable feedback paths for others, play activities do not ordinarily have the goal of producing a reliably stabilized product for someone else. Whether someone succeeds or fails in these activities, consequently, doesn’t really matter to anyone except the individual (as well as, in a competitive game, the other people playing the game).17 Success makes a difference to the individual, because it provides feedback on the extent to which the necessary control of perceptions has been achieved. Children’s play thus opens up opportunities to practice and perfect important skills in settings relatively free of the social pressure of others depending on the success of their endeavors. Children spend hours testing their physical skills against the challenges of playground equipment, for instance, or pretending to take on adult roles. The playing of games provides opportunities for adults and children alike to test the limits of their control. In many games, chance plays an important part, as a successful outcome is sought in the face of unpredictability. In games that mix chance and skill, winning or losing often hinges on how well the competitor can play the game, that is, controls his or her perceptions within the circumscribed framework of the game environment. In gambling and other games that depend primarily on chance, however, a winning streak may give the competitor an illusory but exhilarating feeling of control, even when skill has little or nothing to do with it. Moreover, losing streaks can create such strong disturbances to illusions of control that the competitor may counter the disturbances by doubling down on the game. In games where real skill predominates, practice in playing the game at the limits of one’s control helps to optimize the control systems required for playing the game. Optimized control systems can have a higher loop gain and thus control the target perception more tightly. Computational models of control loops, however, demonstrate that a system’s loop gain can be increased only so far before control becomes unstable, showing the kinds of oscillatory patterns that 17. By this argument, professional sports, as well as team sports at the high school or college level, take on the character of work instead of play, because the athlete’s success in controlling the perceptions necessary to win the game provides feedback paths for the control of perceptions by many other people, such as parents of the athletes, fans dedicated to supporting a team, or people who have placed bets on the outcome of the contest.

Social structure and control Chapter | 9

287

Norbert Wiener recognized as signifying the presence of control systems in the human body, when he first argued that the “purpose tremors” some people exhibit are evidence of malfunction in human neural systems that are organized as control systems (Wiener, 1948, 15). However, many intensely pleasurable activities (most notably sexual intercourse, but also the buzz from intoxication and the tremors of extreme physical exertion) also have this oscillatory quality, as the physiological control systems involved in these activities operate at the limits of their capacity. Thus, visceral vibrations often accompany feelings of intense excitement and exhilaration, and this controlsystem understanding of play and fun as an exploration of the limits of control can help us understand why that occurs.

Differences among types of new members of social structures The challenges that young children face in aligning their perceptions with the feedback paths provided in their social structural environments may be substantial, but they pale in comparison to the challenges faced by immigrants from one country to another (Wexler, 2006). The task facing children is to build up a new hierarchy of perceptual control systems layer by layer by exploring the opportunities provided and limitations imposed by the feedback paths they encounter. An immigrant, by contrast, enters an unfamiliar set of social structures with a perceptual hierarchy already in place, but one that is out of alignment with the social environment of the host country from bottom to top. Reorganization must take place at all levels of the hierarchy: from the level of physical actions, such as the contortions of mouth and tongue necessary for speech in the new language and the physical gestures that ordinarily accompany conversations with others, to the abstract level of perceptions of the self. An immigrant who has made the transition successfully may feel like he or she has become an entirely different person in the process. Individuals who simply enter a new social structure in the same overall cultural environment (because the new social structure links to many of the same GVC’s for cultural perceptions as their old ones) are likely to have an easier time than either children or immigrants in aligning their perceptions with the new structure. For instance, people moving from one job to another or one school to another within the same country can use many of their control systems for culturally habitual ways of acting and thinking in the new settings, leaving less need for reorganization up and down the perceptual hierarchy. The entrance of any new member into a social structure, however, whether the person’s need for adjustment of perceptions is massive or minimal, changes the balance of the virtual reference points of the structure’s collective control processes and thus becomes a source of dynamism for the social structure as a whole. With new members replacing old ones, social structures

288 SECTION | C Collective control and communication

never stay the same for long. And even when membership is stable, social structures change through innovation, as differences between members’ perceptions lead to conflict and the invention of new ways of doing things. We turn to these topics next.

Innovation and change in social structures Mismatches between self and living environment Because each person’s perceptual hierarchy is uniquely different, the physical and cultural environments provided by social structures will provide some individuals with more resources than others have for controlling their perceptions. The physical environments of social structures comprise an array of stabilized physical feedback paths, and their cultural environments include GVC’s for the standardized manufacture and use of common physical objects, for cultural protocols and patterns of routine and ritualized actions, and for social institutions like values, norms, and common-sense knowledge. People possessing perceptual hierarchies well aligned with the physical and cultural environments in which they find themselves, and in possession of the necessary atenfels for access to the physical environments, will thus find it easy to use these resources (stabilized feedback paths) to control their most important perceptions, especially perceptions that relate to their personal identities and feelings of self worth. Other people, however, may find they lack the necessary resources for controlling key perceptions in their hierarchies. Groups of people whose perceptual hierarchies are well matched to the resources available in the physical and cultural environments of their social structures have some clear advantages. They possess social power, in the sense that they can use the resources available in their social structures reach their own goals more easily than other people can. Conversely, people whose personally important goals don’t match the means available to them for meeting those goals are obviously disadvantaged. High among the goals pursued by many advantaged groups is a firm resolve to preserve their advantages by using the all the available resources to stabilize their social structures in their current form. Thus, social structures have histories, and their socially most powerful members construct and maintain them over time to suit their own purposes. One commonly used tactic for preserving a group’s advantages is to restrict the access of other groups to the resources necessary for controlling their own valued perceptions. Here, of course, we are talking about discrimination and social inequality, central concerns of the discipline of sociology. People whose perceptual hierarchies are well matched to the stabilized feedback paths in their living environments have little need for innovation. Disadvantaged groups, however, may find themselves in circumstances in

Social structure and control Chapter | 9

289

which innovation becomes a necessary coping mechanism. Groups in contemporary societies who face restricted access to resources for controlling their own valued perceptions include women, people with little money and in lower social classes, ethnic and racial minorities, many young people, immigrants and refugees, people whose sexual practices or gender identities differ from the presumed heterosexual norm, people whose physical or mental capabilities are substantially compromised, and people labeled as criminals or felons.18 The types and degrees of restrictions encountered by these various groups vary widely, of course, but members of each of these groups have less than free access to the use of the stabilized feedback paths of contemporary societies, in part because people with privileged access to those feedback paths seek to monopolize it. Powers has explained the likely outcomes when an individual cannot control his or her perceptions. First, prolonged inability to control produces negative emotions, such as irritation, frustration, anger, or even depression (Powers, 2005, 257). Second, lack of control increases the rate of reorganization of the interconnections between elementary control units in the brain (p. 255), especially affecting those implicated in the lack of control. Reorganization, however, is a somewhat random process, and the results of this reorganization are not necessarily predictable. One possible outcome of a mismatch between people’s control hierarchies and the collectively stabilized feedback paths is that they learn how to function more effectively within their current environment by reorganizing their control hierarchies in the direction of greater similarity to those of more advantaged community members. When immigrants assimilate to the culture of a new country, for instance, this kind of learning has taken place. Another possibility that I have discussed is innovation, whereby people develop unconventional control systems that allow them to control their valued perceptions by using the feedback paths available to them in creative ways. An immigrant community might innovate, for example, by developing a hybrid cuisine that resembles the favorite dishes of the home country but relies on locally available ingredients. A third possible outcome is failure of the reorganization process, leaving the individual still unable to control important perceptions and thus caught in an emotional state of chronic anger or depression. Of course, perceptual hierarchies have many layers, and people whose control hierarchies are not well suited to the available resources may end up assimilating at some perceptual levels, innovating at other levels, and suffering chronic frustration at still others. Because innovation is one likely outcome of reorganization to correct a mismatch between a group’s valued perceptions and 18. Jails and prisons are physical environments specifically designed to minimize the feedback paths available to inmates. Many ex-prisoners, as well, may face restrictions in the availability of feedback paths, such as reduced access to jobs or lack of voting rights (Alexander, 2012, 140e177).

290 SECTION | C Collective control and communication

the available feedback paths, disenfranchised groups like minority youth or immigrants have often been the source of new fads and fashions in dress or popular music, as they express their rebellion by putting to innovative uses whatever feedback paths their social structures make available. In many cases, though, people in discriminated-against groups may find it expedient to assimilate as best as they can to the culture of dominant groups. People in these groups also tend to suffer internal conflict at relatively high rates, sometimes escaping into substance abuse or other forms of “deviance” as a way to cope.

Competition, innovation, and social structures In the competitive milieu of global capitalism, even highly privileged individuals may come to feel that they lack sufficient resources to control the perceptions they want to control. Perceiving their own wealth or economic success as insufficient, they may innovate to get yet more, like the Wall Street bankers whose financial machinations led to the “great recession” of 2008. And because innovations sell new products, many of jobs in capitalist economies require workers to turn out a steady stream of inventive objects, actions, or ideas. Industrial, fashion, and software designers; artists, authors, and composers; media commentators and academic researchers; in these jobs and many others, workers’ paychecks depend on their ability to innovate. The organizations and other social structures that provide jobs for creative workers are often intensely competitive. Of all the people vying for success in these cultural competitions, only a few can have the best sellers, the most newsworthy performances, the most cited articles, the scientific breakthroughs, or even videos that go viral on the Internet. The other competitors, and sometimes the winners as well, may not get enough attention, adulation, and commercial success for them to control the highly valued perception of being the best in their own business. While innovation is necessary to stay competitive, workers in these creative pursuits usually enjoy far more resources to support their endeavors than innovators in less privileged groups. And their perceptual hierarchies are likely to conform in most respects to the prevailing cultural pattern of stabilized feedback paths, which allows them to focus on solving creative problems rather than coping with everyday problems of survival, although it often limits their innovations to narrowly defined commercial channels.

Consequential and inconsequential innovations Social structures are formidably complex, as I have argued, built from layer upon layer of interconnected perceptions maintained by hyper-networks of collective control. Innovations may occur in any of the many perceptual levels of these hyper-networks, as individuals reorganize their perceptions to

Social structure and control Chapter | 9

291

devise new elementary control unitsdinnovative ways to perceive and manipulate their environmentsdand others then observe and imitate the results of these innovations, thus adding new collective control networks to the hyper-networks of social structures. When these new collective control networks merge into the complex overall structures, however, it may lead to adjustments all up and down the interconnected hierarchies of hypernetworks, in somewhat the same way that the control of strong disturbances to a perception at one level of an individual’s perceptual hierarchy may require many readjustments in the control of connected perceptions at other levels. Thus, the changes introduced when an innovation is widely adopted in a social structure may reverberate throughout the rest of the structure’s hyper-network. However, some kinds of innovations are more consequential than others. Most of the innovations routinely produced by people with creative jobs have only a negligible impact on the social structures they serve. Journalists’ daily stories about current events, the popular-culture offerings of the entertainment industries, the constant outpouring of tweets and other trivia on the Internet, the ever-changing high-fashion designs in clothing, the parade of new-and-improved products filling the shelves of big-box stores, and even the earnestly turgid publications of most academics can be easily accommodated within the hyper-networks of contemporary social structures. The innovative perceptions offered by creative workers for public imitation and consumption tend to located in lower levels of the perceptual hierarchy, and these low-level perceptual signals can fluctuate quite a bit without appreciably disturbing higher-level perceptions maintained by the giant virtual controllers (GVC’s) of the social institutions central to the overall social structures. Putting it another way, these mass-produced innovations fit neatly into the competitive logic of the capitalist economic system that pays the salaries of the creative workers, and, in fact, the operation of this system would grind to a halt without a constant flow of such low-level innovations. The GVC’s for high-level perceptions of the core social institutions of social structures operate on a much longer time scale than the collective control systems for the everyday innovations turned out by creative workers. These low-level innovations usually make use of short-lived and easily malleable atenfels, such as words and images flitting across computer screens or physical objects expressly designed to be consumed, thrown away, and quickly replaced. Furthermore, creative workers ordinarily design these innovations, perhaps unconsciously, to be consistent with the longer lasting high-level perceptions that define the cultural environments in which they are created, as when the plots of popular novels or movies recycle well-worn narrative tropes and stock characters. The trick for the creative workers devising such innovative products is to concoct combinations of perceptions uncommon enough to garner wide attention but not so extreme as to offer any great disturbances to assumptions and sentiments controlled by the GVC’s for

292 SECTION | C Collective control and communication

social institutions.19 And because these high-level perceptions are built upon many layers of lower-level perceptions, changes in their component perceptions several layers down are unlikely to be reflected in any substantial change at the higher level. Thus, a constant ferment of innovations at lower perceptual levels is easily compatible with the relative stability of the overall social structure. Sometimes, however, the invention of a new product or idea does leads to far-reaching changes in other collective control networks within the hypernetworks of social structures, if not necessarily at the highest structural levels. Recent innovations in computer technology, for instance, and especially the invention of hand-held devices, have offered new options for controlling perceptions that formerly were more difficult or even impossible to control, and these new atenfels have enabled people to modify their routines of work and entertainment in ways that would have been almost unimaginable a few decades earlier. At a higher perceptual level, some social movementsdfor example, the Civil Rights Movementdhave succeeded in disseminating new perceptions of the social world that have changed people’s high-level perceptions of people in other groups.20 And leaders of political revolutions, by their creative re-envisioning of society, have sometimes been able to institute massive and thoroughgoing changes in the social structures of their countries. Nevertheless, as was seen with the re-emergence of Orthodox Christianity in Russia and the partial return to Confucian values in China, the GVC’s for a society’s high-level perceptions of norms and values are extremely resistant to change, even in the face of political revolutions.

Other sources of social change Innovation and turnovers in membership are not the only sources of change in social structures. Conflict, as sociologists have long argued, can lead to social change, and changes in other structures to which a social structure is connected can also be the occasion for social change. With regard to conflict as a source of change, the PCT perspective does not necessarily confirm the idea that conflict and change go hand in hand. My computational modeling simulations have shown that conflicts in collective control can result in stasis rather than 19. According to affect control theorists and researchers, who use a control-theory framework to study the process by which people construct ongoing events, people ordinarily strive to structure the events in which they participate to be consistent with widely shared cultural sentiments about different kinds of people, actions, and settings, or, in other words, what I would describe as GVC’s for social institutions. (see Heise, 2007). 20. However, some authors argue that in spite of all the achievements of the Civil Rights Movement in the United States, the core social institutions of race that date back to the founding of the country, such as the “white racial frame” of white superiority (Feagin, 2013) and “racial caste system” of punitive control of African Americans (Alexander, 2012), are still in place.

Social structure and control Chapter | 9

293

change by creating a kind of high-tension stability with regard to the stakes of the conflict, at least until one side or the other runs out of resources to apply to the conflict. Examples abound of high-conflict situations in which little change takes place, as when deadlocks between evenly matched political parties bring government business to a standstill or entrenched conflicts between ethnic groups drag on for years (McClelland, 2004, 2014). Nevertheless, despite the fact that conflict sometimes results in stalemate, I would argue that conflict is more likely to lead to change than stability in the long run. In violent conflicts, for instance, the whole point of the violence is to damage or destroy atenfels that the victims of violence use for controlling their perceptions. Whether the aggressor bombs an enemy’s infrastructure, shoots down planes and blows up tanks or other weapons, destroys homes and buildings, places landmines and improvised explosive devices in roads, or inflicts injury and death on people’s physical bodies, the purpose of this destruction of physical objects is to make it impossible for the enemy to continue controlling the perceptions that depended on those resources. Besides attempting to retaliate, however, the enemy is also likely to seek innovative ways to control perceptions that have been rendered inoperative. When an individual’s control of vital perceptions is blocked, as I argued above, neural reorganization results, and innovation is one likely outcome. As combatants seek new ways to defend themselves and rebuild in the face of physical destruction, wartimes often see rapid technological and social change. Stalemated conflicts can also precipitate social change indirectly. In the first place, a stalemate reduces the quality of control of the contested variable because it inhibits the flexibility of the control systems involved in the conflict to respond effectively to disturbances. A stalemated political system, for instance, can have difficulty dealing with unanticipated threats, as we have seen in the failure of the United States government to move rapidly against the threat of climate change. Deadlocked conflicts also leave the individuals involved in the conflict unable to control perceptions that they feel are important, which again increases the likelihood of innovations and thus social change. Beyond the dynamism resulting from conflicts and innovation, social structures are sometimes forced into change by changes in their external environments, both physical and cultural. Most social structures are embedded in other more inclusive social structures, and shifts in the control networks of the more inclusive structures can work their way on down. Similarly, because people usually participate in a variety of unrelated social structures, major shifts in the work done in one social structure may affect the references people carry into their participation in other structures or their ability to carry out the work of other social structures, even when those structures are not nested in the first. Collective control networks, of course, function to create stability in the variables they control, so that social structures can to a certain extent absorb

294 SECTION | C Collective control and communication

changes coming from their environments without experiencing any real change internally. For instance, the policy edicts announced by a government or by an organizational headquarters will often fail to make waves at the local level, as members of smaller units of the organization reinterpret the new policy to be consistent with doing things the way they’ve been done, even while ostensibly implementing the changes. Some external changes are so drastic, however, that business as usual at the local level will no longer suffice, and innovation or at least some substantial shifting of the parameters of control becomes absolutely necessary. Like war, natural disasters can undermine the physical environments on which social structures depend and thus are a source of unavoidable change. For business organizations, market competition from other businesses may have that same disastrous quality. Because social structures and their members have to function in environments that are constantly affected by other dynamic social structures, even as they go about attempting to create their own internal stability, their collective control networks must always be engaged in readjustment, sometimes minimally, sometimes radically. In sum, social change is a fact of social life, because the collective control networks that constitute social structures are inherently dynamic.

Discussion The introduction to this chapter began with my critique of the traditional causal-modeling approach in sociological research and theory, and I listed several important questions about the social world that this conventional approach has left unanswered: • What is the origin of the empirical regularities sociologists have observed, and how do they develop? • What maintains these social patterns, and what are the forces of social change? • And when deviance from these norms is observed, how do we account for it? • More profoundly, exactly what are these social structures that sociologists have endeavored to describe? In this chapter I have offered PCT-informed answers to all of these questions. The empirical regularities that sociologists observe, I have argued, arise from the collective control of perceptions, and especially from the giant virtual controllers (GVC’s) generated in the interactions of large numbers of people using shared atenfels to control similar perceptions. These social patterns are maintained by the work that people do to create and maintain the atenfels necessary for other people to control perceptions in accordance with the norms of the GVC’s in their cultural milieu. Deviance arises because people have different and unique perceptual hierarchies that, given the uniformity of the

Social structure and control Chapter | 9

295

physical and cutural environments produced by collective control, will leave some people better able than others to control their most important perceptions. People unable to control their perceptions satisfactorily will tend to reorganize their perceptions and, thus, to innovate. Social change comes from these innovations and from changes in the external physical and social environments in which social structures are embedded. And, to answer the more profound question of exactly what social structures are, my argument in this chapter has been that social structures comprise hyper-networks of the myriad layers of collective control networks that people use to stabilize the physical and cultural environments in which they live, work, and play. The vision of social structure offered in this chapter is, I admit, radically different from received sociological wisdom. My unorthodox perspective has opened the door to a more scientific approach to sociology, and social sciences more generally, than current approaches in four important respects. First, this PCT approach to theory and research pays closer attention to the empirical world, including the world of physical objects and their meanings, than has been characteristic of most previous theory and research in the social sciences. Second, the PCT approach eschews the metaphorical language, with concepts like “social forces,” to which previous generations of sociologists have often resorted in the face of the boggling complexity of the social worlds we encounter. Instead, the PCT approach substitutes the technical vocabulary of control theory, which, though perhaps initially off-putting, makes possible a scientifically verifiable account of how social interactions work, not just how they look to outside observers. Third, the PCT approach widens the focus of sociological theorizing and research to include all of the phenomena of social life, from the materiality of our physical surroundings to the most ineffably abstract perceptions of our shared culture. Instead of focusing on a few easily measurable indicators, this approach offers a perspective on the social world that is as fully comprehensive and highly articulated as the phenomena we seek to study. Finally, the PCT approach offered in this chapter starts from the insistence that human behavior is autonomous, not caused by outside forces. Many contemporary sociologists will affirm that people have “agency” to act autonomously, while in almost the same breath talking about the external “causes” of social behavior, without ever having squarely confronted the glaring contradiction between those two ideas. The PCT perspective offers a way to reconcile individual autonomy with a rigorously causal approach to behavior by locating those causes in the references people hold internally, not the dictates of the external environment. Granted that my argument shows how the environmental uniformity and stability resulting from collective control can make conformity to the norm the path of least resistance for individuals, it nevertheless shows how individuals retain the autonomy to look for other paths.

296 SECTION | C Collective control and communication

Any sociologists or other social scientists who might seek to implement this PCT approach face a formidably challenging task, and I do not kid myself that the disciplines are likely to shift in this direction anytime soon. For one thing, this new perspective calls for a complete rethinking of our usual methods for doing social research. For a second thing, as my argument has suggested, any such drastic change has to confront the massive physical infrastructure of written texts and research procedures devised with the assumptions of old methods and theories in mind. Third, and most consequential, social scientists encountering so radical a revision of accepted ways of thinking are likely to resist it vigorously, because it disturbs perceptions they have long labored to control. I hope, however, that some readers of this chapter will catch a glimpse of a more scientific, more complete, and more satisfying way to understand our social lives together.

Acknowledgments My thanks to Peter Burke, Audrey Devine-Eller, Karla Erickson, Ross Haenfler, Patrick Inglis, Warren Mansell, Bruce Nevin, Kathleen Oberlin, Sergio Pellis, and Martin Taylor for their comments and suggestions on earlier drafts of this chapter.

References Alexander, M. (2012). The New Jim Crow: Mass Incarceration in the Age of Colorblindness (Revised edition). New York: The New Press. Feagin, J. R. (2013). The White Racial Frame: Centuries of Racial Framing and Counter-framing (second ed.). New York: Routledge. Fine, G. A. (2012). Tiny Publics: A Theory of Group Action and Culture. New York: Russell Sage. Flannery, K., & Marcus, J. (2012). The Creation of Inequality: How Our Prehistoric Ancestors Set the Stage for Monarchy, Slavery, and Empire. Cambridge, MA: Harvard University Press. Giddens, A. (1984). The Constitution of Society: Outline of the Theory of Structuration. Berkeley, CA: University of California Press. Heise, D. R. (2007). Expressive Order: Confirming Sentiments in Social Actions. New York: Springer. Marken, R. S. (2002). Looking at behavior through control theory glasses. Review of General Psychology, 6, 260e270. https://doi.org.10.1037//1089-2680.6.3.260. McClelland, K. (1994). Perceptual control and social power. Sociological Perspectives, 37(4), 461e496. McClelland, K. (1996). The collective control of perceptions: sociology from the ground up. In Presented at the Annual Meeting of the Control Systems Group, Flagstaff, AZ. McClelland, K. (2004). Collective control of perception: constructing order from conflict. International Journal of Human-Computer Studies, 60, 65e99. McClelland, K. A. (2006). Understanding collective control processes. In K. A. McClelland, & T. J. Fararo (Eds.), Purpose, Meaning, and Action: Control Systems Theories in Sociology (pp. 31e56). New York: Palgrave Macmillan.

Social structure and control Chapter | 9

297

McClelland, K. (2014). Cycles of conflict: a computational modeling alternative to collins’s theory of conflict escalation. Sociological Theory, 32, 100e127. https://doi.org/10.1177/ 0735275114536387. Powers, W. T. (1973). Behavior: The Control of Perception. New York: Aldine de Gruyter. Powers, W. T. (2003). PCT and engineering control theory. In Presented at the Annual Meeting of the Control Systems Group, Marymount College, Los Angeles, CA. Available as multiple_control.pdf from www.livingcontrolsystems.com. Powers, W. T. (2005). Behavior: The Control of Perception (second ed.). New Canaan, CT: Benchmark Publications. Powers, W. T. (2011). A hierarchy of control. In R. J. Robertson, & W. T. Powers (Eds.), Introduction to Modern Psychology: The Control-Theory View (pp. 59e82). Hayward, CA: Living Control Systems Press. Sewell, W. H., Jr. (1992). A theory of structure: duality, agency, and transformation. American Journal of Sociology, 98, 1e29. Taylor, M. (2019). The Powers of Perceptual Control: An Inquiry into Language, Culture, Power, and Politics (Unpublished manuscript.). Wexler, B. E. (2006). Brain and Culture: Neurobiology, Ideology, and Social Change. Cambridge, MA: MIT Press. Wiener, N. (1948). Cybernetics or Control and Communication in the Animal and the Machine. New York: Wiley.

Chapter 10

Perceptual control in cooperative interaction M. Martin Taylor Martin Taylor Consulting, Toronto, ON, Canada

Introduction A colleague told me the following story1: We had a large Japanese customer for our [product]. I developed a pleasant conversational relationship with one of their visiting managers. One day as I was walking fast to get to a meeting, at a corridor corner we met unexpectedly face to face. I saw his alarm, bowed in proper Japanese fashion, and his face immediately relaxed with a friendly look of perhaps gratitude in his eyes. Diplomatic protocol is like that. We are both on familiar ground. It gives us time to accommodate the unknown from a secure footing.

This interaction is very simple on the surface, a few seconds of an unexpected encounter, but underneath there is a whole story of interacting cultures that can be explained by the control of perceptions. A brief word about “control” as it is used in engineering and in Perceptual Control Theory (PCT). In everyday speech, you “control” something if you are able to keep it near some state you desire despite opposing influences. In engineering and PCT, that ability would be called “good” or “excellent” control, because in engineering and PCT “control” simply means that you act so as to move the sensed state of the thing (your perception of it) toward a desired state, not necessarily that you are fully successful, or even that your influence on the thing is significant. Influence is not control, but you must be able to influence something if you are successfully to control your perception of it. Here is a trivially simple interaction. Irene: Carl, could you open the window, please. Carl opens the window. 1. B. Nevin, personal communication 2015.11.25. The Interdisciplinary Handbook of Perceptual Control Theory: Living Control Systems IV. https://doi.org/10.1016/B978-0-12-818948-1.00010-1 Copyright © 2020 Published by Elsevier Inc. All rights reserved.

299

300 SECTION | C Collective control and communication

In discussions of Perceptual Control Theory (PCT) it is sometimes assumed that “You can’t control another person,” because both you and the other person control only internal perceptions, and if you try to control what the other person does, the only way to do it is to overpower their muscular output. But did not Irene control Carl in that little scene, and did not my colleague control the Japanese visitor? Yes, Irene did, and so did my colleague. And while they were doing so, Carl controlled Irene, and the Japanese visitor controlled my colleague. None of them controlled the whole other person, but they did control their perceptions of some aspect of the other’s observable behavior. Irene controlled for perceiving Carl to understand that she wanted him to be opening the window, and my colleague, assuming that the Japanese visitor wanted to feel safe and didn’t, controlled for perceiving the Japanese visitor to show that he now felt safe. Carl controlled his perception of Irene’s feelings, perhaps her feeling toward him, perhaps just her feeling that the window should be opened. And the Japanese visitor controlled my colleague’s feeling of discomfort at the Japanese person’s discomfort. So yes, you can control, if not “another person”, at least your perception of some of their actions or displayed feelings. But you can do it only if they allow you to do it (or if they are unaware that you are doing it). Carl need not have opened the window; the Japanese might not have been feeling unsafe and might have objected to my non-Japanese colleague’s use of Japanese forms. In other words, you cannot control another person arbitrarily without the use of force, but you can do it selectively in suitable circumstances. Why would Carl want to open the window after Irene asks, although he apparently did not want to do it before she asked? If he had wanted it opened before she asked, he presumably would have opened it, unless he thought doing so would bother Irene (meaning it might disturb some perception Irene was controlling). What perception or perceptions does he control by acceding to her request? The answer to this question is rather more complex than would appear on the surface. This chapter is about how perceptual control works in such a cooperative interaction, which we call a “protocol”. In a protocol, the actual behaviors of the participants are means to their own personal ends, and can be greatly varied without changing the protocol. For example, Irene might have displayed her desire to have the window closed in many different ways. She could have gestured, fanned herself in an exaggerated fashion, or said simply “Window”, among many other ways she could have let Carl know what she wanted. Carl might have wanted the window to be closed, and tried to find out why Irene wanted it opened, to see whether there was some other way to deal with the conflict. The possibilities are endless, but we will try to make some sense of them.

Perceptual control in cooperative interaction Chapter | 10

301

History: Layered Protocol Theory (LPT) About a quarter-century after Powers et al. (1957) had published their seminal “Prospectus”, which introduced Perceptual Control Theory to the world, my colleagues and I were independently beginning to develop a “Theory of Layered Protocols” (LPT) (some of which will be described later in this chapter). The initial development of LPT occurred almost a decade before I discovered PCT (Taylor, 1986, 1988; Taylor et al., 1984) but when I did, it was quickly obvious that LPT was simply an application of PCT to the interaction between a person and a machine or between two persons, rather than a standalone theory applicable only to “quasi-intelligent” interactions. LPT was initially developed to facilitate the design of a multimodal interface for a relational database with a geographic situation display, the modes being voice, keyboard, and gesture (Taylor et al., 1984; McCann et al., 1988). It quickly developed into a general theory of human interaction with computers or other humans, and was soon used for helicopter interface design (Farrell et al., 1999). When we start explaining and analyzing protocols proper in Part 4 of this chapter, we will use a descriptive device from LPT called a “General Protocol Grammar” or “GPG”. The GPG was initially intended as an aid for developing quasi-intelligent computer interfaces, but evolved into a descriptive tool for natural dialogue, just as the grammar of a language is a descriptive tool for the language. In previous publications, LPT has been taken as given, and explained as a consequence of perceptual control (e.g., Taylor, 1986, 1988, 1989; Taylor et al., 1999; Taylor and Waugh, 2000). This chapter takes the opposite approach, showing how the basic unit of LPT, the “protocol”, is a natural consequence of the control of personal perception. We treat a perceptual control loop not as a unique kind of feedback structure, but as a member of a class of homeostatic negative feedback loops that have an indefinite number of active elements (the standard control loop of the kind normally discussed in Perceptual Control Theory has only two). We do this because we treat a protocol as an instance of a four-element loop. The environment of each participant is no longer a passive entity to be perceived and influenced to bring the perception to a reference value, but is instead an active perceptual controller. In the performance of a protocol, many control loops are in play. Most of them control perceptions of belief and uncertainty in the interacting partners. Neither uncertainty nor belief appears among the levels of the Powers control hierarchy (Powers, 1971, 1973/2005, 1989, 1992, 2008), so control of those perceptions must be considered as a slight extension of the theory, which we justify below. Nothing in this chapter, however, contradicts anything in the

302 SECTION | C Collective control and communication

standard approach to perceptual control. So let us begin a short introduction to PCT. Other chapters in this book offer more extensive introductions, as do many writings by Powers in the books just referenced and in many formal and informal reports and presentations (accessible through http://pctweb.org).

Control and perceptual control When some people hear the word “control”, they take it to be the opposite of “freedom”, and therefore assume that a psychological theory based around “control” is a theory to be opposed on principle. Others may think of “selfcontrol”, a modicum of which is the core of civilized behavior. Too much self-control prevents people from seeing your true feelings and renders you untrustworthy, too little, and you thoughtlessly do things that damage other people and your physical environment. Perceptual Control is neither of those. To get a feel for what Perceptual Control is, imagine that you are trying to hold a tall post upright in hard ground into which the base of the pole can not sink. If you do nothing and the post is not perfectly upright, it will lean and fall down if you leave it. This is not what you want, which is to perceive the pole perfectly upright, so you act to oppose any tendency of the pole to fall in one direction or another. You might be tempted to say that you are controlling the pole’s position, but you are not. You are actually controlling a perception of its position. Perceptual Control Theory says that all your intentional actions are to control your own perceptions and only to control your own perceptions. In controlling your perceptions, you also stabilize some complex property of your environment, for example the apparent orientation of the pole with respect to gravity despite any possible effects of wind gusts. In casual conversation the environmental stabilization entailed in controlling your perception is often called “controlling” that property of the environment. The environment determines whether you live or die; if you are in a desert with a full water-bottle, you may live, but if you rely on drinking the water in a lake you see in the distance, you will die if it turns out to be a mirage. Environmental stabilization is the reason perceptual control matters. Perceptual Control Theory (PCT) takes as a primary assumption that all you can know of the world is either embedded in your genes or has been gained through your senses. The world “out there”, assuming one even exists, might be very different from the world you perceive and might work very differently from the way you perceive it to work. All your perceptions might be created by God-like beings who change them according to the ways they want the world to be when you produce some action. But, and this is a big “But”: If your actions influence your perceptions in a reasonably consistent fashion, then it matters little what is “really” out there, because you can affect the way it seems to be, just as though your perceptions truly reflected reality.

Perceptual control in cooperative interaction Chapter | 10

303

The rock (that seems to be) coming straight at your head might not exist the way you perceive it, but if you don’t move out of (what seems to be) its way, you may thereafter be unable to control any perceptions at all. If, when you set a pole vertically in soft ground, you use a spirit level rather than trusting your eyes, you won’t see it fall so soon after you leave it. Perceptual Control Theory is not about control by perception; it is about control of one’s own perception.

Elements of control The concept of negative feedback control has been known to engineers for many decades, and was applied to the analysis of simple tracking tasks performed by humans at least as early as Gibbs (1954). However, apparently nobody before Powers (e.g., Powers et al., 1957) saw that the concept had general application to psychology. The idea of perceptual control is almost trivially simple, but it often takes a stroke of genius to see the complexities that can be accounted for by the combined action of many simple things working together. Such was the case with the Powers insight that perceptual control could account for a wide range of psychological phenomena. Negative feedback in Perceptual Control Theory is exactly the same as negative feedback in Engineering. To control successfully, in both the engineering and the psychological sense, is to act so that some (controlled) variable (in PCT called a “perception” or “perceptual variable”) is maintained near some possibly varying “reference” value, despite other influences that might disturb it. The difference between the perception and the reference value is called the “error”, which is input to some function called the “output function” whose output is its “action”. The action influences the environment of the control system in some way that affects the perception in a direction that reduces the error, completing a negative feedback loop. If the action increased the error, the loop would be a positive feedback loop and there would be no control. The set of three functions (perceiving, comparing to a reference, and generating output) that together execute the control form an “elementary control unit” (ECU), the part of the diagrams of Fig. 10.1 above the gray area. It takes time for the initial effect of a change in the error value (the difference between the reference and perceptual values) to make its way around the loop until it can return to influence the error value. During this transport lag time, values continue to change all around the loop. This simultaneity of variation all around the loop is easily forgotten, but should be remembered later in this chapter when we discuss the GPG, whose states represent the continuously changing perceptions of various beliefs during interpersonal interactions. A typical short form of describing a perception being controlled together with its reference value is to say that the controller “controls for X”. For example, I may “control for the room temperature to be 18
C”, which means

304 SECTION | C Collective control and communication

FIG. 10.1 An Elementary Control Unit, or ECU, consists of three functions, a Perceptual Function that defines the controlled variable, called a “perception”, a comparator that compares the perception with a reference value for that perception and outputs an “error value”, and an output function that produces influences on the environment of the ECU which complete the feedback loop by altering the inputs to the Perceptual Function. Independent environmental influences also contribute to the Perceptual Function inputs. In Panel (A) the paths and influences in the environment are shown separately, in Panel (B) as single values that represent their combined effects.

that I control my perception of room temperature with a reference value that the temperature should be 18
C. The “control for” way of saying it is a little less cumbersome, though it might mislead some into thinking that the controlled variable is in the environment, which it is not. The variable that must be stabilized is indeed in the environment, but that environmental variable is not accessible to the ECU. Only the perceptual signal derived from it is accessible internally, and only the perception can be controlled.

The Powers hierarchy of control At the level of basic chemistry, an atom is a simple thing, a positively charged nucleus with a cloud of negatively charged electrons around it. Put together, however, a few simple atoms can create an indefinite variety of molecules with an enormous range of properties and behaviors. So it is with the simple ECU. By itself, one ECU does not do much, but a hierarchic structure of them working together can create coordinated behaviors of indefinite complexity and time-scale, as Powers demonstrated over a span of over half a century of work (see the many writings collected in Powers, 1973/2005, 1989, 1992, and 2008). Powers’s Hierarchic Perceptual Control Theory (HPCT) asserts that within the body of any organism, be it a bacterium, a tree, a jellyfish, or a human, there is a hierarchy of interconnected ECUs that together form a complex control system that is responsible for everything that the organism is does intentionally. Fig. 10.2 shows a tiny part of such a hierarchy around the ECU shown in Fig. 10.1. When we come to protocols, we will be dealing with ECUs that work together by influencing each other’s different controlled variable, rather than

Perceptual control in cooperative interaction Chapter | 10

305

FIG. 10.2 The place of the basic ECU in a hierarchic control system of the type described by Powers. Each ECU is in the Environment of those above it. Each ECU sends its output to the reference input of some ECUs in its environment at the level immediately below, and its perceptual value to the inputs of some perceptual functions in the level immediately above it. Powers eventually described eleven levels, each controlling a different kind of perception, and allowed that there could be sub-levels within at least some of the eleven.

by setting references for and accepting perceptions from each other. It must be remembered, however, that these influences always are the result of the actions of a hierarchy of perceptual controllers, not direct “wired” connections. They are accordingly more variable. Whereas in a “wired” hierarchy, the value at the output of a “wire” is the value at the input, in protocol interactions what one party perceives about the other is not always accurate and can never be known by the participants to be accurate. Fig. 10.3 shows two of many levels of perceptual control involved in visiting someone’s house. The visitor wants the door to be opened, so rings the doorbell. The figure shows one level at which the visitor want to hear the bell ringing, which would presumably happen if the bell was working and the visitor pushed the button, and another level at which the visitor wants the button to be pushed, which is accomplished by muscular actions that move the finger to the button and press it. These are only two of many levels, and the figure omits the possibilities of disturbances. The button-pushing control loop is in the environmental feedback path of the bell-ringing ECU, which is itself in the environmental feedback path of a higher-level “want to perceive the door open” ECU that is not shown. As explained below, each loop is an atenfel2 for a higher-level ECU.

2. Atenfel is a short form of ATomic ENvironmental Feedback ELement. The name, together with related ones (molenfel, atenex and molenex for molecular and nexus forms), was invented in discussions with Kent McClelland during the writing of our chapters for this book.

306 SECTION | C Collective control and communication

FIG. 10.3 Two of the levels of perceptual control that might be involved in visiting someone’s house. The visitor expects that ringing the doorbell will result in the door being opened, so wants to hear the bell ringing. It is not, so the visitor acts in order to hear the ring, in this case by setting a reference value of perceiving the button to be being pushed. This is done by moving the finger to the button and pushing, which requires several even lower levels of muscular control subsumed here as “Muscular Actions”.

Things in the environment of one control unit are likely also to be in the environment of other ECUs, so it is unlikely that an action output from an ECU will influence only its own perception. Indeed, the objective of perceiving the bell to be ringing is likely to be that someone else will also be able to hear it, will be controlling a perception that is disturbed by hearing it, and will control their disturbed perception by coming to the door and opening it. When we theorize, we often assume that the whole control loop and all the changing values of its variables are open to view. This position is often called “the Analyst’s viewpoint”. Fig. 10.3 exemplifies this viewpoint, since in practice a watching stranger could not know whether the person seen pushing the button wanted to hear the bell ring, wanted to discover the function of the button, or possibly had a numb finger that was being tested for the return of some feeling. In contrast to the Analyst’s omniscient viewpoint, two other viewpoints are possible, one from inside the controller, the other from an observer outside in the environment. We call the former the “Controller’s viewpoint”, while the latter has two forms, a passive “Observer viewpoint” and an active “Experimenter viewpoint” in which the observer acts to disturb the action of the control loop being observed. Much of late 19th century psychology was done by introspection d taking the “Controller’s View” d whereas after the First World War the supposedly more scientific and rigorous “Experimenter’s View” became paramount. In this chapter, we usually take the Analyst’s View unless we specifically indicate otherwise. Someone performing a protocol can d and in a complex protocol will d take any or all viewpoints as the protocol proceeds. A third party who does not participate in the protocol can take only the passive

Perceptual control in cooperative interaction Chapter | 10

307

Observer View, and is always subject to the problem embodied in the PCT mantra: “You can’t tell what someone is doing by watching what they are doing”. To illustrate this last point, consider the following interaction between two people sitting on a park bench, overheard by a casual bystander: Person A: What time is the party? Person B: Six o’clock. To the bystander, what happened seems quite obvious. The two strangers executed a “Question-Answer” protocol. They must be planning to meet for some food and fun some evening, the place and date being known to both, but the time unknown to Person A. But in this example, the two happen to be spies. Person A knows she is to meet someone, but is unsure who. Person B tells her that the person she is to meet is someone they have previously designated as Number 6. The bystander was correct about what protocol they performed, but not about its content. The example illustrates a point important for the analysis of protocols in general: the content of a protocol is quite independent of its type (in this case a Question-Answer), and the form of the content is completely arbitrary, provided that it allows the participants to control their perceptions by means of each other’s actions. Person A might have posed the Question by looking quizzical, and to provide the Answer, Person B might have casually extended six fingers as though examining the fingernails, while even an observant bystander might not have noticed that they interacted at all. A great deal of what happens in a protocol can be hidden from casual outside observers if the participants want it hidden. Prisoners often do just that, to keep their interactions private from the guards.

Perception of uncertainty You look at little Johnny and say to yourself “He’s shooting up really fast. I wonder if he has grown too tall for child fare on the bus? I believe he has.” You know that the bus has a post with a marker at 4 ft (about 1.2 m) against which the driver can check the height of children claiming child fare. What are you perceiving, and can you control it? What do you mean by “wonder” and by “believe”? You perceive yourself to be uncertain3 d you wonder d whether Johnny is too tall to be allowed to ride for the child fare, so you measure his height with some instrument such as a measuring tape. The limit is 4 ft, and you measure

3. I do not here propose a mechanism for the perception of uncertainty. I introduce it because it is a perception that we obviously have. Since it is not included in the Powers hierarchy, it must be considered to be a deviation from HPCT. The evidence for the existence of a perception of uncertainty is that such a perception can be conscious, that we can have a reference value for it, and that we can act to control its value.

308 SECTION | C Collective control and communication

Johnny at 3 ft 10in (1.17 m), so now your uncertainty as to whether he will be allowed child fare is much reduced, and you believe the opposite of what you weakly believed before making the measurement. Your prior weak belief that Johnny was too tall has changed into a strong belief that he is not too tall. Your uncertainty would have been equally reduced if you had found him to be 4 ft 2in (1.27 m), but your resulting belief would be different. As with so many other perceptions, if you have a particular desired (reference) value for your perception of uncertainty about something, you may have a means to influence your uncertainty perception closer to that desired value. If your reference is to be less uncertain than you are, you can act either to get more information or to change the situation, or both d a gambler might try to learn about the horses in the race, or might bribe a horse trainer to drug opposing horses so that they would run more slowly. But a handicapper who adjusts the weights being carried by the horses in the same race controls for maximum uncertainty about which horse will win. The uncertainty one perceives is not a property of one’s environment like the intensity of a light, the sequence of sounds when a mechanical clock strikes the hour, or even the political direction of a party for which one might vote. Uncertainty is always about something, and that something is a perception, whether it be Johnny’s height relative to an arbitrary value, the name of the capital of Tajikistan, or whether that person in the distance is a friend. Uncertainty always involves the possibility of different perceptual values based on the same data. If there is only one possible perceptual value, there is no uncertainty about that perception. “Belief” about something is a term usually used when there is some uncertainty about it. But whereas non-belief and non-disbelief about something are much the same, both implying high uncertainty, low uncertainty corresponds both to the very different states of strong belief and to strong disbelief in the truth of that something. The control of uncertainty and belief perceptions, especially perceptions of another person’s beliefs, is at the heart of the analysis of human interaction using protocols, which we now introduce. Because the control of perceptions of various beliefs is central to the analysis of protocols, we use a shorthand notation to avoid having to use long and complex verbal descriptions (Box 10.1). For example, if Irene believes P to be true, we write Irene(P ¼ B). The notation is recursive, so that if “P” is “Carl perceives X to be uncertain” d Carl(X ¼ U) d we can use the synonym Irene(Carl(X ¼ U)¼B).

Protocols This chapter takes for granted most of the complexities of the simultaneous control of many perceptions by one person, and concentrates on the analysis of a particular pattern that occurs when two people interact in a well practiced way. We call this pattern a “protocol”. Protocols, as interactions, have a wide

Perceptual control in cooperative interaction Chapter | 10

309

BOX 10.1 A notation to describe belief states in a protocol A shorthand notation is required to avoid a plethora of long-winded statements along the lines of “John believes that Beth does not believe that John is uncertain about X”. To indicate that John has some belief about the truth of a statement X we write John(X). We write John(X ¼ B), John(X ¼ U), or John(X ¼ D) where “B” stands for Belief, D stands for Disbelief and U stands for Uncertainty. At some point in an interaction, Jill may believe John believes X, in which case the proposition in question is not “X” itself, but “John believes X00 , so the notation for “Jill believes that John believes X00 is Jill(John(X ¼ B) ¼ B). If the analysis allows for an arbitrary state of a particular belief, we use a question mark, as in John(Jill(X ¼ ?) ¼ B) to represent “John believes it does not matter what Jill believes about X”. This is distinct from John(Jill(X) ¼ B), which means “John believes Jill has a belief about X”. For example John(Jill(The plums John bought are red ¼ ?) ¼ B) implies that John believes it doesn’t matter to Jill what color plums he bought, whereas John(Jill(The plums I bought are red) ¼ D) might imply that John doesn’t believe Jill even knows he bought plums. In summary: l John(X): John has a belief about X l John(X ¼ B): John believes X to be true. l John(X ¼ U): John is uncertain about whether X is true l John(X ¼ D): John believes X is not true l John(Jill(X) ¼ B): John believes “Jill has a belief about X00 l John(Jill(X ¼ U) ¼ B): John believes Jill is uncertain whether X is true. l . and so forth. Occasionally, we may need to represent weak belief or biased uncertainty. These states are represented by lower-case “b” or “d” in the appropriate places.

range of complexity. Some examples of protocols at a level that typically involves turn-taking4 (only because turn-taking eases description in text such as this): l

l l

On meeting a friend: “Good Morning. How are you today” “I’m fine. And you?” “Very well, thank you”. The same protocol implemented differently: “Hey buddy” “Hi”. At the table: “Arthur, could you pass the salt, please”; Arthur passes the salt.

4. Most protocols do not involve turn-taking. At low levels, usually only one of the partners acts except when there is misunderstanding, while at high levels, both partners act simultaneously throughout the interaction. All “Dialogue Analysis” is done with protocols at a turn-taking level.

310 SECTION | C Collective control and communication l

l

l

l l

l

l

l

The same protocol: John waves a finger toward the salt container; Arthur passes the salt. In conversation: “Who is that with Zoe?” “I don’t know. I’ve never seen him before.” The same protocol: Jane nods toward a couple. Joan looks and turns to Jane with a puzzled expression. In a shop: “I think the blue.” “That one is $69.95.” “I’ll take it”. In the same shop a short while later: Customer silently shows the blue object and a credit card to cashier. Cashier passes to the customer a card reader whose display reads “Tap or insert card”. Customer taps the card on the reader. Cashier takes back the reader, which now reads “Authorized”, and says “Thank you” to the customer, the only verbal event in the performance of this “pay for goods” protocol. At school, a chemistry teacher tries to get a student to understand the concept of valence over a span of several weeks, during which time the student does experiments and asks questions. Eventually the student aces her chemistry exam. A three-person protocol: Donald: “Mary, would you ask your child to stop making that racket.” Mary: “Bobby, (Bobby looks up) could you play a little more quietly, because Mummy is trying to have a conversation” Bobby: “Sorry, Mummy.” (Bobby quiets down). Donald: “That’s good, Bobby. Thank you.” (We analyze this “triadic” protocol in Part 5 of this chapter.) Agitator with a bullhorn: “Let’s go get’em”. Crowd moves to attack “them”.

These examples suggest a few of very many possible protocols. The first pair are almost rituals in which the actual word sequences are almost precisely prescribed (collectively controlled), though those words will differ among cultures and will depend on the roles the two parties are playing. They might be the same people in both, but in contexts of different formality. The others clearly are not ritualized. In another culture some of the example protocols might not exist, but other protocols, unknown in this culture, would serve similar functions. Tapping a credit card on a reader to pay for a purchase is likely to be unavailable in some countries, but any culture that uses money has a protocol for paying for goods and services. The words used do not define a protocol. In some cultures, the participants in the first example might be enquiring about each other’s health, but in others, they would only be greeting each other and displaying continued goodwill, and would be surprised if either followed the “How are you” question with, “My hip is much better, but I think I’m starting a cold”. The protocol between the chemistry teacher and the student extends over weeks, whereas some of the others are complete in seconds.

Perceptual control in cooperative interaction Chapter | 10

311

All protocols have the same loop structure through two persons (and occasionally more, as in the last two examples). In a dyadic protocol, one person (the “initiator”), to control some perception, acts so as to disturb a perception controlled by another person (the “continuer”). The continuer acts in a way that reduces the error in the perception controlled by the initiator, and the error reduction changes the initiator’s action so as to reduce the disturbance initially introduced, thus bringing the controlled perceptions in both participants closer to their reference values. The initiator’s control uses the continuer as a kind of tool (an atenfel), with the continuer’s acceptance of that role incidentally requiring the use of the initiator as a complementary tool. Acceptance of the “tool” role is an essential aspect of a protocol. In Shakespear’s “Othello”, Iago drops a handkerchief so that it will be found by Othello in a context that would disturb his perception of Desdemona’s faithfulness. Othello’s action in controlling that perception is what Iago wants to see, but Othello knows nothing of the perception Iago controls by the handkerchief drop, so Iago and Othello are not participating in a protocol interaction. Indeed, if Othello had perceived that it was Iago who placed the handkerchief, quite different perceptions in him would have been disturbed, and his later angry control actions would probably have been directed at Iago rather than at Desdemona.

Classes of protocol In any culture, human or non-human, protocols can be grouped into classes to which we can apply labels. Four broad classes might be labeled “Getting Help”, “Helping”, “Teaching”, and “Learning”. Irene’s successful effort to get Carl to close the window is an example of “Getting Help”. An outside observer might easily confuse “Helping” with “Getting Help”, since the actions of both parties look the same if the observer does not know who initiated the protocol. “Teaching” or “Informing” broadly covers situations in which the initiator controls a perception of the state of someone else’s beliefs about the world. The encounter with the Japanese visitor that introduces this chapter belongs to this class, as my colleague had a reference for the Japanese visitor to perceive his environment as non-threatening. A request such as “What time is it?” would initiate a protocol of the “Learning” class, in which the questioner hopes that the continuer’s action will alter the questioner’s perception of the current state of the world. Protocols are not limited to these four classes, but these four have many different protocol subtypes. For example, one might define a “Trading” protocol class in which the participants each provide the other with something they want. “Trading” could be seen either as a separate class, or as a hybrid of “Helping” and “Getting Help”. “Contracting” might be a form of “Trading”, in

312 SECTION | C Collective control and communication

which promises are traded, rather than objects or services. Protocol classes can be further refined indefinitely. For example, a subclass of “Learning” might be “Question-Answer”, while a subclass of Trading might be “Buy-Sell”, in which one of the partners transfers only money to the other, in contrast to “Barter”, in which no money changes hands. The fact that we can identify classes of protocol according to what the participants control for does not reduce the flexibility inherent in the control processes involved in performing the protocol, but legal restrictions may do so. For example, there may be a requirement to perform some kind of ritual if a “Contracting” protocol is to be completed, so that any member of the public can attest to just what promises were included in the contract. The “Contracting” protocol itself is not thereby turned into a ritual, but it does indicate that rituals may be incorporated into the performance of a protocol. Buying a house, then, would be a “Buy-Sell” supported in part by a “Contracting” that is supported in part by a ritual, whereas buying an ice-cream would be a simple “Buy-Sell”. We treat support of one protocol by another, and the implied multi-level structure of protocol systems, toward the end of this Chapter.

Feedback loops and control loops To lead into the question of perceptual control in cooperative interactions among people, we must discuss control itself, taking a wider view than is usually done when introducing Perceptual Control Theory. We start with a discussion of loops that have more active elements than the simple control loop of which the Powers hierarchy is constructed. We need this discussion, because at an absolute minimum when two people interact cooperatively, at least two ECUs are in the same loop, each affecting the operation of the other, and each ECU has two active elements. We start with loops that have an indefinite number of active elements.

Generalized feedback loops In Fig. 10.1, a control loop was shown as a single ECU that accepts multiple inputs from a mysterious environment and provides output to that environment, with the loop being completed through the environment. The ECU has two active elements, the Perceptual Function and the Output Function. In most illustrative analyses of PCT systems, the perceptual function is shown as a unity multiplier. Instead, here we give the two functions equal prominence. The analysis of control usually used in PCT discussions subsumes all the possible complexity of the environmental feedback path into a single function of time, as though there were only a single path from output to input with a Complex Environmental Variable (the small circle in the “Environment” part of Fig. 10.1B) in the middle. Moreover, the potentially complicated time function that describes how a single output impulse might return to influence

Perceptual control in cooperative interaction Chapter | 10

313

FIG. 10.4 A Rube Goldberg control system for maintaining the weight of dry sand in a leaky pan, showing how the different signal paths in a control loop can be of very different kinds.

the perceptual signal is often reduced to a simple equivalent transport lag. We keep much of this enormous simplification, but add the active components of the partner’s ECU into the environmental feedback path. We will need these and more when we finally get to discussing protocols. Fig. 10.4 illustrates a complicated way of maintaining a certain weight of sand in a leaky pan. l

l

l

l

A is a balance that senses the difference between the weights on its two scale pans. When the right pan, which holds a set of metal weights, is heavier, the indicator arm presses a microswitch, which turns on a red light. B senses the red light and if it is visible, B produces a magnetic field in a coil. The field stops when the red light shuts off. C is a gate that allows water down a flume when it is opened by the magnetic field. D is a water wheel driven by the water coming down the flume. The wheel drives a belt that delivers sand onto the left pan of the scale, which leaks the sand away over time and would become empty if the sand were never replenished. A continuous supply of sand runs onto the belt, but falls off back into the supply pit if the belt is not moving.

Provided that the belt can deliver sand faster than it leaks out of the pan, this loop keeps just enough sand in the left pan of Scale A to nearly balance the metal weight in the right pan.5 It is a homeostatic loop, and a negative feedback loop that has four identified active elements, each of which produces an “output” that differs qualitatively from its “input”. When we talk of control, we are interested only in cases in which energy sources independent of the loop variables power the active elements. The 5. It actually will never achieve a stable value, but will follow a stable sawtooth-like oscillation around the reference value, the size and shape of the sawtooth depending on how fast the water wheel spins up and how fast sand is applied to the pan when the belt is moving at top speed.

314 SECTION | C Collective control and communication

FIG. 10.5 The components of element A of the Rube Goldberg loop. The sand pile is currently heavier than the scale weight, but the pan is leaking sand, and soon the scale will tilt the other way, setting the microswitch and turning on the red light.

separate energy sources allow the active elements to transform the kind of influence that is input to a possibly different kind that is output. The “Rube Goldberg” loop has separate energy supplies to the photon detector, to a magnet, to a water source, and to the driver of the sand belt. We shall refer often to variants of this loop in what follows. Each of the active elements could be subdivided into smaller units, as is usually the case with macroscopic systems. For example, A (the scale) could be described as a set of active elements: (1) a balance beam A1, which (2) closes or opens a microswitch A2, which (3) allows or prevents current flow through the red light A3 (Fig. 10.5). The combinations of active element groups around the loop into four unitary components is an analytic convenience, permissible unless an influence from outside is possible inside the combination, or the analyst is concerned with some effect within the combination. If, for example, one was interested in the sound of the click of microswitch A2, perhaps because the red light was failing to go on and one was trying to diagnose the reason, then it would be appropriate to consider A1, A2, and A3 separately. Otherwise, one can treat just “A” as a unit that produces a red light if there is not enough sand in the left pan.

Atenfels We can treat the Rube Goldberg loop as a control loop that maintains the weight of sand on the left pan of the scale near a reference value provided by the scale weight on the right pan. The scale A has all the elements of an Elementary Control Unit (ECU) if we take its components as distinct, while the rest of the loop is in its environment, as suggested in Fig. 10.6. The environment, however, is not just a simple connection of the output of the ECU to a possibly disturbed CEV and thence to its sensory input. Many changes of character happen to the signal between the output (the red light being on or off) and the sensory input (the weight of sand on the left scale pan). Quite apart from what happens inside the elements we have arbitrarily identified as unitary, the feedback signal is first an optical photon stream, then in turn a magnetic field, a stream of water, and sand on a traveling belt. Any of these could be independently disturbed. The operations of the different devices

Perceptual control in cooperative interaction Chapter | 10

315

FIG. 10.6 The homeostatic system seen as a controller of the weight of sand in the left scale pan, showing the sequence of elements that constitute the environmental feedback path. Although each of them could be further broken down, the analysis of the loop can treat B, C, and D each as a unitary element or together as a simple feedback path.

B, C, and D might become stronger or weaker; the water wheel might have increased friction or the water stream might depend on recent rainfall, or the sand supply rate onto the belt might change up or down. None of this matters so long as the belt continues to be able to deliver sand to the scale pan faster than the rate at which the sand drains out of the pan. If that is true, the weight in the sand pan will oscillate near that of the weight in the scale pan, no matter what happens to the various components of the environmental feedback path. The controller implemented by the loop knows nothing of sand. Its “perceptual function” reports as a perceptual signal only the difference between the weights in the two pans in the form of a pointer deflection, whether what is being supplied is sand, gold dust, pebbles, or even a liquid. What is controlled is a perception of weight difference, not a quantity of what may or may not be sand. The atenfels (elementary units of the feedback path) labeled B and C could be replaced by a human operator of the water gate. Although as Analysts we can see that the operator would be able to open and close the water gate by flipping a switch (Fig. 10.7), he cannot. He knows nothing of gates, water, or

FIG. 10.7 The red-light sensor that produces a magnetic field could be replaced by a human operator who is like a cat always on the wrong side of the door, because he wants the red light to be off if it is on and on if it is off. The operator’s only possible action is to open and close the water gate.

316 SECTION | C Collective control and communication

sand. All he knows is that there are two lights, the red light switched by the microswitch and a console light that we Analysts know to signify the openclosed state of the water gate, and he wants the lights both to be on or both to be off. He controls his perception of the relationship between the two lights, with a reference value of “same”, and that is all he does. It makes no difference to the sand control system (for which the ECU is the scale with the microswitch) whether its environment contains a human operator or the BC sequence of machines, provided that the human operator efficiently controls his perception of the relation between the red light and his console light. If he stops controlling that perception, the sand control system also fails. If the sand is not deliverable faster than it leaks away, the human will not be able to control the relationship between the lights. The human who controls the relationship between the lights is just as much an atenfel for the sand controller as the scale is an atenfel for his control of the light relationship. Each is in the environmental feedback path of the other, and neither can control his/its perception if the other’s control fails. This kind of co-operation is at the heart of the protocols we shall be discussing.

Four-element loops The standard control loop of Fig. 10.1 has two active elements, but negative feedback loops can be much longer, as suggested by Fig. 10.4 and by either version of the Rube Goldberg system. Although more complex loops do occur in protocols, we limit our discussion to the simple four-element loop with no branching paths (Fig. 10.8), because even when it is not the only loop in a protocol, it is usually the dominant one. More complex protocol structures can be important, but are beyond the scope of this Chapter. In our analyses, we treat as a single active element of the loop any sequence of components that is not interrupted by an outside influence. In Fig. 10.8, an active element consists of the combination of a 4 symbol followed by the circled “g” symbol. It has two inputs, one “d” from outside the loop, and one “p” from the output of the preceding active element in the loop. Its output goes to the next active element around the loop. We ignore possible side-effects, but they should not be forgotten. A “d” input corresponds to either the reference signal or the disturbance signal in a control loop diagram.6 In Fig. 10.8, the different p-variables, or “signals” communicated between the elements, are treated as simple numeric values, even though the concepts they represent may be qualitatively different. In the Rube Goldberg sandpile stabilizer example, there are eight places where the qualitative nature of the 6. In a “standard” diagram of the control loop, the comparator subtracts the perceptual signal from the reference signal. In the generic diagram, this effect is accomplished by a sign inversion of the previous “g” function. For feedback to be negative, there must be an odd number of such sign inversions around the loop.

Perceptual control in cooperative interaction Chapter | 10

317

FIG. 10.8 (A) The standard control loop of Fig. 10.1B, relabeled as a two-element generic feedback loop. (B) A four-element generic loop. Each element “perceives” the output of the previous one, plus a possible influence (“disturbance” or in panel (A) the “reference”) from some outside source.

signal changes, but this does not matter, as the changes all cancel out in a complete circuit of the loop. Fig. 10.9 shows a feedback loop representation of the “human operator” version of the Rube Goldberg example of Fig. 10.7. Neither the scale nor the human operator in Fig. 10.7 and Fig. 10.9 could influence their own perceptual variable if the other stopped controlling their completely different perceptual variable. Each is an atenfel for the other’s control (as are the loop components labeled “Coupling Constants” in Fig. 10.9).7

Protocol representation We are finally in a position to deal with protocol structure and protocol function. When the partners belong to the same “culture”, whether it be the culture of a family, a profession, a religion, a trading community, a street gang, a nation, or even a hunting pack of wolves, the members have collectively developed ways making their interactions user-friendly, even if the participants are opponents.8 The participants in the interaction truthfully or deceptively display to each other some of their internal states and reference values well

7. We use the term “coupling constant” because we are concentrating on the control of only two variables, one by the scale and one by the human operator, not because there is any formal distinction between these and pathways internal to the scale and the human. The coupling constants are in the environment of both, which is important for an outside observer, but not for the analysis of the structure as a four-element control loop. 8. Diplomatic protocols allow people from different cultures, even enemies or potential enemies, to interact effectively. As the diplomat, George Ignatieff said: “I learned that protocol is really a language, a set of rules and conventions which enable people of different nationalities, social backgrounds and political persuasions to feel comfortable with each other” (Ignatieff, 1985).

318 SECTION | C Collective control and communication

FIG. 10.9 A specific example of a four-element control loop, showing the components of the Rube Goldberg apparatus with a human operator. The dashed lines that would normally complete each of the two local control loops are instead implemented by the long loop through the other ECU. The state of the red light is not directly influenced by the operator’s control of its relationship with the console light, which determines the “Desired state of red light” in the Figure, and the rate of sand delivery is not directly influenced by the angle of the scale arm.

enough to allow the partners to use each other’s actions in controlling some perceptions. These displays determine the progress of the protocol. To this point we have developed only a control loop representation, but this is neither the only possible nor usually the most convenient way of looking at a protocol. Indeed, in earlier presentations of Layered Protocol Theory, a protocol was only represented in graphic form, as a network called a “General Protocol Grammar” (GPG). Perceptual control was either ignored or treated largely as an implementation mechanism. We now redevelop the GPG, treating it as a natural consequence of perceptual control of uncertainty and belief.

A protocol example Carl is standing by a closed window, while Irene sits in a wheelchair across the room. Scenario 1: Irene presses a switch on the arm of her wheelchair, and the window beside Carl silently opens. Scenario 2: Irene says “Carl, it’s a bit stuffy in here. Would you mind opening the window.” Carl silently opens the window and Irene says “Thanks. You’re a dear.” In the first scenario, the switch clearly provided an atenfel for Irene’s control of her perception of the window state. In the second scene, Carl did what the switch would have done, just as the human operator controlling a perception of the red light state provided an atenfel for controlling the weight

Perceptual control in cooperative interaction Chapter | 10

319

in the pan of the Rube Goldberg apparatus. But did Carl only provide an atenfel for Irene’s control of her perception of the window state? What, the PCT Analyst might ask, is the difference between these two scenarios? In both, Irene controlled a perception of the window state, with a reference that it be open and in both scenarios, her action brought her controlled perception to its reference value. But how? In scenario 1, Irene could perceive and control the state of the switch, but in scenario 2 she could not know what actually happened inside Carl. Nevertheless, in both scenarios the window opened. Irene may not know what happened inside Carl, but the Analyst has the advantage of being able to hypothesize possibilities about both Irene and Carl, to see which best fits the observation encapsulated in the description of each scenario. One might even be correct. So let us begin with the most obvious, which is that Irene in both scenes controlled a perception of “air quality”, initially in error because Irene had a reference value of “fresh” while she perceived “stuffy”. The HPCT analysis of Scenario 1 is self-evident, so we consider only Scenario 2. Irene perceives that an open window could provide an atenfel for controlling her air quality perception. She is not currently in a convenient position to open it, but Carl is. She has previously reorganized so that she is often able to get Carl to do what she wants. To use him as an atenfel, she disturbs some perception she perceives him to control, perhaps his perception of her contentment, which has worked on previous occasions. If he is indeed controlling for her to be contented, then when Irene displays that she is discontent he will act to increase her contentment if he knows how, provided that it does not conflict with his control of some other perception. For Carl to increase her appearance of contentment, thereby decreasing error in his own controlled perception, his action must decrease error in her controlled perception, which should change her contentment display. In this example, he must do it by actions that influence her perception of the window. Fig. 10.10 illustrates this situation as a four-element loop with a timeline, starting with Irene’s air quality perception. Another possibility: Irene does not control a perception of “stuffiness”, but controls a perception of Carl’s feeling toward her. She tests whether Carl is controlling for her contentment. If he is, then her disturbance should result in his opening the window. He does, so Irene perceives him to be controlling for her contentment and by thanking him allows him to perceive that his action did increase her contentment, thus reducing the chance of his control structure being changed by reorganization. Next time Irene wants something done, she may again be able achieve her objective by disturbing Carl’s perception of her contentment. Yet another possibility: Irene perceives that Carl is not especially well or ill disposed toward her and she controls for perceiving a situation in which she

320 SECTION | C Collective control and communication

FIG. 10.10 The control diagram for an intermediate stage of the interaction in which Irene asks Carl to open the window because it is stuffy. The values are shown for the moment before Irene thanks Carl, while the air still feels stuffy. In this situation, Irene cannot or will not open the window, and Carl cannot control his perception of Irene’s contentment without influencing her state. The only way either Carl or Irene control their perceptions is through the four-element loop, which allows both or neither to eliminate their perceptual error.

can increase his goodwill, by giving him an occasion to do something that would allow her to use the expression “You’re a dear”. What of the possibilities on Carl’s side? Apart from the obvious possibility mentioned above, here are a couple of others. Carl may not hear Irene’s request, but coincidentally wants to talk to someone outside, for which an atenfel is an open window. Irene’s “Thank you” might come as a surprise, if he hears it at all. Or Carl may control for Irene to perceive him as annoyed at her request, but also, independently of Irene, controls for the room’s air quality, so he opens the window abruptly and turns away. There are multiple possibilities on both sides, most of which involve at least one of the pair controlling a perception of something about the other. In the situation as described, Irene would not have said either of the things she did if she did not want to influence some perception in Carl, and even though Carl said nothing to Irene, many of the different perceptions Carl might be controlling by opening the window are perceptions of Irene. Even his failure to speak might display to Irene something that affects a perception she controls, as would his turning away.

The General Protocol Grammar: introduction The “obvious” explanation above of the interaction between Irene and Carl describes one possible example of a protocol class we called “Getting Help”. Irene got help in controlling her perception of the room air quality.

Perceptual control in cooperative interaction Chapter | 10

321

In general, whatever the protocol, Irene (a name I now use as a generic label for the protocol initiator) controls her perception of something about Carl. She has reorganized so that she has various means (atenfels) for influencing this perception, one of which is that a certain way of acting to disturb some perception Carl controls will result in Carl acting to reduce the error in her perception (e.g. Carl opens the window, or Carl says he now understands what she has been trying to tell him). This is the basic form of simple protocols, a four-element loop. What Irene wants from the protocol has been called the “Primal Message” in previous writings on Layered Protocols, a practice we shall continue. The protocol is successful when Carl understands what she wants, Irene believes he understands, and Carl believes Irene believes he understands. She may be controlling for Carl to perform some action, or she may simply want him to change something about the way she perceives that he currently perceives the world. Whatever she wants him to believe or to do, when he understand what she wants him to understand, she has successfully sent the “Primal Message”. Whether he actually does what she wants is a separate issue. If he does, the question of his actual understanding becomes irrelevant. He might have done it as part of controlling some perception entirely independent of Irene, but so far as Irene is concerned, the controlled perception that led to her initiating the protocol is now at its reference value. She need not act further to ensure that he actually understands what she wanted, since he did it anyway. The classic “Getting Help” protocol, in which both parties think that they immediately understand one another, results in the externally observable actions illustrated in Fig. 10.11, which shows the same interaction between Irene and Carl. Fig. 10.11 is a first step in representing graphically a “General Protocol Grammar” (e.g., Taylor, 1988; Taylor et al., 1999; Taylor and Waugh, 2000), which in its full form applies to every dyadic protocol (a protocol involving only two participants). “Sending” the “Primal Message” is the execution of the entire protocol represented in the Figure as a network. In this and later Figures, nodes are shaped differently according to which partner is acting to influence the state of the other.

FIG. 10.11 The externally observable actions in a basic protocol. Circles labeled IreneX are points where Irene’s actions change the state, while squares labeled CarlX are points where Carl’s actions change the state. Both may act simultaneously.

322 SECTION | C Collective control and communication

The “General Protocol Grammar” (GPG) is one way of representing an analyst’s view of the possibilities inherent in an interaction, in much the same way as a grammar of a natural language describes an analyst’s view of the possibilities for the ways people might use the language. It is not a prescription for what should happen, but a description of what usually (but not always) happens.9 Underlying the GPG, as with a natural language grammar, is a set of assumptions about why what is observed is observed, including: l

l

l

l

that people can “display”; they can act to allow some aspects of their internal states, such as their state of happiness, to be perceived by others, truthfully or deceptively. that people always control their own perceptions and nothing else, that people can and do perceive some of the states displayed intentionally or as side-effects by other people, and that they can act to control those perceptions; that perceptions of others may be wrong, in the sense that a person may wrongly be perceived to be displaying happiness when they do not actually feel happy, that among the states displayed may be strengths of belief and uncertainty.

As a descriptive technique, the GPG is not restricted to describing protocol interactions. It can also be used to describe an ordinary perceptual control process. As an illustration, we use it to describe someone called Paul moving a moderately heavy rock by pushing it across a surface (Fig. 10.12). In Fig. 10.12 the Rock “displays” only its location, a property of the rock that Paul can perceive. Whereas Fig. 10.10 shows in one diagram all the processes involved in the interaction but has no sense of time, Fig. 10.11 and Fig. 10.12 are more suggestive of a time-line of observer-visible outputs while indicating little or nothing of the processes involved. Each looks like a state transition network diagram, seeming to suggest a discrete sequence of events, but the suggestion

FIG. 10.12 The basic protocol as a control process in which Paul pushes a Rock to a new location, stopping when the Rock reaches his reference location for it.

9. It can, however, be a prescriptive device for designing human-computer interaction, in much the same way that a dictionary or a grammar book can be a prescriptive device when teaching “proper” use of a language.

Perceptual control in cooperative interaction Chapter | 10

323

is deceptive. Just as PCT emphasises that signal flow is continuous around a control loop and that all the variables change simultaneously, so also the rock moves while Paul pushes. Therefore in the GPG of Fig. 10.12, the left and the middle arrows represent simultaneously active arcs10. Paul starts out fully at the node labeled PaulS, but as the rock nears his reference for its location, he decreases his fuzzy membership in PaulS and increasingly occupies Paul2. The membership of a protocol user in a node is related to (but not necessarily equal to) the probability that he or she will next use an arc exiting from that node. When the rock is for his purpose close enough to its reference location, he is completely at Paul2 and stops pushing. The “Acknowledgment” (stopping pushing) that the Rock is in the desired place does occur in time after the other processes. An “Acknowledge” occurs last in an ordinary protocol interaction, too, because it is intended to terminate the protocol. Each arc in the graph represents an “arc-message”, a message that is a Primal Message in a lower-level supporting protocol that implements the arc. The labels in Fig. 10.12 “Paul pushes” and “rock moves toward target location” represent such supporting, lower-level, protocols. In the case of Paul and the rock, these lower-level “protocols” are trivial, but in the more general case of interpersonal interaction they can be very complicated, involving several supporting levels of protocols. An “arc-message” becomes the Primal Message of a supporting protocol because it represents action to change some perception of a state of the partner, in the same way as does the Primal Message of the main protocol. How an arc-message is implemented is irrelevant to the progress of the supported protocol, provided that it actually does its job. The “Normal Feedback” arc-message of a cooperative interaction in Fig. 10.11 is a good example. The “Normal Feedback” message is an action by a human Continuer (Carl) allow him to perceive that the Initiator (Irene) knows what Carl now believes he understands of the Primary Message d the content that the Irene wants Carl to understand. Sometimes, Irene expects Carl to understand immediately, and Carl need do nothing unless he has not understood. “Doing Nothing” is called a “Null” instantiation of an arc, meaning that there are no supporting protocols because the relevant perceptions are already at their reference values. Sometimes Carl wants to display to Irene only that he has understood the Primal Message, without wasting time for both of them by describing what he understood. We call such a content-free message a “Neutral” instantiation that has a very simple supporting protocol. Its supporting protocol probably has

10. The “Primal Message” in this illustration is the moving of the rock from where it was to where Paul wants it, because the rock does not “understand”. It simply does what the laws of physics demand. As always, however, Paul controls his perception of the Rock’s “display”.

324 SECTION | C Collective control and communication

FIG. 10.13 The General Protocol Grammar representation showing continuing correction of remaining error in Paul’s perceptual control of the location of the Rock. The figure-eight pattern represents the normal time course of error correction. If some other influence were disturbing the Rock’s position, it would be shown entering RockP beside the Edit arrow.

“Null” arcs everywhere except the Primary Message that is effectively “Got it”. At higher protocol levels, most Normal Feedback messages are intended to let the Initiator know just what has been understood (and what has not). These are “Inform” instantiations. Those may have complicated supporting protocols with much depth. Less frequently, Carl may think that Irene displayed something with an obvious interpretation that is not the interpretation the Carl believes she intended. In that case, Normal Feedback would be a “Correction” or “Correct” instantiation. For example, if Irene referred to “the intersection of 12th and Jones” when Carl does not believe these two streets intersect, he might say “Don’t you mean 12th and Johnson?” Next we add error-correction arcs and nodes to the GPG.

Extending the GPG: error correction Paul moves the Rock on a surface with low but non-negligible friction, so that after a sharp push it moves some distance, slows, and stops. Paul can’t push while walking on the slippery surface. Paul tries to move the Rock to its desired location with the first push, but it stops some distance away from his reference for his perception of its position. What then? Presumably he will walk to the new place and give the Rock another push, and will continue this cycle of push and walk until the Rock is satisfactorily close to where Paul wants it. In the standard two-element control loop diagram, the processes involved are all exactly as before, but what about the GPG representation? Fig. 10.13 adds an “Edit-Accept” sub-loop that handles this situation. As always, the graphical notation neither requires nor precludes simultaneous and continuous activity on these and other arcs. The Edit-Accept loop simply represents the entire control loop in the functional diagram for what happens after the Primary Message and Normal Feedback have occurred. It does not duplicate the GPG elements of Fig. 10.12, because the Edit-Accept messages represent changes in state, whereas the “main line” messages represent the actual states.

Perceptual control in cooperative interaction Chapter | 10

325

If an external disturbance affects the perception of a message by either partner, the result is likely to be that the message fails to change the recipient’s perception in the way the sender intended. If the sender cannot “Edit” to correct the difficulty, this presents a new problem, or “Problem”, which we now consider, still in the context of moving the Rock.

Representing problems If the rock, no longer on an icy surface, but now stuck in soft ground, does not move when first pushed, Paul might try pushing harder, following one of the mantras told to schoolchildren: “If at first you don’t succeed, try, try, try again.” But when that doesn’t work, another mantra comes into play: “If what you are doing doesn’t work, try something else”. In the development of an individual control hierarchy, “trying something else” is sometimes called reorganization. In a simplified approach to reorganization, only the weights that interconnect the ECUs at different levels change; the structure of the hierarchy does not change. If that were all there was to reorganization, the original method of influencing the controlled perception, which had been successful on previous occasions, would be lost in favor of the new method that succeeds on this one occasion. Clearly, this is not what happens in everyday life. When one learns to use a lever to move a rock, one does not lose the ability to move a less recalcitrant rock by pushing or lifting it using one’s hands.11 Stated in different words, reorganization provides to a mature organism both new atenfels and improvements of old ones. In the GPG, this process of adapting atenfels or of reorganizing to create new atenfels is represented as another loop that is used until an atenfel is found that works. In the rock example, this new loop is started by a “Problem” arc to a new node called “PaulP” (for “Paul’s Problem”) and continues through the same “RockP” node as is used by the Edit-Accept loop (Fig. 10.14). When the problem of finding a working atenfel has been solved, normal control can continue around the Edit-Accept loop. Perhaps Paul has hired a backhoe and a competent operator, and very soon the Rock is at Paul’s reference value for its location. This is about as far as we can usefully go in describing the GPG representation when the “Continuer” is as simple as a Rock, which can display nothing to the Initiator except its position. The GPG is in any case overkill for describing simple control, and “moving the Rock” has now served its purpose in introducing some of the main concepts that are important for describing more complex interactions between and among humans, animals, or computers. 11. If the new mechanism becomes the primary means of control and the one that failed to work remains unused, the old way may indeed be reorganized out of existence, but this does not always happen. It is said that one never forgets how to ride a bicycle, even after many years of not having tried.

326 SECTION | C Collective control and communication

FIG. 10.14 The GPG including a “Problem” loop that requires the originator to try a new way of moving the Rock when pushing fails to move it. In a protocol, this loop is used when the Continuer does not perceive correctly what the Originator wants (or rather, when the Originator perceives that the Continuer has misunderstood what is wanted).

Protocols proper People are not rocks. They control their own perceptions. You can push a rock and it will move in the direction of your push, or not at all, but if you push a person, you are likely to get a push back, or worse. One person’s individual control actions may influence the CEVs that correspond to perceptions controlled by other people. Sometimes these cross-influences form a fourelement control loop, and some of those four-element loops operate to the benefit of both parties. Of those, we give the name “protocol” to ones in which the cross-influences are intentional. There is a particular set of controlled perceptions, perceptions of each others’ beliefs, which we claim to be the same for every protocol, no matter of what type or at what level of Primal Message complexity, just as the components of an Elementary Control Unit are the same no matter what kind of perception is being controlled, or how they are implemented. Let us remind ourselves of some assumptions so basic as to be easily forgotten. Many years ago I called them the “Three Independences of intelligent communication” (Taylor, 1993): 1. Independence of design: the partners were not designed in order to work together. 2. Independence of input: Neither partner can know everything that the other has perceived over its lifetime that affects its present state. 3. Independence of action: Neither partner can know what perceptions the other is controlling to produce observed actions, nor can they know the states that produce reference values for those perceptions. These independences are usually taken for granted, but they must be overcome if communication is to be at all possible, a problem that does not exist for simple control. If a person is moving a rock, independence (1) holds, independence (3) does not, and independence (2) probably does not matter,

Perceptual control in cooperative interaction Chapter | 10

327

unless the rock’s prior experience includes something like being bolted to bedrock in some previous encounter with a person. So in the case of the rock, only independence (1) need be overcome, and to overcome it in most cases involves only using whatever the Actor has learned through reorganization about controlling the perception of the locations of heavy objects. When the partner is more complex than a rock, such as a computer, an animal, or another person, independences 2 and 3 become important. Because of the second independence, not only does Irene not know what Carl knows from a lifetime of experience, she does not know what he might be perceiving during the performance of the protocol. The third independence means that Irene cannot not know everything Carl wants at any moment. Most protocols are not designed, but over much reorganization in interacting individuals, they have evolved to work effectively. The result is that protocols might seem to have been designed to overcome all three independences. We examine how this may be.

Protocol function: control of belief The three independences mean that both partners have some belief and some uncertainty about what the other knows, is controlling for, and can do. As discussed above, uncertainty and belief are themselves perceptions, albeit ones not incorporated in the Powers hierarchy of levels. When I am talking to you, I may believe weakly that you failed to understand what I intended to get across, or strongly that you misheard what I believe I said. Likewise, I may believe strongly that you are clear about what I intended you to understand. The level of one’s belief in the truth of a statement may change with any new observation. Though I disbelieved something yesterday, today I may learn something that makes me wonder, and tomorrow I may observe something else that makes me believe what yesterday I disbelieved. A short discussion of the approach to belief and the approximation to truth is contained in Box 10.2. Each node of the GPG represents a profile of belief perceptions controlled by the person represented by the node. The arcs represent intended patterns of changes in the other’s beliefs. If the nodes represent states of the dialogue, the arcs represent differentials in the dynamics of the dialogue. The main reference state for both participants in a cooperative dialogue is to believe that the continuer has correctly perceived the Primal Message. This is the first of three propositions that determine the flow of the protocol. Fig. 10.15 shows a different view of the same GPG as is shown in Fig. 10.14, including (dotted arrows) more ways of accessing the “ProblemResolve” loop, which has been separated out and relabeled “Problem Reorganization Sub-loop”. This GPG diagram is still much simpler than the complete 25-arc GPG described by, for example, Taylor et al. (1999), but it is sufficient to illustrate some perceptual controls involved in a protocol interaction.

328 SECTION | C Collective control and communication

BOX 10.2 Perception of proximity to “truth” In the main text, we base the analysis of protocols and the traversals of state through the GPG on changes in the participants’ perceptions of the level of their beliefs that certain propositions are true. “True” has an alternative of “false”, so logically a proposition must be one or the other. Concepts such as “nearly true” or “true in most respects” are ill-defined. However, in everyday control of perception, proximity often is important. It may be true that you are not “at home” when you are outside on your front doorstep, but it might matter whether you are close or in another continent. If someone believes you are “at home” when you are actually on your front doorstep, that might be true enough for their purposes. Beliefs about three propositions are at the core of protocol analysis. Let us examine one of them, P1. That the recipient has understood the message. We again use Irene and Carl as respectively the initiator and the continuer of the protocol. Using the notation described in Box 10.1, we argued that the reference values for Irene and Carl were, respectively: Irene(P1 ¼ B), Irene(Carl(P1 ¼ B) ¼ B), and Irene(Carl(Irene(P1 ¼ B) ¼ B) ¼ B), and. Carl(P1 ¼ B), Carl(Irene(P1 ¼ B) ¼ B) and Carl(Irene(Carl(P1 ¼ B) ¼ B) ¼ B). Between them, Irene and Carl control six perceptions of belief about P1. The problem is that if P1 is taken to be the proposition “Carl has understood exactly what Irene intended”, which can only be “true” or “false”, then it is almost guaranteed to be false. Furthermore, neither Carl nor Irene could ever know if by chance it happened to be true. Belief perceptions can never reach such reference values. Not only that, but as Irene and Carl interact, neither would have any way to bring any of the perceptions closer to its reference value, since the belief is exactly “false” until suddenly it becomes exactly “true” (if it ever does). They would be groping in the dark. In everyday parlance, however, we often say things like “She more or less understands” or “He’s way off-base”. Such statements seem to imply that perception of quality of understanding can be fuzzy. We can replace perception of belief in the strict truth of a proposition by perception of its membership in the class “true”. Irene(P1 ¼ B) as a reference condition translates into Irene(P1 ¼ x) where “x” takes on a value near 1.0. As a measure of Irene’s actual belief in P1, value of “x” could vary from near zero to near unity as the protocol progresses. If Irene thinks Carl’s understanding of what she intends is “way off base” x is near zero; if she perceives that “He’s got it; he’s really got it”, x is very near unity; and if she perceives “He’s getting there but has a ways to go” x is somewhere in the middle. In the analyses in the main text, these fuzzy ranges of perceptual value for the approximation of the proposition to “truth” can always be substituted for the letter-symbol ranges of belief perception, B, D, and U. For the second- and third-level recursions such as Irene(Carl(P1 ¼ B) Brisk silent reading

200

6.28

300

e3.28 PART | V Synthesis

TABLE 19.2 Velocity encoded in neural firing.

How the brain gets a roaring campfire Chapter | 19

e3.29

receptive field (RF) in the cortex. A human analog of this situation may be one where a person scans across a line of text, making adjustments in one’s reading speed. A simplification adopted for this example is that vertical lines of certain spatial frequencies mimic the more complex stimuli of letters being sequentially scanned. The magnitudes in the last column are utilized below to illustrate the need for bi-directional reference signals, due to neuronal firing being constrained by half-wave rectification. As outlined in this description of Table 19.2, the data were collected about firing rates in the cortex of a cat, while gratings were moved across its visual field at different speeds. Typically, animals and humans do their own visual scanning, which can be at different rates of speed, leading to spatial frequency signals in certain portions of their brains. This suggests an analog scenario where a person is reading text on a computer screen from two feet away. For the purposes of this illustration, the alteration of letters and spaces on a line could be considered spatial frequency lines of various spacings appearing in sequence across a moving receptive field window. A brisk rate of silent reading would cross the line of text in about 3 s, comparable to the data in the bottommost row of Table 19.2. Establishing that rate of scanning, assuming comparable neurophysiology between a cat and a human, would lead to certain neurons firing at 300 imp/s. If one wanted to slow down this reading rate by half, to capture a more difficult section of text, this would correspond to a firing rate of about 220 imp/s (according to the same set of data presented by Glezer), and indeed that would be the magnitude of the reference signal needed for whatever input was producing the 300 imp/s rate. By sending an inhibitory signal (by way of the TRN) of that magnitude to a corresponding comparator, the difference from 300 imp/s leads to a net error signal of 80 imp/s, which becomes the degree of correction needed for that particular micro control system. It would need to feed in an inhibitory way into output systems further down the line. The same comparator, however, cannot call for correction in the opposite direction. Suppose the rate of movement of the letters past the same striate RF were much slower, at that 100 imp/s rate. Now sending a desired reference standard of 220 imp/s to the same comparator would either reduce the output to zero (if entered as inhibition) or lead to overshoot (if entered as excitation). A method is needed to derive just the difference of 120 imp/s between the two quantities, to bring the current rate of 100 imp/s up to the desired rate of 220 imp/s. This is where the inversion of the perception by the mechanism of the dendritic triads comes into play. A companion set of relay cells operating in parallel but with reversed polarity to the original set provides a way to get correction in either direction. Now consider a bit more realism in this scenario. The striate RF investigated and described above by Glezer (1995) is actually composed of a pair of oncenter and off-center inputs from the geniculate nucleus of the thalamus (p. 27). According to the reading scenario, when a white space passes the center

e3.30 PART | V Synthesis

of the RF, an on-center input from the thalamus is firing, and when a dark letter passes the RF center, an off-center input is firing. But a temporal signal of ‘onthen-off’ is ambiguous. Striate RFs with the above profile of changing the rate of firing according to gratings passing across at different rates of speed could not distinguish between a flickering stimulus versus true movement. To discern if there is actually movement, one must check what is going on in spatially adjacent RFs. There would need to be a series of ‘on-then-off’ signals operating in sequence, one after the other as the person’s eye scans across the line of text. So then, reference signals to alter the speed of such a sequence must send a timed series of signals to lower level comparators. A single ‘on-then-off’ striate RF knows nothing of its neighbors registering their own patterns of ‘onthen-off’ inputs. A higher level of monitoring would be needed, registering inputs from multiple striate cells and controlling their progression into a coordinated train of movement as a similar perception shifts from one to the next, similar to the way adjacent blinking lights with the right timing give the appearance of a single moving light. Recall that this scenario is simulating the ‘on-then-off’ contrast patterns generated by the spaces and letters as one’s eye scans across a page of text. To change the speed of input of those letters means altering the rate of scanning. So the sequence of timed references generated at this level must ultimately make its way into a sequence of timed transitions for motor linkages controlling opponent-process (i.e., pull-on-one-side, releaseon-the-other) muscle forces moving the eyes. The result so far shows that an ‘on-then-off’ input must be arranged next to an ‘off-then-on’ input, which is itself adjacent to the next corresponding input, and so on for each letter the eye encounters. These alternating inputs become the desired reference standards for the series of on-center and off-center cells at the geniculate level of the thalamus. Each ‘on’ portion corresponds to a given level of illumination, assuming a roughly uniform white background for the text. So the references to the on-center cells are fairly standard, but at an alternating rate, depending on whether the space is between two letters or between two words. Similarly, the ink can be presumed to be of uniform darkness, so given a constant rate of scanning, those ‘off’ portions also arise at a similar rate, but in counter-phase to the ‘on’ inputs. The fact that the letters are not identical narrow gratings is being ignored at present, for simplicity’s sake. While each mini-level in the above scenario generates its own type of perception based on the forms of neural input at that level, many of these can be brought under control by common outputs affecting the rate of eye scanning across the page. So then, the retinal patterns of contrast give rise to the oncenter and off-center cells of the thalamus. Those adjacent thalamic geniculate cells offer possibilities for various patterns of alternating ‘on-then-off’ signals in the cortex. If those combinations are arranged and controlled in sequence, then patterns of movement can be generated or, if needed, shifted into different rates of speed. These and many other types of perceptual input functions are explored in detail in Chapter 7. The present discussion suggests

How the brain gets a roaring campfire Chapter | 19

e3.31

some of the ways those steps may be coordinated. Its main proposal is that the thalamus is an important juncture where cortical and subcortical connections meet, so that perceptual signals and reference signals can be compared.

Thalamic routing The previous subsections of this chapter have put forth the proposal that thalamic relay cells can function as PCT-style comparators, calculating the difference between ascending perceptual input and descending reference input. The resulting error signals can be conveyed with high fidelity by means of the tonic mode of relay cell firing. There is also a means available for heightening the gain on those error signals, via burst firing. So it is further proposed, on the basis of the neural anatomy and physiology of the thalamus and structures related to it, that these relay cells can oscillate in predictable and meaningful ways between tonic and burst modes of firing. It remains to give some indication of where these kinds of signaling have their respective effects. It is important to realize that an error signal, from a PCT standpoint, signifies how much output is needed to correct things, for a given perception. Whenever a preference is specified for how much of a given perception is wanted, the existing state of that perception essentially defines how close it has already come to the reference standard. So to calculate the difference between those two quantities becomes the amount of further correction needed. Apart from half-wave rectification, it would not matter algebraically where the minus sign is placed, as long as the result is the net difference. The above method of inverting the perception is only needed because the nervous system only propagates positive action potentials. Likely a functionally equivalent method would be necessary at other spots in the nervous system where bi-directional control might be needed. This sets the stage for understanding how the output of a PCT comparator has to function. Sherman (2007) makes the point, “the anatomical fact that many or all inputs relayed by thalamus are branches of axons that also target motor structures requires some further consideration” (p. 421). This begins to speak to the output side of this discussion. Jones (1998) lists a variety of target structures in the brain for thalamic outputs, including the basal ganglia striatum. Wei, Bonjean, Petry, Sejnowski and Bickford (2011), in their investigations of the pulvinar, a higher order thalamic relay nucleus, lend support to conceiving of the burst mode of firing as a form of gain control. They also mention a significant output connection from the pulvinar to the striatum of the basal ganglia structures. A similar connection between a different relay nucleus e the ventrobasal nuclei, conveying somatosensory information e and the basal ganglia striatum is noted by Llina´s, Lesnik, and Urbano (2002). Specifically, this thalamic relay projection occurs in conjunction with collateral projections from layer 5/6 of the cortex. Groenewegen (2003) discusses co-occurring thalamic and cortical

e3.32 PART | V Synthesis

inputs to the medium-sized densely spiny neurons of the striatum, which function as a gate via coincident detection, and without which the spiny neuron would not become active. This type of coincident convergence of inputs from the thalamus and cortex is also noted by McFarland and Haber (2000). Such co-occurring inputs are schematically portrayed at the bottom of Fig. 19.4 above. It is worth noting that this kind of AND-gate could serve the role of a contingent decision node in Powers’ (1973) conception of how program perceptions operate. This temporal co-occurrence serves as a form of addressing, where if a related input from cortical layer 5/6 takes place at the same time as a relay input, then the striatum cell will fire. This is particularly the case if the relay cell is firing in burst mode. As noted above, when there is not much difference between a reference input and its corresponding perceptual input, the difference between them may not amount to a very substantial error signal, despite its potentially being significant for other reasons. Burst firing offers a way around that dilemma, by heightening the gain and detectability of such a signal. It has been found that thalamic inputs in their burst mode of firing convey a potent activation of a cortical column of cells (Swadlow & Gusev, 2001; Swadlow et al., 2002). This may be one mechanism by which the cortex provides an ongoing reference signal to the subcortical TRN and its inverted projection to a thalamic relay nucleus. If cortico-thalamic input is not sustained, there may not be sufficient time to bring the corresponding perception to its reference state, through the motor output structures of the brain, before the descending input to the thalamus changes. Thus, there may well be a need to keep reactivating the cortical reference signals. Obviously, thalamic relay output, even once it has become a PCT error signal, continues to have its original topographical connections with certain cortical columns. The burst mode of firing described above for the error signal may provide a nonlinear way to activate that cortical area. Larkum, Zhu, and Sakmann (1999) describe axonal and dendritic coincidence detection, from different cortical layers, facilitating supra-threshold firing of pyramidal neurons. Llina´s et al. (2002) propose that a particular type of coincidence detection e between what are called nonspecific thalamic inputs to cortical layer 1 and specific thalamic inputs through layer 4 e provides a mechanism for recursive cortical firing. They state among their findings: “Beyond coincidence detection, the powerful coactivation of specific and nonspecific inputs resulted in a temporally tuned return output to the thalamus and, thus, in the generation of recursive dynamic loop activity” (Llina´s et al. 2002, p. 449). This may provide a means for ongoing reference signals to be supplied. It is worth noting, among these same authors’ tracing of thalamic connections to the apical dendrites of layer 6 pyramidal cells, that both direct excitatory and indirect interneuron inhibitory connections are made through layer 4 to the same sets of dendrites. This may serve the function of canceling out the effect when the relay signal cycles out of burst mode firing and into

How the brain gets a roaring campfire Chapter | 19

e3.33

tonic mode. In other words, coincidence detection would take place for the relay signal in burst mode, resulting in reactivating the cortical column and renewing its calculation or addressing of a reference signal. But then when switched to tonic firing, the relay cell projections would remain sub-threshold and not complicate matters for the cortex, because of competing excitatory and inhibitory projections. Admittedly, this latter notion is a speculative part of the proposal, suggesting a role for those offsetting connections.

Summary: thalamic control loop signals Make it a roaring campfire The summary provided here focuses on the visual changes that comprise constituting perceptions for a roaring campfire. Obviously, there would be a series of body movements, to bring about the changes for the visual system. Those implementations would occur by actions taken in the environment, and their perceptual results would percolate back up as a hierarchy of visual perceptions, from the retina to the thalamus to various cortical sites. The focus of this summary, however, is on the visual mechanisms, not the movementrelated mechanisms. In exploring examples to envision these aspects in action, it is important to remember the notion of higher-order thalamic nuclei (Guillery, 1995; Sherman, 2007). The majority of thalamic relay nuclei are not dealing directly with sensory information from the periphery, but rather cortico-thalamic driving inputs from layer 5 of the cortex. And if the central proposal of this chapter is correct, there are corresponding reference inputs coming from layer 6 of the cortex as well, turning thalamic relay neurons into PCT comparators. This permits consideration of at least some higher level perceptions according to a PCT schema, which make use of higher order relay nuclei of the thalamus. To get a roaring campfire from a visual standpoint, essentially the brain wants more brightness, more contrast, and more motion. This is like a movie director calling for “Lights, camera, action.” One aspect of wood catching on fire to produce more brightness is that the retina and thalamus would register more On-zones from the change in light intensity. For instance, as On-center ganglion cells project to the thalamus, the corresponding zones of the linear Xcells in the Lateral Geniculate Nucleus (LGN) also go on. To the extent there is flickering in the flames, both On- and Off-zones would register in turn. In order to actively control for more brightness, the brain would send a reference instruction to those thalamic cells, specifying more firing. Here is one place where an excitatory reference input would be paired with an inhibitory perceptual input, through the mechanism of the dendritic triads discussed above. Those dendritic connections invert the effect of the ascending perceptual signal as it enters a thalamic relay cell acting as a PCT comparator. The error signal that results would eventually lead to an increase in that cell’s firing.

e3.34 PART | V Synthesis

Another aspect of more brightness is broader patches of light. This involves lower spatial frequencies designating larger areas of luminance. According to Glezer (1995, p. 42), spatial frequencies are formed in the brain by different degrees of overlap of On- and Off-cells, creating partially overlapping light and dark bars. Such thalamic cells project to cells of the striate cortex, which have been modeled with Gabor filters.7 A given image is constructed from first finding “the period of the first harmonic that is near the size of the image” (Glezer, 1995, p. 103). A module showing maximal output and containing those Gabor filter characteristics is found, and stored as the initial reference standard (p. 108). This may occur via the mechanism of auto-associative addressing of reference specifications, discussed in Chapter 6. The task for a brain wanting a broader campfire then becomes one of designating progressively lower spatial frequencies as the first harmonic fundamental, as a new set of goals for the corresponding striate cortex modules. These wider spatial frequencies would specify broader collections of On-cell inputs from thalamic neurons in the LGN, again setting the direction of subtraction in the comparator via the dendritic triads. An associated aspect of expanding these patches of light is repositioning their edges. This would involve two features, (a) the borders between textures, and (b) the high spatial frequencies outlining fine details of an image. Texture is a way of measuring similar output from adjoining areas, helping to differentiate a figure from its background. The brightly burning portion of a fire has a smoother visual texture than the irregularities of glowing clumps of charred wood. Thus, moving the borders outwards means a larger patch of campfire. Within a common texture, autocorrelation functions determine similarity, as mapped by oriented Gabor filters (Fogel & Sagi, 1989). Borders between textures are constructed as gradient differences, with derivatives of striate outputs projected to the larger RFs of the peristriate cortex. Thus, to send references for expanding patches of textured light is to seek to decrease the rate of change signals at the current edges. To the extent more complex perceptions are routed through higher order thalamic nuclei, via the corticothalamic-cortical projections discussed above, reducing such derivative signals means the new L6 cortical specifications are transmitted through the thalamic reticular nucleus (TRN), to invert the sign of the reference signal. In parallel with lowering the gradient of a common (flaming) texture is the matter of turning what are currently moving edges into the broader body of the flames. This means changing high spatial frequency details into low spatial frequency patches. This may involve registering such areas as sustained firings of neurons rather than transient ones, because there is less variation of 7. To recap material presented earlier in these chapters, a spatial frequency represents the number of black-white oscillations in a degree of visual space, measured in cycles per degree. A Gabor function shows this as sinusoidal cycles within a Gaussian envelope, with parameters defining width of the envelope, and period and phase of the cycle.

How the brain gets a roaring campfire Chapter | 19

e3.35

luminance from the middle of a fire than at its edges. Stationary versus drifting gratings have shown two broad pathways from retina to thalamus to visual cortex, with temporal features modeled by means of low-pass filters for semistatic images (called a parvocellular pathway), and bandpass filters for moving ones (called a magnocellular pathway) (Perrone & Thiele, 2002; Perrone, 2006). Such a distinction becomes more significant below, in summarizing how faster rates of motion are obtained. Here, it is simply worth noting that a change of reference, seemingly from the striate cortex back to the LGN, may control for and thus produce sustained firing rates from RFs denoting former edges of the flames. However, it is not sufficient to just get more sustained luminance from the center of a fire. A larger fire is not the same as an actively dancing and “roaring” fire. What is needed is not simply more brightness, but more motion. This means more transient signals of luminance onset and offset, along the magnocellular pathway mentioned above. Thalamic Y-cells are geared to detect sudden changes, registering peripheral areas of the retina, as well as the fast-Off retinal ganglion cells. To increase the firing of the thalamic neurons means subtracting any existing perceptual signal from an excitatory reference signal, and thus the afferent input would arrive via the dendritic triad mechanism. As will be discussed below, the reference is likely coming from the middle temporal (MT) cortex. The transient signals just discussed are part of a mechanism that constructs a rate-of-change perception, as a ratio between neurons demonstrating transient versus sustained responses. As discussed earlier, these temporal signals represent bandpass versus low-pass filtering, respectively, in the striate cortex. Their ratio is modeled as the cross-over point of these two signals, representing the relative speed of the oriented edge being measured. What this refers to is the visual oscillation at the edges of the campfire. A roaring campfire means more vertical and lateral motion for those edges. As Perrone (2005) has demonstrated, it is possible to adjust the measure of speed in the MT cortical neurons by scaling the transient output from the striate cortex by means of a sensitivity factor. To multiply the MT speed entails dividing the striate transient amount. When that is done with quantities in a log scale, that amounts to using subtraction. In other words, an inhibitory reference sent to a lower level can vary the speed tuning of the MT neuron. The PCT microscope suggests that the comparator for such a calculation operates by way of a thalamic relay nucleus. That is to say, a feedforward projection from striate layer 5 to a higher order thalamic relay is compared with a reference signal from the higher MT layer 6, which is routed through the TRN to make it inhibitory. This results in specifying a faster speed or rate of change from the transient neurons. Those transients are themselves measuring the onset motion of edges in a given orientation, which correlates with how the edge of a flame is moving. According to Perrone (2004, p. 1735), the transient neurons in the

e3.36 PART | V Synthesis

striate cortex manifest directional selectivity. Therefore, a flame moving back and forth, even if it maintained a fairly vertical orientation, would need at least two transient sensors to help compile the rate of that movement in either direction. This means that any reference from the MT cortex for altering the speed of that oscillation would need to be branched to reach both sets of sensors. To designate a faster rate of change must come from a slightly higher level than the MT locale that measures speed itself, which amounts to a change in the slope of the Weighted Intersection Mechanism modeled by Perrone (2004). When a given surface moves across the aperture center of a neuron’s RF, its apparent movement is orthogonal to the orientation of the surface’s edge. This can be visualized via a slanted stimulus such as a back-slash (\), as compared with L-shaped portions of a RF. Whether the stimulus is moving vertically or horizontally to the right, it would cause subsequent points in the RF lying along the L to go on at the same time. Thus, at that level it is impossible to detect the actual direction of movement. To all appearances, it is moving in a forward-slash (/) direction. This creates difficulties if it is part of a larger pattern moving at some other oblique angle, such as an ascending flame. It takes multiple sensors to measure the movements of the components as well as the overall movement of the pattern. This issue arises with the campfire, because the wedge-shaped points of the flame move up at a 90 direction while their apparent edges move closer to 45 or 135 directions. To increase the rate of change of the ascending pattern means sending references to increase the rate of change of the components, which constitute that pattern. This would be accomplished, so suggests the PCT microscope, via inverted MT component inputs being routed as perceptual features through a higher level relay nucleus of the thalamus, by way of dendritic triads, and compared with reference inputs coming from the MT pattern neurons. The polarity of the comparison would result in faster changes in those component edges, leading to a faster apparent ascent of the flickering flames. This raises the issue of the brain wanting more contrast, not just brightness and motion, in order to achieve a roaring fire. This includes wanting more points and angles among the moving MT pattern neurons. The ascending points and angles of the flames are comprised of sharper outlines from higher spatial frequencies. According to Glezer (1995), this is a template-matching process occurring in or from the inferotemporal cortex (ITC), as the initial sense from the low spatial frequencies is combined with an active search and specification of modules with the right higher spatial frequencies and the right kinds of contrast. Auto-associative addressing (see Chapter 6) in the cortex may be involved in selecting those “right” modules. The notion of “form invariance” discussed in Chapter 7 involved a continuing perceptual sense, for instance of a flame, despite other transformations such as motion or brightness. The reference for getting more flame perceptions would act as a template against which composite perceptions from

How the brain gets a roaring campfire Chapter | 19

e3.37

MT pattern neurons or elsewhere could be compared. A mismatch, computed in a high order thalamic nucleus, would lead to negative feedback actions to bring those perceptions closer to their specifications. Glezer (1995) notes, “the same spatial frequency can have different meanings, because in different modules it represents different harmonics” (p. 112). This means it is important to anchor each perceptual reference with the right spatial frequency fundamental before the harmonics are determined. The task could be conceived as follows: (a) first find the module or perceptual function that resonates with the most clarity, matching what is currently going on; (b) then start to change the module with a reference toward a preferred perceptual state. This essentially sets up a pursuit tracking task, where the reference specification is a moving target. For instance, if the center of heat of a campfire is called zone A, immediately above that in zone B would be steady flames with a fairly wide angle, while further up in zone C would be taller and more narrow occasional tongues of fire. So the visual pursuit tracking task for a more vigorous campfire is to move zone B higher up, using spatial frequency fundamentals that begin higher than the main body of the campfire but not as high as the narrow flames in zone C. A similar pursuit tracking task takes place in zone C, starting with even higher fundamentals in terms of the width of the existing flames. As the campfire gets fed with fuel, visually zones B and C are tracking from their existing fundamental spatial frequencies into modules with lower fundamentals. Within the flame perceptions themselves, any given edge is formed by compiled harmonics, at whatever orientation a moving flame might generate. So there may need to be multiple reference specifications to cover those various orientations, establishing the fundamental spatial frequency for each sub-image of the dancing flames. With the fundamental set, other neurons within a given module measure the contrast differentials at harmonic intervals. More contrast means starker edges with a higher net amplitude difference between adjoining areas. This amounts to squaring off the gradations within the spatial frequencies, as the odd harmonic channels from lower levels are superimposed on one another in the area of the ITC. To sum up, more brightness means: more On-zones, broader patches, lower spatial frequencies, repositioned edges, reduced gradients, and more sustained firings. More motion means: more transient signals, edge oscillation, MT speed tuning, scaling transient neurons, slope of MT component cells, and rate-of-change of MT pattern cells. More contrast means: point or angle templates, anchored spatial frequency fundamentals, pursuit tracking to lower fundamentals, and squaring off harmonic gradations. A roaring campfire is quite busy indeed.

e3.38 PART | V Synthesis

A final survey through that PCT microscope A case has been made in this chapter that the thalamus is one site in the brain where there is an array of cells functioning as PCT style comparators. Initial consideration of this proposal followed the lead of a sign-reversal as cortical output made its way through the thalamic reticular nucleus to project onto thalamic relay cells. The import of each type of input onto such cells has been traced, with PCT in mind, to see if plausible functions could be ascertained. A method was discerned for those cells to supply bi-directional control, by way of dendritic triads compensating for the neural constraint of half-wave rectification. The details of tonic versus burst firing were also examined, with the seeming result that burst firing could amplify the gain when the error-signal output was negligible. While relay nuclei of the thalamus are generally considered to be afferent structures for ascending inputs, over 90% of geniculate inputs actually derive from non-peripheral sources (Sherman, 2006). The connections reviewed in this chapter give prominence to what seem to be descending circuitry that comprises cortical outputs. Several lines of evidence point to the basal ganglia striatum as a key target of that descending thalamic relay projection. This chapter has argued that it is a PCT error signal being relayed to the striatum, signifying how much more correction is needed for the driver perceptual input entering the thalamic relay cell. Because the partially-corrected absolute value of that error signal may be relatively small, in its tonic form of firing, the signalto-noise ratio may benefit from the burst mode of relay cell firing, heightening gain and detectability in the striatum. Indeed, direct collaterals from layer 5 of the cortex onto spiny neurons of the striatum may provide a form of coincident addressing in the basal ganglia in concert with burst firing from thalamic relay cells. The burst firing may serve a different purpose when it is directed along the original topographical relay channel ascending to the cortex. Evidence suggests that burst firing from the thalamus can potently activate its corresponding cortical column, and this chapter proposes such a mechanism for recursive re-addressing of an ongoing reference signal from that cortical area. This may especially be the case when specific thalamic inputs through cortical layer 4 co-occur with nonspecific thalamic inputs distally through cortical layer 1. When a reference signal is coming from that cortical column, recycling an error signal back up to the cortex could become problematic when the thalamic relay cell is back in tonic mode. In this instance, layer 4 projections within the column that are directly excitatory and indirectly inhibitory may serve to cancel out the effect, by keeping the net signal sub-threshold. These proposals about how the thalamus may function are consistent with the neuroanatomy and neurophysiology as reported in the research literature. Their unique aspects come from using Perceptual Control Theory as a microscope, as it were, to make sense of some of the complexity and bring

How the brain gets a roaring campfire Chapter | 19

e3.39

clearer focus to certain puzzling features in the literature. A great many physiological details have been included in this survey. Such was necessary to try to demonstrate that the PCT concepts would still work as proposed. In the process, it has emerged that the so-called relay areas of the thalamus may well operate in a dual manner. They convey and even shape some of the sensory input on its way to the cortex. But they also sit at a juncture point where mid-level control systems are regulating the state of their perceptions. The point of intersection is the role ascribed to comparators, within a PCT analysis. As such, these thalamic relay areas appear to send negative feedback output on its way from the cortex to lower implementing levels of the nervous system. Granted, the mechanisms are a great deal more complicated than the bare-bones requirement of subtraction for a PCT wiring diagram. Yet, no control theory principles have been violated, and indeed a site has been identified for heightening error-signal gain in a control loop, by means of the burst mode of firing of thalamic neurons. The surprisingly varied forms of connection onto thalamic relay cells have each been found to have a meaningful interpretation and potential role to play for the overall operation of these structures. It remains to be seen whether these proposed functions are indeed what is happening within the complicated neural wiring of the thalamus. The hope here is that these formulations would be examined and tested, as further research into the thalamus is designed. If this PCT microscope can stimulate a deeper appreciation and understanding of neural functioning, then its purpose will have been achieved.

Acknowledgments I wish to thank Dominique Berule and Warren Mansell for their wise counsel and guidance in framing and improving the presentation of material in this chapter.

References Bettler, B., & Kaupmann, K. (2001). Structure of GABA-B receptors. In H. Mohler (Ed.), Pharmacology of GABA and Glycine Neurotransmission (pp. 299e309). Berlin: SpringerVerlag. Blitz, D. M., & Regehr, W. G. (March 24, 2005). Timing and specificity of feed-forward inhibition within the LGN. Neuron, 45(6), 917e928. Cox, C. L., Zhou, Q., & Sherman, S. M. (July 30, 1998). Glutamate locally activates dendritic outputs of thalamic interneurons. Nature, 394, 478e482. Crabtree, J. W. (1992). The somatotopic organization within the rabbit’s thalamic reticular nucleus. European Journal of Neuroscience, 4(12), 1343e1351. Crandall, S. R., & Cox, C. L. (July 15, 2013). Thalamic microcircuits: presynaptic dendrites form two feedforward inhibitory pathways in thalamus. Journal of Neurophysiology, 110(2), 470e480.

e3.40 PART | V Synthesis Destexhe, A., & Sejnowski, T. J. (2002). The initiation of bursts in thalamic neurons and the cortical control of thalamic sensitivity. Philosophical Transactions of the Royal Society of London - B, 357, 1649e1657. Fogel, I., & Sagi, D. (1989). Gabor filters as texture discriminator. Biological Cybernetics, 61, 103e113. George, D. (2008). How the Brain Might Work: A Hierarchical and Temporal Model for Learning and Recognition. Doctoral Dissertation. Stanford University. Available at: http://alpha.tmit. bme.hu/speech/docs/education/02_DileepThesis.pdf. Glezer, V. (1995). Vision and Mind: Modeling Mental Functions. Mahwah, NJ: Lawrence Erlbaum Associates. Godwin, D. W., Vaughan, J. W., & Sherman, S. M. (September 1996). Metabotropic glutamate receptors switch visual response mode of lateral geniculate nucleus cells from burst to tonic. Journal of Neurophysiology, 76(3), 1800e1816. Govindaiah, G., & Cox, C. L. (December 27, 2006). Metabotropic glutamate receptors differentially regulate GABAergic inhibition in thalamus. Journal of Neuroscience, 26(52), 13443e13453. Groenewegen, H. J. (2003). The basal ganglia and motor control. Neural Plasticity, 10(1e2), 107e120. Grubb, M. S., & Thompson, I. D. (June 1, 2005). Visual response properties of burst and tonic firing in the mouse dorsal lateral geniculate nucleus. Journal of Neurophysiology, 93(6), 3224e3247. Guillery, R. W. (1995). Anatomical evidence concerning the role of the thalamus in corticocortical communication: a brief review. Journal of Anatomy, 187(Pt.3), 583e592. Available at: http:// www.ncbi.nlm.nih.gov/pmc/articles/PMC1167461/pdf/janat00131-0069.pdf. Hawkins, J., & Blakeslee, S. (2004). On Intelligence. New York: Henry Holt and Company. Hill, S., & Tononi, G. (2003). Thalamus. In M. N. Arbib (Ed.), The Handbook of Brain Theory and Neural Networks (pp. 1176e1180). Cambridge, MA: MIT Press. Huguenard, J. R. (1998). Anatomical and physiological considerations in thalamic rhythm generation. Joural of Sleep Research, 7(Suppl. 1), 24e29. Jones, E. G. (1998). Chapter 1: the thalamus of primates. Handbook of Chemical Neuroanatomy, 14, 1e298. The Primate Nervous System, Part II. Koch, C. (November 1987). The action of the corticofugal pathway on sensory thalamic nuclei: a hypothesis. Neuroscience, 23(2), 399e406. Lam, Y.-W., & Sherman, S. M. (January 2010). Functional organization of the somatosensory cortical layer 6 feedback to the thalamus. Cerebral Cortex, 20, 13e24. Larkum, M. E., Zhu, J. J., & Sakmann, B. (March 25, 1999). A new cellular mechanism for coupling inputs arriving at different cortical layers. Nature, 398, 338e341. Lee, S.-H., Govindaiah, G., & Cox, C. L. (July 2007). Heterogeneity of firing properties among rat thalamic reticular nucleus neurons. The Journal of Physiology, 582(1), 195e208. Llina´s, R. R., Lesnik, E., & Urbano, F. J. (2002). Temporal binding via cortical coincidence detection of specific and nonspecific thalamocortical inputs: a voltage-dependent dye-imaging study in mouse brain slices. Proceedings of the National Academy of Sciences of the United States of America, 99(1), 449e454. Available at: http://www.pnas.org/content/99/1/449.full. McCormick, D. A., & Feeser, H. R. (1990). Functional implications of burst firing and single spike activity in lateral geniculate relay neurons. Neuroscience, 39(1), 103e113. McFarland, N. R., & Haber, S. N. (May 15, 2000). Convergent inputs from thalamic motor nuclei and frontal cortical areas to the dorsal striatum in the primate. Journal of Neuroscience, 20(10), 3798e3813.

How the brain gets a roaring campfire Chapter | 19

e3.41

Ohara, P. T., & Lieberman, A. R. (1981). Thalamic reticular nucleus: anatomical evidence that cortico-reticular axons establish monosynaptic contact with reticulo-geniculate projection cells. Brain Research, 207(1), 153e156. Ohara, P. T., & Lieberman, A. R. (1985). The thalamic reticular nucleus of the adult rat: experimental anatomical studies. Journal of Neurocytology, 14(3), 365e411. Perrone, J. A. (2004). A visual motion sensor based on the properties of the V1 and MT neurons. Vision Research, 44, 1733e1755. Perrone, J. A. (2005). Economy of scale: a motion sensor with variable speed tuning. Journal of Vision, 5, 28e33. Perrone, J. A. (November 15, 2006). A single mechanism can explain the speed tuning properties of MT and V1 complex neurons. Journal of Neuroscience, 26(46), 11987e11991. Perrone, J. A., & Thiele, A. (2002). A model of speed tuning in MT neurons. Vision Research, 42, 1035e1051. Powers, W. T. (1973). Behavior: The Control of Perception. Chicago: Aldine Publishing. Powers, W. T. (2008). Living Control Systems III: The Fact of Control. Bloomfield, NJ: Benchmark Publications. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960a). A general feedback theory of human behavior: Part 1. Perceptual & Motor Skills, 11, 71e88. Powers, W. T., Clark, R. K., & McFarland, R. L. (1960b). A general feedback theory of human behavior: Part 2. Perceptual & Motor Skills, 11, 309e323. Ramcharan, E. J., Gnadt, J. W., & Sherman, S. M. (2005). Higher-order thalamic relays burst more than first-order relays. Proceedings of the National Academy of Sciences of the United States of America, 102(34), 12236e12241. Schroedinger, E. (1967). What Is Life? with Mind and Matter & Autobiographical Sketches. New York: Cambridge University Press. Sherman, S. M. (1996). Dual response modes in lateral geniculate neurons: mechanisms and functions. Visual Neuroscience, 13, 205e213. Sherman, S. M. (November 2004). Interneurons and the triadic circuitry of the thalamus. Trends in Neurosciences, 27(11), 670e675. Sherman, S. M. (2006). Thalamus. Scholarpedia, 1(9), 1583. Available at: http://www. scholarpedia.org/article/Thalamus_anatomy. (Accessed 29 January 2014). Sherman, S. M. (2007). The thalamus is more than just a relay. Current Opinion in Neurobiology, 17(4), 417e422. Sherman, S. M. (2013). The function of metabotropic glutamate receptors in thalamus and cortex. The Neuroscientist, 20(2), 136e149. Sherman, S. M., & Guillery, R. W. (1998). On the actions that one nerve cell can have on another: distinguishing “drivers” from “modulators”. Proceedings of the National Academy of Sciences of the United States of America, 95(12), 7121e7126. Neurobiology. Available at: http://www. pnas.org/content/95/12/7121.full.pdf. Sillito, A. M., & Kemp, J. A. (October 24, 1983). The influence of GABAergic inhibitory processes on the receptive field structure of X and Y cells in cat dorsal lateral geniculate nucleus (dLGN). Brain Research, 277(1), 63e77. Softky, W. (1996). Modeling Thalamus as a Non-rectifying Predictive Comparator [Unpublished manuscript, June 1996]. Available at: http://www.rni.org/softky/thalamus.pdf. (Accessed 11 February 2014). Swadlow, H. A., & Gusev, A. G. (2001). The impact of ‘bursting’ thalamic impulses at a neocortical synapse. Nature Neuroscience, 4(4), 402e408.

e3.42 PART | V Synthesis Swadlow, H. A., Gusev, A. G., & Bezdudnaya, T. (September 1, 2002). Activation of a cortical column by a thalamocortical impulse. Journal of Neuroscience, 22(17), 7766e7773. Wei, H., Bonjean, M., Petry, H. M., Sejnowski, T. J., & Bickford, M. E. (November 23, 2011). Thalamic burst firing propensity: a comparison of the dorsal lateral geniculate and pulvinar nuclei in the tree shrew. Journal of Neuroscience, 31(47), 17287e17299. Wikipedia contributors. (September 14, 2015). Membrane potential. In Wikipedia, the Free Encyclopedia. Retrieved from https://en.wikipedia.org/w/index.php?title¼Membrane_ potential&oldid¼681014986. Yin, H. H. (2014). How basal ganglia outputs generate behavior. Advances In Neuroscience, 28. Article ID 768313.

Appendix 1

Key websites for further reading on perceptual control theory In order to access wider reading on PCT, many websites are available. Three of these are particularly broad in scope and provide links to the wider network of resources: www.livingcontrolsystems.com, managed by Dag Forssell, which includes a range of unpublished papers by William T. Powers, along with his computer demonstrations, and a freely downloadable Book of Readings from a wide range of authors across disciplines. www.pctweb.org, which includes a range of introductions of PCT, and links to diverse academic papers, books and videos, that have utilised PCT across a wide range of disciplines. www.iapct.org, which is the website for the International Association of Perceptual Control Theory. In particular it includes a page of links to the many additional websites that host papers, computer demonstrations and videos on PCT: http://www.iapct.org/websites.html.

631

Appendix 2

Leading figures in perceptual control theory A note from the editor Whilst I endeavored to ensure that the strongest contemporary work on PCT is represented in this volume, there are a number of individuals who are not represented here who have utilized PCT for many years, and for the sake of completeness, deserve recognition outside the chapters of the book. This appendix provides a brief overview of these individuals and their key publications. It replicates and extends a summary from http://www.pctweb.org/lead/ leading.html Bruce Abbott Bruce Abbott is an influential experimental psychologist at Indiana-Purdue University with a keen interest in PCT. He worked with Bill Powers to develop the computer simulations within Living Control Systems III using object-oriented programming. Tom Bourbon Tom Bourbon trained in experimental psychology, and he applied PCT to human factors, brain function and education (with Tim Carey), as well as devising a number of studies of human tracking based on PCT that investigated the role of ‘helping’ and the individual specificity of perceptual control. Tim Carey Tim Carey is a professor at the Centre for Remote Health, Flinders University, Australia. He developed Method of Levels (MOL) therapy, closely supported Bill Powers (see www.methodoflevels.com.au). He leads the dissemination and evaluation of MOL, as well as applying PCT to research methodology, patientperspective health care, and school systems.

633

634 Appendix 2

Gary Cziko Gary Cziko is professor emeritus of educational psychology at the University of Illinois. He has written two of the most accessible introductions to Perceptual Control Theory e, Chapter 8 in Without Miracles, and The Things We Do. Philip Farrell Following his work with Martin Taylor, Dr Farrell has utilised PCT within his work for the Canadian National Defence Headquarters. His work focuses on modelling Team Information Processing using PCT to help organisations interpret and internalise instructions and reach common goals. Ed Ford Ed Ford leads an organization to train educators and parents in using the Responsible Thinking Process (RTP), based on PCT, to assist the young in conducting themselves more effectively, relating better with others, and becoming more responsible. See responsiblethinking.com. Dag Forssell Dag Forssell is a mechanical engineer and manager who has been responsible for video-archiving PCT conference since the 1990s, in addition to publishing a range of PCT books via Living Control Systems Publishing. His large catalogue of work is available at www.livingcontrolsystems.com. Perry and Fred Good Perry Good is a popular speaker, trainer, corporate coach, and author. She co-authored A Connected School, based on PCT, with Shelley Roy. Fred Good founded NewView Publications to publish and disseminate books, seminars and workshops on PCT. Wayne Hershberger Wayne Hershberger is the editor of the volume Volitional Action: Conation and Control, which brings together a wide range of theoretical and empirical papers on PCT alongside studies of the neuroscience of volitional action. He has also utilized PCT in work on animal locomotion, learning, and neuroscience. Fred Nickols Fred Nickols is the director of Distance Consulting, who utilize PCT in a range of management consultancy operations as well as widely disseminating PCT (https://nickols.us/controltheory.html).

Appendix 2

635

Richard Pfau Richard Pfau is a psychologist and human resource specialist with a background in engineering who recently authored a practical guide to selfmanagement based on PCT - Your Behavior. Mary Powers Mary Powers, the late wife of Bill Powers, was a person-centred counsellor. She wrote a number of key articles on PCT, and detailed illustrations of PCT hierarchies (see http://www.livingcontrolsystems.com/intro_papers/mary_pct.html). Richard Robertson Dick Robertson is former professor at Northeastern Illinois University where he taught control theory from the 1970s. He was involved in PCT since its inception in the late 1950s. He is also a clinical psychologist and rehabilitation researcher. He has been a member of the Control Systems Group since its inception and is the author of several significant papers on PCT. He worked alongside Bill Powers to produce the first student textbook of PCT: An Introduction to Modern Psychology: The Control Theory View. Shelley Roy Shelly Roy is a recognised leader in human resources development based on PCT, consulted on a regional, state, national, and international basis on leadership, human behavior, change and transition, effective communication, models of teaching, interagency collaboration, effective schools research, generational diversity and other areas related to change and building productive teams. Phil Runkel Phil Runkel was the Professor Emeritus of Psychology and Education at the University of Oregon. Following long discussions with Bill Powers, he produced a reinterpretation of human psychology using PCT in the book People as Living Things: The Psychology of Perceptual Control. He is also the author of Casting Nets and Testing Specimens: Two Grand Methods of Psychology. This book critiques in detail the existing methodologies within psychology and points to the advantages of two alternative methodologies, which are both consistent with PCT. Sara Tai Sara Tai is a senior lecturer and consultant clinical psychologist at University of Manchester, UK. Since 2006, she has worked on the dissemination and evaluation of MOL across a wide range of contexts, including psychiatric inpatient wards, schools, and outpatient primary and secondary mental health services.

636 Appendix 2

Further reading Carey, T. A. (2012). Control in the classroom: An adventure in learning and achievement. Living Control Systems Publications. Carey, T. A. (2017). Patient-perspective Care: A New paradigm for health systems and services. London: Routledge. Cziko, G. (1997). Without miracles: Universal selection theory and the second Darwinian revolution. MIT press. Cziko, G. (2000). The things we do: Using the lessons of Bernard and Darwin to understand the what, how, and why of our behavior. MIT press. Good, E. P., Grumley, J., & Roy, S. (2003). A connected school. Chapel Hill, NC: NewView Publications. Hershberger, W. A. (1989). Volitional action: Conation and control. Amsterdam, The Netherlands: North-Holland. Pfau, R. (2017). Your behavior: Understanding and changing the things you do. Paragon House. Robertson, R. J., & Powers, W. T. (1990). Introduction to modern psychology: The control-theory view. Control Systems Group. Runkel, P. J. (1990). Casting nets and testing specimens: Two grand methods of psychology. New York: Praeger. Runkel, P. J. (2003). People as living things: The psychology of perceptual control. Living Control Systems Publishing.

Index Note: ‘Page numbers followed by “f ” indicate figures and “t” indicates tables’.

A Acceptability judgments, 434 Accessibility, 510, 512 Adaptive Cascade Model, 160, 165 Agent based modeling (ABM), 562e563, 562f Age-related regression periods, 203e204 Analysis-by-synthesis, 573 Animal contests controlled variable, 76 lessons from fighting, 76e86 problems needing novel solutions, 87e93 thoughts on interpretation, 84e86 thoughts on methodology, 78e84 Anterior forebrain pathway (AFP), 215e216 Artificial cerebellum, 609, 615, 618 Artificial cognitive systems, 560e562 Artificial Intelligence (AI), 27, 423, 512, 518, 526, 552, 557e559, 587f, 599, 606e607 Artificial Life (AL) approach, 518e519 Associative memory and organization of brain, 249e254 Atenfels and facilitation of feedback paths, 238e239 and human interaction, 248e255 associative memory and organization of brain, 249e254 meaning, symbols, language, and culture, 254e255 symbols and meaning from PCT perspective, 248e249 to perceptions, 236e238 Auto-associative method, 133, 146 Autobiographical memory, 600, 620e623 Automatic behavior, 598

B Band-pass temporal filter, 151e152 Bayesian conditional probability equations, 133

Behavior, 4e5, 11, 19, 23e24, 26, 32e34, 42e45, 201e202, 210, 214, 463e464, 588, 601e602 Behavior-based robotics, 546e548, 559e560 Behaviorism, 24e25, 353e354, 437, 467e468, 487 Bi-directional comparators, 146 Blended frames, 177e180

C Canonical babbling, 373e375, 431e432 Carver & Scheier’s self-regulation theory, 593e594 Category perceptions, 112, 365, 430 Causal discovery methods, 64e68 Causal inference, 50e51, 64, 66e67, 69 Circular causation, 29e30, 355e356, 356f Classic automatic control, 565e567, 566f Classifier words, 388e390, 405, 412, 414e415 Closed-loop negative feedback process, 9e10 ‘Cognitive’ control processes, 216e217 Cognitive psychology, 353e356, 353f, 436e440, 443e444 “Cognitive revolution”, 24e25 Collective complex environmental variables (CCEV’s), 243, 260 Collective control, 366e367 Collective control hyper-network, 261, 261f, 263f, 264, 279 Collective control networks, 255, 257e258, 265e271 and social groups multiple overlapping, 257e264 scale and stability of, 256e257 in social structures, 265e268 Collective control processes cooperation and conflict, 241e242 giant virtual controllers, 243 Common knowledge, 402e405, 412e415, 441e442

637

638 Index Comparator function, 31, 146 Compatible cortical mind-sets, 128e141 Complex Environmental Variable (CEV), 234e236, 238e240, 243, 245, 312e313 Compromise CV constant, 84 Computationalism, 351, 439e441 Computational modeling, 168, 466e467, 481e485, 488e489 Computer languages, 441e442 Conceptualization, 105, 131, 135, 178e179, 489e490 Configuration, 84, 93, 111, 123, 129 Confirmation bias, 360, 389, 444 Conflict, 591, 604 with assumptions, 512 Connectionism, 439e441 Consciousness, 443e444, 590, 599, 604 Contemporary intelligent systems, 563e564 Controlled behavior, 598 Controlled variable (CV), 11, 76, 354e355, 589, 602e603 in protocol, 336e338 testing for, 16e20 Control systems, 9, 53e64 Control Systems in Organizations, 467e468 Control theory, 9e14, 470e471, 594 reverse engineering a robot, 14e16 reverse engineering from forward engineering perspective, 11e12 testing for controlled variables, 16e20 Control-theory analysis tool kit atenfels, 234e236 atenfels and facilitation of feedback paths, 238e239 matching atenfels to perceptions, 236e238 mirror world, 239e240 “Control theory glasses”, 233, 243e244, 257 Convenient fiction, 119 Conversational analysis, 511e512 Conway’s model, 600 Co-occurrences, 119e120, 139 Cooperative interaction, 299 control and perceptual control, 302e312 classes of protocol, 311e312 elements of control, 303e304 perception of uncertainty, 307e308 powers hierarchy of control, 304e307 protocols, 308e311

feedback loops and control loops atenfels, 314e316 four-element loops, 316e317 generalized feedback loops, 312e314 layered protocol theory (LPT), 301e302 protocol representation error correction, 324e325 general protocol grammar, 320e324 protocol example, 318e320 representing problems, 325 protocols proper, 326e346 controlled variable in protocol, 336e338 protocol as communication, 344e346 protocol function, 327e336 protocol levels, 341e344 triadic protocols, 338e341 Correction normal feedback, 335 Cortical implementation, 114e118 Countering disturbances, 504e505

D Deliberative architecture, 559 Demonstration, 330, 354, 399, 590, 603e604 Dependent variable (DV), 229, 353 Discourse analysis, 407te409t, 411 Discourse coherence, 402e404 Discrete behavior patterns, 43, 201 Dissemination, 510, 592e597 Distributed Adaptive Control (DAC), 561, 561f Distributional constraints, 426 Divergence, 592e597 Dynamic stability, 105

E Ecological approach, PCT, 597e598 “Effective stimulus”, 33e34, 36 Elementary control unit (ECU), 234e235, 235f, 303, 314e315 Elements of control, 303e304 ELIZA, 512 Embodied cognition, 597e598 Emergent reliability, 158e159, 162e163, 180 Empirical linguistics, 429e436 Enacted reliability, 158e159, 180 Entropy, 421e424 Environmental stabilization, 302 Epistemological considerations, 353, 442e447 Error correction, 324e325, 324f

Index Eshkol Wachmann Movement Notation (EWMN), 78e81, 79f Event coding, 598 Expansion, 186, 341, 372

F Faithfulness axiom, robust violations of, 49 causal discovery methods, 64e68 causal inference, 50e51 control systems, 53e64 digression on disturbances, 60 fundamental problem, 68e69 sufficient conditions for zero correlation, 69e71 zero correlation between variable and its derivative, 51e53 “Feedback function”, 234 Feedback loops and control loops atenfels, 314e316 four-element loops, 316e317 generalized feedback loops, 312e314 Feedforward mechanism, 610e611 Figure-ground contrasts, 162e165 ‘Fixed Action Patterns’ (FAPs), 213e214 Formal organizations, 265, 273 Formants, 381, 382f, 384e385 Form invariance, 172e173, 187, 188te189t Forward engineering, 9, 11e12, 20 Forward models, 152, 611e612 Fourier analysis model, 171e172 Free energy principle, 598e599 Fuzzy set of words, 387, 418

G Gabor functions, 163 Gaussian function, 163, 171, 178 ‘General linear model’, 4, 45, 489 General Protocol Grammar (GPG), 301, 318, 320e324 Generative linguistics, 351, 436e439 Giant virtual controllers (GVC’s), 267, 275, 291 GOFAI predecessors, 562 Good Old-Fashioned Artificial Intelligence’ (GOFAI), 559 Gooing, 371, 373e374 Grasping, 45, 131, 374, 597e598

639

H Hard sciences, 4, 597 Harmonic composition, 172e173 Hawkins’ model, 115e116, 118 Hierarchical control, 533e538, 534f Hierarchical organization, 30, 216e217 Hierarchical perceptual control, 567e569, 567fe568f Hierarchical Perceptual Control Theory (HPCT), 304, 529 Hierarchical temporal memory (HTM), 102e103, 108, 115, 128, 146 functional chunks, 121t Higher-level perception, 30, 240, 251e252, 254, 265e266, 278, 359 High loop gain, 29, 39e40, 238 Human activities collective control and levels of perception, 244e245 to physical artifacts, 243e244 structural frameworks for dyadic interactions, 245e247 Human cognition, 217e218 Human interaction atenfels and, 248e255 associative memory and organization of brain, 249e254 meaning, symbols, language, and culture, 254e255 symbols and meaning from PCT perspective, 248e249 Hyper-network of collective control networks, 258e259, 259f, 264e265 Hypothesized control systems, 15e16, 19, 44, 92

I Imagination, 357e360, 358f, 360f Imagination mode, 619 awareness, 619 Independent variable (IV), 353 Industrial-organizational psychology, 464e487 backlash, 471e472 computational modeling, 481e485 conceptual papers, 469e470 conceptual work, 472e473, 476e477 discontents, 464e466 emergence, 468e472 empirical work, 470e471, 473e475, 477e479

640 Index Industrial-organizational psychology (Continued ) established views, 464e466 history, 467e487 organizational psychology, 464e467 organizational research methods (ORM), 482 permeation challenge, 488e490 self-efficacy studies, 479e480 self-regulation, 490e491 ‘Test-Operate-Test-Exit’ model (T.O.T.E.), 465e466 thinking and learning, 485e487 Inferotemporal cortex (ITC), 112e113, 171e172, 185, 191 Inform normal feedback, 335 Innate Releasing Mechanisms (IRMs), 213 Input functions, 616e617 visual, 154e173 “Insufficiency” principle, 26 Intelligent communicative systems agent based modeling (ABM), 562e563 analysis-by-synthesis, 573 artificial cognitive systems, 560e562 behavior-based robotics, 559e560 classic automatic control, 565e567, 566f contemporary intelligent systems, 563e564 ‘deliberative’ architecture, 559 Distributed Adaptive Control (DAC), 561f GOFAI predecessors, 562 Good Old-Fashioned Artificial Intelligence’ (GOFAI), 559 hierarchical perceptual control, 567e569, 567fe568f iterative search, 574 ‘Mutual Beliefs, Desires, Intentions, and Consequences’ (MBDIAC), 571 optimization, 574 (H)PCT, 571, 576 ‘PREdictive SENsorimotor Control and Emulation (PRESENCE), 569e570, 570f robot, 558 synthesis-by-analysis-by-synthesis, 573 Intensity, 111, 155 Interactive cognitive subsystems (ICS), 599e600 Internal modeling, 593, 617e619 Intra-group categories, 200 Intrinsic error, 371 Intrinsic systems, 614e615

Introversion, 376 Inverse models, 611e612 Iterative search, 574

L Language acceptability judgments, 434 canonical babbling, 373, 375 changes, 418e421 cognitive psychology, 353e354, 353f, 436e439 collective control, 366e367 common knowledge, 402e404 computationalism, 439e441 computer languages, 441e442 connectionism, 439e441 consequences, 429e447 controlled variable (CV), 354e355 dependent variable (DV), 353 discourse analysis, 407te409t discourse coherence, 402e404 empirical linguistics, 429e436 entropy, 424 epistemological considerations, 442e447 expansion, 372 faculty, 369 generative linguistics, 436e439 gooing, 371 imagination, 357e360, 358f, 360f independent variable (IV), 353 information, 404e411 latency period, 435e436 linear order, 395e396 linguistic information, 400e418, 401f, 421e429 mainstream, 436e442 memory, 361e363 metalinguistic sameness, 434e435 method of levels (MoL), 446 methodological summary, 368e369 nonverbal perceptions, 415e418 objective information, 390 operator-argument dependencies, 392e394 Operator Grammar, 390e400 perceptions, 369e390, 370t, 372t perceptual hierarchy levels, 363e366, 365f category, 364 configuration, 363 event, 364 intensity, 363 principle, 364 program, 364

Index relationship, 364 sensation, 363 sequence, 364 system concept, 364 transition, 363 phonation, 371 phonemic distinctions, 378e383 pronunciation, 383e386, 385f reductions, 396e400, 397t repetition, 390 robots, 441e442 scientific method, 352 selection and likelihood differences, 394 subjective meanings, 415e418 subject-matter domain knowledge, 390 sublanguage, 411e415 ‘successive-state analysis’, 355e356 ‘Test-Operate-Test-Exit’ scheme (T.O.T.E.), 355e356, 465e466 Theory of Mind (ToM), 374 variation, 418e421 words, 376 dependencies, 387e389 morphemes, 386e387 order, 389 selection, 389 shapes, 389 Latency period, 435e436, 438 Lateral geniculate nucleus (LGN), 154, 161, 191 Layered protocol theory (LPT), 246, 301e302, 318, 331 ‘Leaping Hurdles’, 212 Likelihood constraint, 426e429 Linear order, 395e396 Linguistic information, 400e418, 401f, 421e429 Lower-level collective control network, 262f

M Mainstream, 436e442 Manage Your Life Online (MYLO), 512 McGurk effect, 360, 386, 400 Memory, 361e363 location, 362 Memory-based prediction, 545 Metalinguistic sameness, 434e435 Method of Levels Therapy (MOL), 446 accessibility, 512 conflict with assumptions, 512 countering disturbances, 504e505 evidence, 509e510

641

history, 506 learning, 508 negative feedback, 504 reorganization, 505e506 scientific validity, 510e512 steps, 507e508 therapy manuals, 513 Middle-level collective control network, 262f Middle temporal (MT) cortex, 151 ‘Mirror’ effect, 374 “Mobile inverted pendulum” (MIP), 14 ‘Modal Action Patterns’ (MAPs), 213e214 Model-based approach, 522, 523f Model-based control, 523fe524f, 552 Model-based methodology, 525 Model-based predictive control, 544e545 Modeling behavior, 542e543 Modeling dynamics, 548e551 Motivated action theory (MAT), 477 “Motor hierarchy”, 175 Motor redundancy, 611e612 Multiple-goal pursuit model (MGPM), 489e490 Muscular mechanisms, 615 ‘Mutual Beliefs, Desires, Intentions, and Consequences’ (MBDIAC), 571

N Name cells, 115e117 NASREM telerobot control system, 541e542 Negative feedback, 29, 504 control, 5, 45 loop, 10, 10f Neural components, 179, 615 Neural processing, 180e185 Neuro-psychotherapy, 593 Neuroscience, 23 calculation problem and control, 27 control of input, 28e30 determined by neural output, 26 input/output analysis and behavioral illusion, 32e34 misunderstanding control, 30e32 observer’s bias, 32 paradigm and crisis, 24e26 proper study of behavior, 42e45 sensorimotor transformations, 37e42 Sherrington’s analysis of reflex, 36e37 Neutral normal feedback, 335 Nonverbal perceptions, 254, 415e418 Notice lateral sequences, 123e124 Null normal feedback, 335

642 Index

O Occupational psychology, 593e595 Off signals, 159e162 On signals, 159e162 Open-loop methods to study closed-loop features, 149e154 Open-loop system, 521f Operator-argument dependencies, 392e394 Operator Grammar, 351e352, 390e400 Optimality Theory, 402 Order-reduction representations, 590e591, 604 Organizational Behavior and Human Decision Processes (OBHDP), 468 Organizational rationale, 106e108 Organizational research methods (ORM), 482 Output signals integration, 618

P Parallel argument position, 398 Parent-infant interaction, 212e213 Parsimony plausibility, 416e417 Partial-ordering constraint, 426 Partnerwise orientation, 79 Perception-based robotics, 518e520 Perceptions, 369e390, 370t, 372t Perceptual control, 103e111 hierarchy, 589, 603 Perceptual hierarchy levels, 363e366, 365f category, 364 configuration, 363 event, 364 intensity, 363 principle, 364 program, 364 relationship, 364 sensation, 363 sequence, 364 system concept, 364 transition, 363 Personality psychology, 593e595 Phonation, 371, 373e374 Phonemic contrast constraint, 426 Phonemic distinctions, 372e373, 375, 378e383 Posterior parietal cortex (PPC), 175, 192 Powers hierarchy of control, 304e307 ‘PREdictive SENsorimotor Control and Emulation (PRESENCE), 569e570, 570f Pronunciation, 383e386, 385f

Protocols, 308e311 classes of, 311e312 as communication, 344e346 controlled variable in, 336e338 example, 318e320 function control of belief, 327e330 moving through GPG, 333e336 propositions, 330e331 R-Display and interrupt, 331e333 levels, 341e344 representation error correction, 324e325 general protocol grammar, 320e324 protocol example, 318e320 representing problems, 325 triadic, 338e341 Protocols proper, 326e346 controlled variable in protocol, 336e338 protocol as communication, 344e346 protocol function, 327e336 protocol levels, 341e344 triadic protocols, 338e341 “Proximal physical stimuli”, 235 Psychological therapies, 593 Psychotherapy Grawe’s psychological therapies, 593 method of levels (MOL), 593 reality therapy, 592

R Receptive fields (RFs), 191 Reduction constraint, 427 Reductions, 396e400, 397t Reference signals, 132e135 Reflexology, 4 Relationship-level perceptions, 388e389 Relative distance, 80 Reorganization, 505e506, 589, 603, 619 “Reorganization of Conflict” (RoC) scale, 511 Repetition, 390 Repetition constraint, 427 Rerouted perceptual memory, 590, 603 Research design, 621e622 Resolution of conflict, 591, 605e606 Reverse engineering, 9 control systems, 11e12 from forward engineering perspective, 11e12 living control systems, 16e20 a robot, 14e16

Index Robotics behavior, 542 behaviour-based robotics, 546e548, 548f comparing approaches, 544e551 hierarchical control, 533e538, 534f Hierarchical Perceptual Control Theory (HPCT), 529 hierarchies, 538e542, 539f memory-based prediction, 545 model-based approach, 522, 523f model-based predictive control, 544e545 modeling behavior, 542e543 modeling dynamics, 548e551 models and control, 520e526 open-loop system, 521f perception-based robotics, 518e520 perceptual control, 527e533, 528f purposes, 526e527 traditional AI approach, 519 Rube Goldberg system, 316

S Scientific method, 352 Scientific validity, 510e512 Second-order operators, 393 Self-driving cars, 525 Self-efficacy studies, 479e480 Self-regulation, 490e491 Sensorimotor coordination, 173e190 Sensorimotor transformations, 37e42 Servos, 31 Signal-to-noise sensitivity, 155e158 Smaller-scale social structures, 267 Social change, sources of, 292e294 Social cognitive/self-efficacy theory (SC/SET), 480e481 Social cognitive theory (SCT), 472e473 Social psychology, 593e595 Social stability collective control and social structural levels, 272 cultural environments and social structural levels, 274e275 high and low culture and layers of perception, 275e277 stabilization of physical environments, 272e274 two faces of social structure, 277e278 Social structures, 264e271, 278 collective control networks, 265e271 embedding and interleaving of, 268e269 hyper-network, 279

643

innovation and change in competition, innovation, and social structures, 290 consequential and inconsequential innovations, 290e292 mismatches between self and living environment, 288e290 socialization of new members of learning by imitation and play, 285e287 redundancy of feedback paths and reorganization of perceptual hierarchies, 283e285 types of new members of social structures, 287e288 work and, 280e283 Society for Industrial and Organizational Psychology’s (SIOP), 471 Society of mind, 599 Sociology affect control theory, 595e596 identity control theory, 596 Soft sciences, 4 Song motor pathway (SMP), 215e216 Spatial frequencies modeling, 167e170 Spatial-frequency hypothesis, 176e177 Stabilized cultural environments, 274e275 Stable feedback paths, 280e282 Standard PCT model of perceptual control system, 230 ‘Stimulus-response psychology’, 4 String contiguity, 398 Strong inference, 483 Structural reliability, 158 Subjective meanings, 415e418 Subject-matter domain knowledge, 390 Sublanguage, 411e415 Successive-state analysis, 355e356 Sudden Infant Death Syndrome (SIDS), 207 ‘Sustained’ V1 neuron, 152 Synthesis-by-analysis-by-synthesis, 573

T Temporal flow of experience, 118e128 Test for the Controlled Variable (TCV), 17, 20, 203, 231, 336e337, 602e603 ‘Test-Operate-Test-Exit’ scheme (T.O.T.E.), 355e356, 465e466 Theory of Mind (ToM), 374, 377 Thermodynamics, 423 “Through control theory glasses”, 16e17 Time-binding process, 618 Traditional AI approach, 519

644 Index Transdiagnostic MOL therapy, 509 ‘Transient’ V1 neuron, 152 “Transition probabilities”, 122 Triadic protocols, 338e341 Triangle Faithfulness, 67e68

U Ultra-stabilized collective control processes, 243 Umwelt, 202 Unconscious scripted responses, 470

V Vancouver’s self-regulatory theory, 594e595 Variation, 418e421 Vision, empirical strategy of, 598 Visual input functions, 154e173 Visual tracking systems, 42 Vowel quadrilateral, 380, 380f

W Weighted Intersection Mechanism (WIM) model, 165e166 Wiener model, 30e31 Words, 376 dependencies, 387e389 morphemes, 386e387 order, 389 selection, 389 shapes, 389 Word sequences, 399e400 Word-sharing constraint, 427

Z Zero correlation sufficient conditions for, 69e71 between variable and its derivative, 51e53