528 102 68MB
English Pages [60] Year 2021
Table of contents :
Spectrum0521_cover1
Spectrum0521_cover2
Spectrum0521_01
Spectrum0521_02
Spectrum0521_03
Spectrum0521_04
Spectrum0521_05
Spectrum0521_06
Spectrum0521_07
Spectrum0521_08
Spectrum0521_09
Spectrum0521_10
Spectrum0521_11
Spectrum0521_12
Spectrum0521_13
Spectrum0521_14
Spectrum0521_15
Spectrum0521_16
Spectrum0521_17
Spectrum0521_18
Spectrum0521_19
Spectrum0521_20
Spectrum0521_21
Spectrum0521_22
Spectrum0521_23
Spectrum0521_24
Spectrum0521_25
Spectrum0521_26
Spectrum0521_27
Spectrum0521_28
Spectrum0521_29
Spectrum0521_30
Spectrum0521_31
Spectrum0521_32
Spectrum0521_33
Spectrum0521_34
Spectrum0521_35
Spectrum0521_36
Spectrum0521_37
Spectrum0521_38
Spectrum0521_39
Spectrum0521_40
Spectrum0521_41
Spectrum0521_42
Spectrum0521_43
Spectrum0521_44
Spectrum0521_45
Spectrum0521_46
Spectrum0521_47
Spectrum0521_48
Spectrum0521_49
Spectrum0521_50
Spectrum0521_51
Spectrum0521_52
Spectrum0521_53
Spectrum0521_54
Spectrum0521_55
Spectrum0521_56
Spectrum0521_cover3
Spectrum0521_cover4
What’s Cooler Than Putting a Radio on Venus? Having it work for years P.24
The Power Struggle in Wireless How Open RAN could boost tech innovation P.40
Artificial Emotional Intelligence
The 2021 IEEE Medal of Honor Jacob Ziv made data compression bit-perfect P.48
FOR THE TECHNOLOGY INSIDER MAY 2021
Your Qubits. Controlled. Meet the next generation of signal generators setting new standards for the control of superconducting qubits.
Your Benefits ∏High-fidelity qubit operation Get the most out of your quantum device with the excellent spectral purity provided by super-heterodyne signal generation. ∏Efficient workflows Operate at frequencies up to 8.5 GHz and free from mixer calibration to maximize system up-time. ∏Turnkey feature set Accelerate quantum computer tune-up with rapid pulse reconfiguration and low-latency feedback. ∏Scalable system approach Rely on accurate channel synchronization across your Zurich Instruments Quantum Computing Control System.
Zurich Instruments
Contact us today www.zhinst.com
VOLUME 58 / ISSUE 05
MAY 2021
24 Venus
Calling
Silicon carbide radio circuits can handle the heat of our hellish sister planet. By Alan Mantooth, Carl-Mikael Zetterling & Ana Rusu
TOP: NOEMOTION; BOTTOM: JACOB ZIV/TECHNION
Building an AI That Feels
32
Emotionally intelligent AI could lead to virtual therapists, productivity assistants, and other systems that better understand humans. By Mary Czerwinski, Javier Hernandez & Daniel McDuff
The Clash Over 5G’s First Mile A fundamentally new way of building radio access networks has the entire cellular industry thinking how best to adapt. By Michael Koziol
Conjurer of Compression IEEE 2021 Medal of Honor recipient Jacob Ziv’s pioneering compression algorithms help pack data into WinZips and cat GIFs today. By Tekla S. Perry
40
48
NEWS Debiasing AI (p.8) Qubit Control Chips (p.10) EV Trucks (p.11)
6
HANDS ON 16 Analog TVs make great platforms for hardware hacking. CROSSTALK 20 Numbers Don’t Lie (p.20) Internet of Everything (p.22) Macro & Micro (p.23) PAST FORWARD Prelude to a Tattoo
56
ON THE COVER: Illustration by Eddie Guy
MAY 2021 SPECTRUM.IEEE.ORG 1
BACK STORY
Cure in 30-45 seconds
Under 405 nm light source
User friendly
Alternative to UV curing
Variety of applications
Bonds heat sensitive components
Suitable in thick sections Up to 0.125 inch
No mixing required
One component systems
Specific grades offer optical clarity, dimensional stability and biocompatibility
learn how to use an
LED curing adhesive
Scan to watch
154 Hobart St., Hackensack, NJ 07601 USA +1.201.343.8983 • [email protected]
www.masterbond.com
2 SPECTRUM.IEEE.ORG MAY 2021
H
opefully you’ve noticed something different about your print IEEE Spectrum this month. It has a new look, the result of a comprehensive redesign whose results will soon be seen on our website as well. For this, we’re particularly indebted to the IEEE New Initiatives Committee for its generous support. This new conception came from one of the best design firms in the business: Pentagram. We were fortunate indeed to work with a trio consisting of partner Luke Hayman [top photo] and designers Austin Maurer [middle] and Laura McNeill [bottom]. This issue is the first with the new design. A publication redesign these days is a big, complex affair. It encompasses not just traditional visual details such as type fonts, color palettes, and article templates but also myriad others that specify a website’s architecture: how the site is designed to help a reader navigate through it and find what he or she wants in as smooth and pleasurable a manner as possible. Designing a technology and science magazine brings unique challenges, our Pentagram team noted. “Sometimes I feel like it’s easy, at least with science and technology stuff, to go too far, and it almost becomes a caricature of itself—too futuristic,” says Maurer. “Back a couple of decades ago, people would do very futuristic-seeming things.” So for Spectrum, the team had an overarching goal: “We wanted it to feel very technical and engineering and precise, and to nod to the subject matter, but without it becoming...a caricature.” Realizing that goal was a matter of getting a lot of details right. Consider typography: The team picked two different families of type fonts for the new Spectrum, chosen in part because they evoked “code screens and engineering,” Hayman explains. “We’ve also introduced a monospaced font, which is very much from the world of code.” A monospaced font is one in which all characters are the same width. They were once very popular in computing because they could be displayed on the monitors then available. Engineers used to designing starkly utilitarian objects—D-to-A converters, transformers, heat sinks—may wonder: When a publication is redesigned, how do you know if the project was a success? For Hayman, the answer is in plain sight. “A great design, I think we all know it. We have magazines we love…. On the way back from the mailbox, you’re already reading it.”
TO WATCH A VIDEO OF OUR INTERVIEW WITH THE PENTAGRAM TEAM, GO TO spectrum.ieee.org/pentagram-may21
FROM TOP: PENTAGRAM (2); ZÖE GLENN
use LED Curing ADHESIVES?
A Magazine Reborn
CONTRIBUTORS
MARY CZERWINSKI EDITOR IN CHIEF Susan Hassler, [email protected] EXECUTIVE EDITOR Glenn Zorpette, [email protected] EDITORIAL DIRECTOR, DIGITAL Harry Goldstein, [email protected] MANAGING EDITOR Elizabeth A. Bretz, [email protected] SENIOR ART DIRECTOR Mark Montgomery, [email protected] PRODUCT MANAGER, DIGITAL Erico Guizzo, [email protected] SENIOR EDITORS Evan Ackerman (Digital), [email protected] Stephen Cass (Special Projects), [email protected] Jean Kumagai, [email protected] Samuel K. Moore, [email protected] Tekla S. Perry, [email protected] Philip E. Ross, [email protected] David Schneider, [email protected] Eliza Strickland, [email protected] DEPUTY ART DIRECTOR Brandon Palacio, [email protected] PHOTOGRAPHY DIRECTOR Randi Klett, [email protected] ONLINE ART DIRECTOR Erik Vrielink, [email protected] NEWS MANAGER Mark Anderson, [email protected] ASSOCIATE EDITORS Willie D. Jones (Digital), [email protected] Michael Koziol, [email protected] SENIOR COPY EDITOR Joseph N. Levine, [email protected] COPY EDITOR Michele Kogon, [email protected] EDITORIAL RESEARCHER Alan Gardner, [email protected] ADMINISTRATIVE ASSISTANT Ramona L. Foster, [email protected] CONTRIBUTING EDITORS Robert N. Charette, Steven Cherry, Charles Q. Choi, Peter Fairley, Maria Gallucci, W. Wayt Gibbs, Mark Harris, Jeremy Hsu, Allison Marsh, Prachi Patel, Megan Scudellari, Lawrence Ulrich, Emily Waltz EDITOR IN CHIEF, THE INSTITUTE Kathy Pretz, [email protected] ASSISTANT EDITOR, THE INSTITUTE Joanna Goodrich, [email protected] DIRECTOR, PERIODICALS PRODUCTION SERVICES Peter Tuohy MULTIMEDIA PRODUCTION SPECIALIST Michael Spector ASSOCIATE ART DIRECTOR, PUBLICATIONS Gail A. Schnitzer ADVERTISING PRODUCTION +1 732 562 6334 ADVERTISING PRODUCTION MANAGER Felicia Spagnoli, [email protected] SENIOR ADVERTISING PRODUCTION COORDINATOR Nicole Evans Gyimah, [email protected] EDITORIAL ADVISORY BOARD, IEEE SPECTRUM Susan Hassler, Chair; David C. Brock, Robert N. Charette, Ronald F. DeMara, Shahin Farshchi, Lawrence O. Hall, Jason K. Hui, Leah Jamieson, Mary Lou Jepsen, Deepa Kundur, Peter Luh, Michel Maharbiz, Somdeb Majumdar, Allison Marsh, Carmen Menoni, Sofia Olhede, Wen Tong, Maurizio Vecchione EDITORIAL ADVISORY BOARD, THE INSTITUTE Kathy Pretz, Chair; Qusi Alqarqaz, Philip Chen, Shashank Gaur, Lawrence O. Hall, Susan Hassler, Peter Luh, Cecilia Metra, San Murugesan, Mirela Sechi Annoni Notare, Joel Trussell, Hon K. Tsang, Chenyang Xu MANAGING DIRECTOR, PUBLICATIONS Steven Heffner EDITORIAL CORRESPONDENCE IEEE Spectrum, 3 Park Ave., 17th Floor, New York, NY 10016-5997 TEL: +1 212 419 7555 FAX: +1 212 419 7570 BUREAU Palo Alto, Calif.; Tekla S. Perry +1 650 752 6661 DIRECTOR, BUSINESS DEVELOPMENT, MEDIA & ADVERTISING Mark David, [email protected] ADVERTISING INQUIRIES Naylor Association Solutions, Erik Henson +1 352 333 3443, [email protected] REPRINT SALES +1 212 221 9595, ext. 319 REPRINT PERMISSION / LIBRARIES Articles may be photocopied for private use of patrons. A per-copy fee must be paid to the Copyright Clearance Center, 29 Congress St., Salem, MA 01970. For other copying or republication, contact Managing Editor, IEEE Spectrum. COPYRIGHTS AND TRADEMARKS IEEE Spectrum is a registered trademark owned by The Institute of Electrical and Electronics Engineers Inc. Responsibility for the substance of articles rests upon the authors, not IEEE, its organizational units, or its members. Articles do not represent official positions of IEEE. Readers may post comments online; comments may be excerpted for publication. IEEE reserves the right to reject any advertising.
IEEE BOARD OF DIRECTORS PRESIDENT & CEO Susan K. “Kathy” Land, [email protected] +1 732 562 3928 Fax: +1 732 981 9515 PRESIDENT-ELECT K.J. Ray Liu TREASURER Mary Ellen Randall SECRETARY Kathleen A. Kramer PAST PRESIDENT Toshio Fukada VICE PRESIDENTS Stephen M. Phillips, Educational Activities; Lawrence O. Hall, Publication Services & Products; Maike Luiken, Member & Geographic Activities; Roger U. Fujii, Technical Activities; James E. Matthews, President, Standards Association; Katherine J. Duncan, President, IEEE-USA DIVISION DIRECTORS Alfred E. “Al” Dunlop (I); Ruth A. Dyer (II); Sergio Benedetto (III); Manfred “Fred” J. Schindler (IV); Thomas M. Conte (V); Paul M. Cunningham (VI); Miriam P. Sanders (VII); Christina M. Schober (VIII); Rabab Kreidieh Ward (IX); Dalma Novak (X) REGION DIRECTORS Eduardo F. Palacio (1); Barry C. Tilton (2); Jill I. Gostin (3); Johnson A. Asumado (4); James R. Look (5); Timothy T. Lee (6); Jason Jianjun Gu (7); Antonio Luque (8); Alberto Sanchez (9); Deepak Mathur (10) DIRECTOR EMERITUS Theodore W. Hissey IEEE STAFF EXECUTIVE DIRECTOR & COO Stephen Welby +1 732 562 5400, [email protected] CHIEF INFORMATION OFFICER Cherif Amirat +1 732 562 6017, [email protected] PUBLICATIONS Steven Heffner +1 212 705 8958, [email protected] CHIEF MARKETING OFFICER Karen L. Hawkins +1 732 562 3964, [email protected] CORPORATE ACTIVITIES Donna Hourican +1 732 562 6330, [email protected] MEMBER & GEOGRAPHIC ACTIVITIES Cecelia Jankowski +1 732 562 5504, [email protected] STANDARDS ACTIVITIES Konstantinos Karachalios +1 732 562 3820, [email protected] EDUCATIONAL ACTIVITIES Jamie Moesch +1 732 562 5514, [email protected] GENERAL COUNSEL & CHIEF COMPLIANCE OFFICER Sophia A. Muirhead +1 212 705 8950, [email protected] CHIEF FINANCIAL OFFICER Thomas R. Siegert +1 732 562 6843, [email protected] TECHNICAL ACTIVITIES Mary Ward-Callan +1 732 562 3850, [email protected] MANAGING DIRECTOR, IEEE-USA Chris Brantley +1 202 530 8349, [email protected] IEEE PUBLICATION SERVICES & PRODUCTS BOARD Lawrence O. Hall, Chair; Sergio Benedetto, Edhem Custovic, Stefano Galli, James Irvine, Clem Karl, Hulya Kirkici, Fabrizio Lobardi, Aleksandar Mastilovic, Sorel Reisman, Gaurav Sharma, Isabel Trancoso, Maria Elena Valcher, Peter Winzer, Bin Zhao IEEE OPERATIONS CENTER 445 Hoes Lane, Box 1331 Piscataway, NJ 08854-1331 U.S.A. Tel: +1 732 981 0060 Fax: +1 732 981 1721
IEEE SPECTRUM (ISSN 0018-9235) is published monthly by The Institute of Electrical and Electronics Engineers, Inc. All rights reserved. © 2021 by The Institute of Electrical and Electronics Engineers, Inc., 3 Park Avenue, New York, NY 10016-5997, U.S.A. Volume No. 58, Issue No. 5. The editorial content of IEEE Spectrum magazine does not represent official positions of the IEEE or its organizational units. Canadian Post International Publications Mail (Canadian Distribution) Sales Agreement No. 40013087. Return undeliverable Canadian addresses to: Circulation Department, IEEE Spectrum, Box 1051, Fort Erie, ON L2A 6C7. Cable address: ITRIPLEE. Fax: +1 212 419 7570. INTERNET: [email protected]. ANNUAL SUBSCRIPTIONS: IEEE Members: $21.40 included in dues. Libraries/institutions: $399. POSTMASTER: Please send address changes to IEEE Spectrum, c/o Coding Department, IEEE Service Center, 445 Hoes Lane, Box 1331, Piscataway, NJ 08855. Periodicals postage paid at New York, NY, and additional mailing offices. Canadian GST #125634188. Printed at 120 Donnelley Dr., Glasgow, KY 42141-1060, U.S.A. IEEE Spectrum circulation is audited by BPA Worldwide. IEEE Spectrum is a member of the Association of Business Information & Media Companies, the Association of Magazine Media, and Association Media & Publishing. IEEE prohibits discrimination, harassment, and bullying. For more information, visit https://www.ieee.org/web/aboutus/whatis/policies/p9-26.html.
Czerwinski started out as a cognitive psychologist studying productivity and stress. Today, as research manager of Microsoft’s Human Understanding and Empathy group, she works on AI that can understand those issues. As she and Microsoft colleagues Daniel McDuff and Javier Hernandez describe in “Building an AI That Feels” [p. 32], the team is riffing on Microsoft’s mission to help everyone achieve more. They go further, aiming to help every person achieve “their best self.”
EDDIE GUY Guy, a New Jersey–based illustrator, specializes in 3D-rendered images that seem to pop off the page. For this issue’s cover art, which accompanies an article about emotionally intelligent AI [p. 32], he decided to depict an empathetic humanoid robot. His own domestic AI systems, in the form of virtual assistants like Siri, show a bit of emotional awareness. “If you’re rude enough, sometimes the assistant will respond in a snippy tone,” he says.
JEREMY HSU IEEE Spectrum contributing editor Hsu began writing about quantum computing for us in 2013. Only in the last few years, though, has Spectrum’s coverage tilted toward the tech’s grittier details, as Hsu’s contribution does this month in “Fast, Cold, and Under Control” [p. 10]. Typically CMOS chips run warm and qubits run cold—which means they operate in different realms. But, he says, researchers are bringing them closer together.
ALAN MANTOOTH Mantooth of the University of Arkansas is a Fellow of the IEEE. After a project on electronics for extreme environments with the Jet Propulsion Laboratory, he began to note “all the places where we could not yet put electronics and what an advantage it would be if we could,” he says. In “Venus Calling” [p. 24], he, Ana Rusu, and Carl-Mikael Zetterling—colleagues at KTH Royal Institute of Technology—explain why silicon carbide is what’s needed for the extreme environment of Venus.
MAY 2021 SPECTRUM.IEEE.ORG 3
SPECTRAL LINES
Ellume and Relativity Space Win IEEE Spectrum Awards A COVID-19 home test and 3D-printed rockets were notable achievements BY SUSAN HASSLER
I
t was back in 2005 that the first IEEE Spectrum Awards were presented to Nesscap Co., a South Korean maker of ultracapacitors, and to Microsoft TV, for Internet Protocol television. In those days, the IEEE Spectrum Awards were given out in partnership with EE Times, at the Annual Creativity in Electronics (ACE) Awards. In 2015, we were delighted to be able to join IEEE’s Honors Ceremony to present the IEEE Spectrum Awards. The first
of these yearly awards, for Technology in the Service of Society, recognizes a standout achievement in IEEE’s core mission: to foster technological innovation and excellence for the benefit of humanity. The second, the Emerging Technology Award, honors an emerging technology that has the potential to transform an entire industry. This year, the IEEE Spectrum Awards will be presented at the 2021 IEEE Vision, Innovation, and Challenges Summit &
Honors Ceremony, which will be held virtually on 11–13 May 2021. The winning companies are chosen by IEEE Spectrum’s editors. Every January, we review notable technology achievements that have appeared in our coverage and elsewhere over the past year or so. We then debate the merits of the nominees before reaching a final decision. It was particularly difficult this year to settle on a company for the category of Technology in Service of Society. The pandemic has inspired engineers worldwide to rise to the challenge of designing technologies to help defeat COVID-19. We had many deliberations and considered dozens of nominees, but we agreed unanimously that this award should go to Ellume, of East Brisbane, Australia, for its COVID-19 Home Test. Founded in 2010 by Dr. Sean Parsons, Ellume makes diagnostic health care products that are designed to work with mobile devices. Ellume’s home test is the first to be authorized by the FDA and the first to become commercially available—over the counter and without a prescription. The test will initially be rolled out in the United States and Australia and will sell for about US $30. For our awardee in the category of Emerging Technology, not even the sky’s the limit. After considering dozens of serious contenders, we are pleased to present the award to the rocket startup company Relativity Space, from Long Beach, Calif., for its Stargate factory technology, which makes it possible to 3D print a small-lift launch vehicle in weeks instead of months. Apart from its electrical systems, the entire rocket is made up of 3D-printed parts. Tim Ellis is the CEO, and he cofounded Relativity Space in 2015. It’s the world’s first “all in one” rocket factory. By fusing 3D printing, artificial intelligence, and “intelligent” robotics, Relativity Space is changing the future of space exploration. Its first 3D-printed rocket, Terran 1, is scheduled to launch later this year.
WITH THIS ISSUE we are delighted to debut Spectrum’s new print look, the magazine’s first redesign in nearly 10 years.
The project helps ensure that Spectrum continues to align with IEEE’s mission of attracting new members, and of supporting the interests of current members in new and emerging technologies. We collaborated with the internationally renowned design company Pentagram on the print project, and are also working with them on revamping our website and mobile edition, now scheduled to be rolled out in July of this year. The logo and typography have been refreshed; the layouts are crisp and contemporary. We hope you enjoy your new and improved Spectrum! To read more about the design process, check out this month’s Back Story.
4 SPECTRUM.IEEE.ORG MAY 2021
EDWARD CARREON PHOTOGRAPHY/RELATIVITY SPACE
Relativity Space CEO Tim Ellis stands alongside part of the company’s 3D-printed rocket, Terran 1, scheduled to launch this year.
simulation case study
IoT calls for fast communication between sensors Developing the 5G mobile network may not be the only step to a fully functioning Internet of Things, but it is an important one — and it comes with substantial performance requirements. Simulation ensures optimized designs of 5G-compatible technology, like this phased array antenna. learn more comsol.blog/5G
The COMSOL Multiphysics® software is used for simulating designs, devices, and processes in all fields of engineering, manufacturing, and scientific research.
THE LATEST DEVELOPMENTS IN TECHNOLOGY, ENGINEERING, AND SCIENCE
Adidas Futurecraft. Strung manufacturing robots weave highperformance shoes for high-speed runners.
When Footwear Becomes Footware Adidas explores complex geometries of support for the world’s fastest feet BY BRIAN T. HOROWITZ
6 SPECTRUM.IEEE.ORG MAY 2021
F
or athletes trying to run fast, the proper shoe can be essential to achieving peak performance. For athletes trying to run as fast as humanly possible, a runner’s shoe can also become a work of individually customized engineering. This is why Adidas has married 3D printing with robotic automation in a mass-market footwear project called Futurecraft.Strung. Using a customized, 3D-printed sole, a Futurecraft.Strung manufacturing robot can place some 2,000 threads from up to 10 different
ADIDAS
INDUSTRIAL ROBOTICS
MAY 2021
2,000
Number of threads in one section of a Futurecraft.Strung running shoe
sneaker yarns in one upper section of the shoe, according to an athlete’s requirements. Skylar Tibbits, founder and codirector of the Self-Assembly Laboratory and associate professor of design research in MIT’s department of architecture, says that because of its small scale, footwear has been an area of focus for 3D printing and additive manufacturing, which involves adding material bit by bit. “There are really interesting complex geometry problems,” he says. “It’s pretty well suited.” Adidas began working on the Futurecraft.Strung project in 2016. Then two years later, Adidas Futurecraft, the company’s innovation incubator, began collaborating with digital design studio Kram/Weisshaar. In less than a year the team built the software and hardware for the upper part of the shoe, called Strung uppers. “Most 3D printing in the footwear space has been focused on the midsole or outsole, like the bottom of the shoe,” Tibbits explains. But now, he says, Adidas is bringing robotics and a threaded design to the upper part of the shoe. The company bases its Futurecraft. Strung design on high-resolution scans of how runners’ feet move as they travel. This more flexible design can benefit athletes in multiple sports, according to an Adidas blog post. It will use motion capture of an athlete’s foot and feedback from the athlete to create a design specifically for the athlete’s unique gait. Adidas customizes the weaving of the shoe’s “fabric” (really more like an elaborate woven string figure, a cat’s cradle to fit the foot) to achieve a close and comfortable fit, the company says. What they call their “4D sole” consists of a design combining 3D printing with materials that can change their shape and properties over time. In fact, Tibbits coined the term “4D printing” to describe this process in 2013. The c ompany takes customized data from the Adidas Athlete Intelligent Engine to make the shoe, according to Kram/Weisshaar’s website.
The “fabric” for these shoes is really an elaborate string figure, a cat’s cradle for the foot.
Expect Adidas Futurecraft.Strung shoes in the marketplace by later this year or early 2022.
“With Strung, for the first time we can program single threads in any direction, where each thread has a different property or strength,” says Fionn Corcoran- Tadd, an innovation designer at Adidas’s Futurecraft lab, in a company video. He says that each thread serves a purpose. “This is like customized string art for your feet,” Tibbits says. Although the robotics technology the company uses has been around for many years, what Adidas’s robotic weavers can achieve with thread is a matter of elaborate geometry. “It’s more just like a really elegant way to build up material, using robotics to combine the fibers and yarns into these intricate and complex patterns,” Tibbits says. Robots can of course create patterns with more precision than a person weaving them by hand could, as well as rapidly and reliably change the yarn and color of the fabric pattern. Adidas says it can make a single upper in 45 minutes and a pair of sneakers in 1 hour and 30 minutes. It plans to reduce this time to minutes in the months ahead, the company says. An Adidas spokesperson says that sneakers incorporating the Futurecraft. Strung uppers design are a prototype, but the company plans to bring a Strung shoe to market in late 2021 or 2022. Adidas Futurecraft sneakers are currently available with a 3D-printed midsole. Adidas plans to continue gathering data from athletes to customize the uppers of sneakers. “We’re building up a library of knowledge, and it will get more interesting as we aggregate data of testing and from different athletes and sports,” the Adidas Futurecraft team writes in a blog post. “The more we understand about how data can become design code, the more we can take that and apply it to new Strung textiles. It’s a continuous evolution.”
MAY 2021 SPECTRUM.IEEE.ORG 7
ARTIFICIAL INTELLIGENCE
Engineering Bias Out of AI Machines that learn the worst human impulses can still relearn BY PRACHI PATEL
A
rtificial-intelligence systems today increasingly determine whether someone gets a job or a loan, how much they’re paid, how they’re treated by doctors and hospitals, and how fairly they’re dealt with by government, law enforcement, and the justice system. If the AI behind these autonomous decision-making systems is deeply biased—more likely to discriminate
8 SPECTRUM.IEEE.ORG MAY 2021
against people on the basis of race, religion, sex, or other criteria—then that problem can be very difficult to weed out. “But it is also an opportunity,” says Alexandra Ebert, chief trust officer at the Vienna-based synthetic-data startup Mostly AI: a chance for businesses, data scientists, and engineers to begin the hard but important work of extricating bias from AI data sets and algorithms. Removing bias from AI is not easy
because there’s no one cause for it. It can enter the machine-learning cycle at various points. But the logical and most promising starting point seems to be the data that goes into it, says Ebert. AI systems rely on deep neural networks that parse large training data sets to identify patterns. These deep-learning methods are roughly based on the brain’s structure, with many layers of code linked together like neurons, and weights given to the links changing as the network picks up patterns. The problem is, training data sets may lack enough data from minority groups, reflect historical inequities such as lower salaries for women, or inject societal bias, as in the case of Asian-Americans being labeled foreigners. Models that learn from biased training data will propagate the same biases. But collecting high-quality, inclusive, and balanced data is expensive. So Mostly AI is using AI to create synthetic data sets to train AI. Simply removing sensitive features like race or changing them—say, increasing female salaries to affect approved credit limits— does not work because it interferes with other correlations. Instead, the startup uses a deep neural network that learns the patterns and connections in a data set and automatically generates a whole new individual—for instance, one “who behaves like a female with higher income would behave, so that all the data points from a person match up and make sense,” Ebert says. The synthetic data sacrifices accuracy slightly compared to that of the original data but is still statistically highly representative. Other startups like London-based Synthesized have thrown their hats into the synthetic data game. Mostly AI and a few other companies are now in the process of launching an IEEE standards group on synthetic data, Ebert adds. Researchers have also developed several tools to help reduce bias in AI. Tool kits like Aequitas measure bias in uploaded data sets, while others like Themis–ml also offer a few ways to reduce that bias using bias-mitigation algorithms. A team at IBM Research led by Kush Varshney has brought these efforts together to create a comprehensive open-source tool kit called AI Fairness 360, which helps detect and reduce
Photo-illustration by Stuart Bradford
NEWS
Teaching Computers to Unlearn Bias-mitigation algorithms can make predictions of an AI system more fair and less prone to discrimination. Pictured here are results of a ProPublica study of a criminal-sentencing program called COMPAS (Correctional Offender Management Profiling for Alternative Sanctions). COMPAS scores, as ProPublica pointed out, “inform decisions about who can be set free at every stage of the criminal justice system.” In its predictions of the likelihood of inmates’ reoffending, COMPAS got it wrong twice as often with nonwhite inmates as it did their white
fellow prisoners. COMPAS also incorrectly tagged men as high-risk for reoffending three times as often as it did for women. But as IBM researchers show in a demo associated with their AI Fairness 360 tool kit, an algorithm called Reweighing (Algorithm 1 in the chart) greatly mitigates bias related to sex, and even flips the bias with respect to race. Meanwhile, another algorithm called Optimized Preprocessing (Algorithm 2) almost eliminates racial bias— at least according to the present metrics— but fares poorly in debiasing the data related to sex. A third algorithm, Adversarial Debiasing, corrects fairly well for sexrelated bias but does not debias the data very well at all with respect to race.
unwanted bias in data sets and machine-learning models (see infographic). It brings together 14 different bias-mitigation algorithms developed by computer scientists over the past decade, and is intuitive to use. “The idea is to have a common interface to make these tools available to working professionals,” says Varshney. Some of the algorithms in the tool kit massage training data in sophisticated ways. For instance, they change how certain inputs and outputs are paired (a zip code with “yes” on loan approval), or the weights that are given to specific pairs. A technique called reweighing, for instance, gives higher weight to input/output pairs that give the underprivileged group a positive outcome. Yet it’s hard to know how your model will behave in the real world. This makes external audits of an AI system crucial. An AI algorithm can look perfect and yet perform miserably for some stakeholders, says Cathy O’Neil, data scientist and founder of O’Neil Risk Consulting & Algorithmic Auditing. “I test the code as a black box to see how it performs,” she
Bias
Fair
Bias
–0.36
–0.03
–0.18 REMOVING SEX-RELATED BIAS
–0.04
–0.18
0.02
–0.01 REMOVING RACE-BASED BIAS
–0.14 –0.5 Original
AI decisionmaking tools that are both accurate and fair are also more useful for a larger group of potential customers.
0 Algorithm 1
0.5 Algorithm 2
Algorithm 3
says. “I ask the broad question: Who does your algorithm fail? Soon as you ask that question—boom, there’s all sorts of things that can go wrong.” Investing the time and effort to make AI models fairer comes with a cost. So without regulatory pressure, companies find it easier to focus on how accurately their systems reflect the original data. But a Gartner Research study predicted that by 2022, 85 percent of AI projects will deliver wrong outcomes due to bias in data, algorithms, or the teams responsible for managing them. So AI decision-making tools that are both accurate and fair would not only be more useful for a larger group of potential customers, says Mostly AI’s Ebert, but also prevent a reputational hit and customer backlash when bias in proprietary AI shows up. “We need to ensure that AI benefits us,” she says. “But debiased AI can also lead to higher benefits for businesses.” FOR AN EXPANDED VERSION OF THIS STORY, GO TO spectrum.ieee.org/reducingbias-may21
MAY 2021 SPECTRUM.IEEE.ORG 9
NEWS
Fast, Cold, and Under Control Conventional electronics join quantum circuits in the deep freeze BY JEREMY HSU
T
hree of the biggest companies making quantum computers today— Google, Intel, and Microsoft—are betting on supercooled devices operating at close to absolute zero. But there’s a problem: These cold cathedrals of quantum computing cannot tolerate the extra heat given off by the conventional computing chips that control them. This means the classical and quantum- computing components must be separated, despite their marriage by design. The control chips usually reside at room temperature on top of the quantum-computing stack, while the quantum bits (qubits) remain in the coldest depths of dilution refrigerators. The dilution fridges involve helium-3 and helium-4 isotopes to supercool the environment, lowering temperatures from a baseline of 4 kelvins (–269.15 ºC) at the top to about 10 millikelvins at the bottom. Cables running up and down the hardware stack connect each qubit with its control chip and other conventional computing components higher up. Such unwieldy setups with just dozens of qubits would become an “engineering nightmare” if scaled up to the number of qubits necessary for practical quantum computing, says Fabio Sebastiano, research lead for the quantum-computing division at QuTech, in Delft, Netherlands. He compared the approach to trying to connect each of the 10 million pixels in a smartphone camera to their readout electronics using 1-meter cables. That is why these three tech giants have been developing either qubits that operate at warmer temperatures or control chips that operate at colder temperatures—while minimizing heat from power dissipation. The companies hope to shrink the operating temperature difference and possibly unite classical and quantum-computing components in the same integrated chips or packages.
10 SPECTRUM.IEEE.ORG MONTH 10 SPECTRUM.IEEE.ORG MAY 2021 2021
Qubit and Control-Chip Temperatures
3K
4 kelvins
4K INTEL’S Horse Ridge II control chip
3K
GOOGLE’S control chip
2K 2K
MICROSOFT’S generalpurpose Alta cryocompute chip determines what instructions get sent to the Gooseberry control chip, which resides at temperatures 20 times as cold.
1K INTEL’S silicon spin qubits trade warmer temperatures (15 times as warm as comparable qubits) for slightly diminished performance.
1K
0.1K MICROSOFT’S Gooseberry control chip converts classical-computing instructions into voltage signals that can control the qubits next door.
0.01K GOOGLE’S superconducting qubits
0K
FROM TOP: INTEL; ERIK LUCERO; INTEL; MICROSOFT
QUANTUM COMPUTERS
Tesla’s Cybertruck: Beyond the Thunderdome—and beyond 200 kilowatt-hours?
ELECTRIC VEHICLES
EV Trucks Battle for Supremacy Seven models will compete for a vast market BY LAWRENCE ULRICH
TESLA
T
esla rules global electric- vehicle sales, with nearly 500,000 buyers in 2020, some 23 percent of the worldwide battery-EV market. But in the United States, pickup trucks rule the road. The top three trucks alone, the Ford F-150, Chevy Silverado, and Ram 1500—the best-selling models of any vehicle type— found nearly 2 million customers, even in a COVID-stricken 2020. With pickups and SUVs together hogging a record 76 percent of a total 14.5 million sales in 2020, EV makers have gotten the memo, writ large: Tiny EVs are out. Big and burly is in. The plug-in pickup race includes Tesla, General Motors, and Ford, the latter two now vowing to swiftly electrify their fleets. It’s a compe-
tition with the potential to save billions of gallons of gasoline and diesel every year, between pickups’ ubiquity and their fuel-slurping ways. With so many electric trucks on the cusp, first-mover advantage is no guarantee of success: Doing it right, with appealing design, performance, technology, and a sustainable business plan, will prove more critical than being first to showrooms. IEEE Spectrum has analyzed the contenders, based on what’s known so far, and handicapped the field. We set out our top two most promising EV truck releases below. For the other models in the field (including GMC’s Hummer EV, Rivian’s R1T, as well as long shots like the Bollinger B2 and the quirky Canoo Pickup), see Spectrum’s extended coverage online.
Tesla Cybertruck Betting against Tesla in any EV category would be unwise, given its market dominance, Supercharger infrastructure, captive battery capacity, and the euphoric reception it’s received for cars like the Model 3. The Cybertruck’s out-there styling—Asimov meets Mad Max—won’t appeal to some traditionalists. But the promised numbers are eye-popping: More than 805 kilometers (500 miles) of driving range for a $69,900 “tri-motor” Plaid all-wheel-drive model, with a 2.9second launch to 97 kilometers per hour (60 miles per hour), 6,350-kilogram (14,000-pound) towing capacity, and 1,588-kg maximum payload. Analysts foresee a battery of at least 200 kilowatt-hours and Base price: US $39,900 roughly 800 horsepower. A $39,900 single-motor model Expected should bring a 402-km driv- arrival: 2022 ing range and a 6.5-second scoot to 97 km/h; a $49,900 Raven edition brings dual-motor all-wheel drive and 483 km of stamina. A dent-resistant, stainless steel “exoskeleton” body is notable, along with polymer- composite “armor” glass, and an electric roll-down tonneau cover for the bed. Tesla has ordered a record 7,257-tonne (8,000-ton) casting press to manufacture the Cybertruck’s rear body casting at its
MAY 2021 SPECTRUM.IEEE.ORG 11
Ford’s F-150 has been the best-selling U.S. vehicle for 39 years. Could the F-150 EV someday extend the winning streak?
Gigafactory Texas beginning late this year or in 2022. Considering Elon Musk’s hitor-miss record on promises, our money is on 2022.
(1.25 million pounds). Two years later, Ford has begun selling its F-150 * As of press time, Ford had not released PowerBoost, the first full-size, full- pricing information on its F-150 EV. The hybrid pickup—one of Spectrum’s price given here is Kelley Blue Book’s annual Top 10 tech cars for 2021. Some- estimate. time next year, Ford will add a full- electric F-150 to its formidable F-Series lineup. Tech details are minimal, but FOR AN EXPANDED VERSION Ford has confirmed a dual-motor, all- OF THIS ARTICLE, GO TO wheel-drive layout. Expect a spacious spectrum.ieee.org/evtrucks-may21
FORD
Ford F-150 EV As stunts go, this was a good one: In 2019, Ford used a prototype of its upcoming electric F-150 to tow 10 double-decker rail cars, weighing a total of 566,990 kg
crew cab version (with other body styles likely), a hefty battery pack, and a 483-km (300-mile) range, at least for high-end versions. As much as any vehicle, the Ford will test the mainstream U.S. appetite Base price: US $60,000* for all things electric: Ford sells about 900,000 Expected F-Series pickups in a good arrival: 2022 year, making it the United States’ best-selling vehicle for 39 straight years. As noted by Ted Cannis, Ford’s global director of electrification, if Ford can eventually convert just one in nine buyers, that would be 100,000 electric pickup sales a year—a solid start in the truck transition from fossil fuels.
12 SPECTRUM.IEEE.ORG MAY 2021
NEWS
JOURNAL WATCH
Hack-Proofing the “Smart Farm” The world’s population by 2030 is projected to surpass 8.5 billion, with more than 840 million people affected by acute hunger. Clearly global agriculture needs all the efficiencies technology can confer. Xing Yang of Nanjing Agricultural University, in China, and his colleagues surveyed the different kinds of smart agriculture on the horizon, as well as the key technologies and applications specific to them. (Their research was published in a recent issue of the IEEE/CAA Journal of Automatica Sinica.) Yang says the most pressing cybersecurity problems in smart agriculture involve the physical environment, such as plant factory control-system intrusion and unmanned aerial vehicle (UAV) false positioning. The researchers also paid extra attention to agricultural equipment as potential security threats. “The delay caused by long-distance signal transmission also increases the risk of Sybil attacks,” he says. (Sybil attacks, named after a movie about a person with dissociative identity disorder, transmit malicious data from multiple accounts.) Yang is optimistic that the application of existing technologies—such as edge computing, artificial intelligence, and blockchain— can be used to mitigate some of the existing problems. He says that AI algorithms can be developed that might detect the presence of malicious users, while existing industrial security standards can be applied to design a targeted security scheme for an agricultural “Internet of things.” —Payal Dhar
MAY 2021 SPECTRUM.IEEE.ORG 13
THE BIG PICTURE
Floating List of Chores Even the most technologically advanced houses need someone around who’s handy and can keep them in good repair. That’s what NASA astronaut and Expedition 64 flight engineer Michael Hopkins was doing when he suited up and stepped outside the International Space Station. Hopkins, who was part of the crew that reached the ISS aboard SpaceX’s Resilience aircraft in November (the first successful space launch by a private entity), is seen here performing some maintenance on the cooling system and communications gear at his temporary home away from home. PHOTOGRAPH BY NASA
MAY 2021 SPECTRUM.IEEE.ORG 15
TECH TO TINKER WITH
Old consumer electronics often have enough room for major upgrades.
Geek TV Salvage analog tech with Wi-Fi and a Raspberry Pi Zero BY STEPHEN CASS
16 SPECTRUM.IEEE.ORG MAY 2021
L
ike vinyl and cassette tapes, cathode ray tubes are experienc ing a comeback in some circles. Retrogamers want to experience old titles on the fuzzy displays they were designed for. And makers have been deconstructing them for all kinds of projects: In January of 2020, we fea tured a vector graphics clock based on a CRT originally made for an oscillo scope. But oscilloscope CRTs are rela tively rare, while analog television sets
Illustrations by James Provost
The first published picture of the Mandelbrot set, by Robert W. Brooks and Peter Matelski in 1978
can literally be found littering sidewalks. So when I came across a listing on eBay for a cheap portable black-and-white television with a 5-inch-screen that looked in almost mint condition, I thought I’d take on the challenge of tackling this old technology. Unlike with an oscilloscope CRT, you can’t easily control a television CRT’s electron beam directly. Instead, it’s best to feed in a video signal that the in situ circuitry can decode. So I decided to make a Geek TV: Instead of broadcast channels, it would display the latest updates and imagery from places like NASA and the European Organization for Nuclear Research (CERN), as well as some animations. And I wanted it to reflect the original TV’s self-contained simplicity: no extra cables or boxes dangling off it and certainly no keyboard, mouse, or extra switches for control. My “new” TV is a Coby CX-TV1, which has an RCA composite video connector and runs off a DC power-supply adapter, both of which suggested I would have good jumping off points in the internal circuitry for hacking. Once I unscrewed the case and looked at the guts, I was even more encouraged. The internal circuitry is divided into two parts: a main printed circuit board built around a clone of an AN5151 TV chip, and a secondary PCB mounted at a right angle. The TV’s tuning knob turns the shaft of a variable capacitor on this secondary board. I had always planned to use this sidemounted knob as my control for selecting “channels,” but initially I thought I would mount it to a rotary encoder to detect movement. But once I saw how the PCBs were arranged, I had one of my rare moments of inspiration. I disconnected the tuner board from the main PCB, and hooked up a lead of the variable capacitor to an Adafruit MPR121based touch sensor. This capacitive sensor is intended to provide a simple yes/no response to the question of “Is a finger touching a sur-
The TV’s control circuitry runs at 12 volts, so it has to be stepped down to 5 V by a buck converter, which is fed into the Qwiic pHat terminals. A regulator built into the Pi Zero further steps the voltage down to the 3.3 V used internally by the Pi and by the capacitance sensor.
face or not?” But by looking at raw sensor measurements, I could determine the moment-to-moment capacitance of the variable capacitor, and thus the knob’s position. As a bonus, this meant I could leave undisturbed the cute mechanical mechanism that moves a red plastic strip up and down to indicate the channel the TV is tuned to.
I added smarts to the TV with a Wi-FI-enabled Raspberry Pi Zero W that I had lurking at the bottom of a drawer. (If I was starting from scratch, I would probably use a Pi 3 or 4 compute module for faster processing, but I was happy to put the Pi Zero to good use.) Every Pi has the ability to generate composite video in NTSC or PAL format. In the Pi Zero’s case,
MAY 2021 SPECTRUM.IEEE.ORG 17
HANDS ON
an unpopulated socket provides the signal, so it was a few moments’ work to solder in a header. A little bit more soldering work let me run wires from the header to the back of the TV’s RCA connector. As my TV is black and white, I disabled the composite video’s colorburst signal in the Pi’s configuration file for a slightly sharper image. The only major hardware step remaining was the power supply: Most of the TV’s internal circuitry runs at 12 volts, which would immediately fry the Pi, so I used a buck converter to step down the voltage to 5 V. Two more dabs of solder connected it to the TV’s on/off switch and ground (Beware of the much higher voltages used to drive the CRT itself). I wrote the software in Python: A passel of libraries makes it easy to tie together the hardware with higher level routines for extracting and processing images from online sources. First, I
installed Adafruit’s Blinka bundle— which is a backport of Circuit Python, originally developed to squeeze Python into microcontrollers. With Blinka, I can communicate with my touch sensor via an I 2C connection mounted on a SparkFun Qwiic pHat v2.0 board plugged into the Pi’s GPIO header. (The pHat also provides a convenient pair of screw terminals connected to the Pi’s 5-V and ground rails that lets me bring the power in from the buck converter). Getting imagery from online sources requires a combination of techniques. For example, the status dashboard of the Large Hadron Collider at CERN is displayed as a PNG that’s always found at a static URL. The Pi just has to download the image and scale it to the 720-by-480 resolution of NTSC composite video. Another channel on my Geek TV displays NASA’s Astronomy Picture of the
Day. To download this, the Pi has to query an API that returns metadata, and then parse it using a JSON library to get the URL of the daily image. And getting the latest photo from the GOES East weather satellite means parsing HTML using the Beautiful Soup library. I also added some animations: One channel “plays” Conway’s Game of Life. Another channel cranks out increasingly higher resolution versions of the Mandlebrot set. And in between channels, the TV displays pixelated “static.” To finish up, I had to repair some leads to the CRT that had broken be cause of all the flexing that had happened while I was soldering in new connections. Once this was done, it was just a matter of stuffing everything inside the TV case and closing it back up. Now, to get my latest science updates, all I have to do is turn on and tune in.
3
1 2
4 I soldered wires to the ground [1] and on-off switch [2] to tap into the TVs power. For a control input, I disconnected the original tuner’s variable capacitor from the rest of the TV, leaving it attached to the knob shaft [3]. Then I soldered a wire between the capacitor and the touch sensor to feed a serial signal to the Pi Zero W [4].
18 SPECTRUM.IEEE.ORG MAY 2021
K
00
$5
AW AR
DS
U
P
TO
NE W
INNOVATE TRANSFORM OUR FUTURE R&D FUNDING PROGRAM The National Reconnaissance Office Director’s Innovation Initiative (DII) Program funds cutting-edge scientific research in a high-risk, high-payoff environment to discover innovative concepts and creative ideas that transform overhead intelligence capabilities and systems for future national security intelligence needs. The program seeks the brightest minds and breakthrough technologies from industry, academia, national laboratories, and U.S. government agencies. Visit the website for Broad Agency Announcement and Government Sources Sought Announcement requirements.
703.808.2769
www.nro.gov/Business-Innovation-Opportunities
COUNTRIES
OPINION, INSIGHT, AND ANALYSIS
Crosstalk WORLD
CHINA
INDIA
INDONESIA
Indonesia: 5.43 million
EGYPT
TONNES OF HARVESTED FISH
China: 47.6 million
India: 7.06 million
Egypt: 1.56 million
NORWAY
Norway: 1.35 million
JAPAN
Japan: 642,900
SOUTH KOREA
South Korea: 568,400
UNITED STATES
United States: 468,200
0
Farming Fish The 36-fold rise in aquaculture in the past 50 years is unparalleled in the history of food production
D
uring the past 50 years, the world’s population has grown about 2.1 times, and consumption has grown more; for some goods, much more. Between 1968 and 2018 (the most recent year for which we have global data), production of both primary energy and of steel rose 3.4 times, that of meat grew 3.5 times, and that of grain grew 2.6 times. The total number of cars on the world’s roads rose sixfold, and the number of revenue-generating passenger-kilometers flown (an industry metric) grew by about 24 times. But the rise in aquacultural production beats them all.
20 SPECTRUM.IEEE.ORG MAY 2021
20 million
40 million
Carp have been grown in inland ponds in China since ancient times and in parts of Europe since the Middle Ages. Fish have been farmed in Hawaii in ponds built with walls of lava rock in shallow coastal seas; some countries, notably Japan, also used such methods to produce crustaceans, mollusks, and algae. But until half a century ago, these practices were concentrated overwhelmingly in East Asia, particularly China, where the annual harvest (including all freshwater and marine animals and algae) was only about 3.4 million tonnes. That total more than doubled in the 1970s and grew even faster in the 1980s and 1990s before, inevitably, slowing down. Still, by 2018 the live-weight harvest of global aquaculture (20 percent of it protein) was about 114 million tonnes—36 times as much as in 1968. Algae accounted for slightly more than 30 million tonnes, fish from fresh water added about 51 million tonnes, and marine aquaculture about 31 million tonnes, bringing the 2018 harvest of fin fish, crustaceans, and mollusks to about 82 million tonnes. Meanwhile, the haul of wild-caught fish continued to rise until the late 1980s, finally reaching a
NUMBERS DON’T LIE BY VACLAV SMIL
AQUACULTURE OF FIN FISH, CRUSTACEANS, AND MOLLUSKS, 2018
World: 82.1 million
SOURCE: UN FOOD AND AGRICULTURE ORGANIZATION (FAO)
ORGANIZATION (FAO)
Aquaculture Capture Fisheries
100 80 Millions of tonnes
WORLD FISH FARMING GROWTH Aquaculture’s share of total fish consumption has risen from a rounding error to a majority in a relatively short time. SOURCE: UN FOOD AND AGRICULTURE
60 40 20 0 1960
PORTRAITS BY SERGIO ALBIAC
60 million
plateau of around 96 million tonnes a year. This means that about 45 percent of all protein coming from aquatic species now originates in ponds, lakes, pens, and cages where the species are grown, with or without feeding. China accounts for 58 percent of global fish- farming output, and its four traditional favorites— carp of the grass, silver, common, and bighead varieties—still dominate global aquaculture, accounting for just about a third of the total produced live weight. India is in a distant second place at about 7 million tonnes a year. Asia contributes 73 percent of the total, the Americas less than 5 percent, Europe less than 4 percent. Besides carp, the most commonly cultured fish are Nile tilapia, catfish, and Atlantic salmon (now grown not only in European and Eastern Canadian coastal waters, but also in the U.S. Pacific Northwest, Chile, and New Zealand). Whiteleg shrimp and red swamp crayfish are the most commonly cultured crustaceans; cupped oysters, Japanese carpet shells, and scallops lead in the mollusk category.
1970
80 million
About 45 percent of all protein coming from aquatic species now originates in ponds, lakes, pens, and cages where the species are grown.
1980
1990
2000
2010
2015
100 million
The industry’s principal challenge outside Asia is the consumers’ preference for salmon, cod, and tuna, all of them carnivorous. Plant-eating carp can be fed cheap grain or cereal pellets, but salmon or tuna will not grow and mature without ingesting fish oils and fish protein, which must be prepared by catching smaller, less valuable sardines, anchovies, and mackerel and converting them into fish feed. The gain:feed ratio for the Atlantic salmon has been reduced to as low as 1:1.2. Tuna farming, which is much more challenging, is just at the beginning stage, and cod farming is even less advanced. The industry is not without controversies. Densely stocked ponds and pens cause environmental problems, notably the release of organic materials, which promote algal blooms and the introduction of salmon that have been genetically modified to grow faster and use food more efficiently. But fish farming is now a global business worth more than US $250 billion a year, and it will keep expanding, introducing new species to more consumers willing to pay for their favorite fish, crustaceans, and mollusks.
MAY 2021 SPECTRUM.IEEE.ORG 21
CROSSTALK
INTERNET OF EVERYTHING BY STACEY HIGGINBOTHAM
Broadband Beat Cop The FCC’s new Democratic majority is poised to revisit the agency’s ability to regulate the Internet
Y
ou’re about to hear a lot about network neutrality again, now that the Biden administration is in the White House. Last time, the issue was ensuring that Internet service providers (ISPs) wouldn’t discriminate against different sources of legal traffic on their networks. This time around, you should expect the debate to be about whether to give the Federal Communications Commission enough legal standing to regulate the Internet. To be a cop on the broadband beat, so to speak. To give the FCC that power, activists, lawmakers, and regulators are turning to a big stick in the world of telecommunications—Title II of the 1996 Communications Act. The act separated services such as radio broadcasts, television, and the airwaves and wires over which they traveled into two categories: information services and services offered by common carriers. Information services, such as online publications, avoided heavy regulation by the FCC. Common carriers, such as voice telephony, weren’t so lucky. They became subject to policies governing how they set pricing, worked with other networks, and more. The 1996 law was an update to the original Communications Act of 1934, which created the FCC as a new regulatory body. One broad way to think about how the 1996 Act divided what required heavy regulation and what did not is to think about it in terms of infrastructure versus what travels over that infrastructure. In other words, the FCC would give increased scrutiny to infrastructure—wires and airwaves—under Title II of the law. Everything that traveled over those wires or airwaves required less oversight and was lumped under Title I. Back in 2002, broadband was classified as a Title I information service—and therefore not subject to the heavy regulations of a common carrier—as part of a formal FCC order, a decision that was later
22 SPECTRUM.IEEE.ORG MAY 2021
The pandemic showed us that broadband is essential.
upheld in 2005 by the U.S. Supreme Court. But after a slew of incidents where broadband providers started interfering with traffic on their networks to throttle or slow down certain traffic sources, the FCC decided to reexamine how to prevent AT&T, Comcast, Verizon, and smaller ISPs from abusing their power. In 2015, the FCC reclassified ISPs under Title II after an earlier attempt was struck down by an appeals court in 2010. Shortly afterward, under the Trump administration, the agency issued a formal order to return broadband to the information service category. Now, the Democrat-led FCC, along with lawmakers including Senator Ed Markey of Massachusetts, is interested in using Title II to again strengthen the FCC’s regulatory authority. Reclassifying broadband service as a common carrier will empower the agency to collect more data from ISPs about where they provide broadband and the quality of that service. Current requirements are limited to census blocks that are less granular than the service quality for individual users, and don’t require the ISP to discuss the price consumers pay. Reclassification could also lead to reformation of the universal service rules that govern fees paid by ISPs that fund projects to bring voice and broadband to Americans who might otherwise not have access. Most importantly, it will give the FCC the power the agency needs to stop broadband providers from using their monopoly power to extract fees while delivering poor service to consumers. The pandemic showed us that broadband is essential. Between Comcast trying to implement data caps or kids using their school’s Wi-Fi while sitting in the parking lot for remote schooling, we’re seeing how vital more FCC regulation is. So, while you may feel we’ve already had this debate too many times before, the fact remains that we need a cop on this particular beat.
Photo-illustration by Edmon de Haro
MACRO & MICRO BY MARK PESCE
Fine-Grained Power The recent power shortfalls in Texas exposed the need for better control
R
ecently, my electricity provider started offering email notifications when the price of electricity rises above a preset threshold. As amply demonstrated during F ebruary’s cold-snap-induced grid collapse in Texas, the spot price of electricity can rise precipitously when demand exceeds supply, which in that case led to some unlucky folks receiving bills for tens of thousands of dollars. It’s unlikely that I’d ever end up in a similar predicament, but on a hot summer’s day the wholesale price of electricity can rise to more than 100 times what’s normal—and a notification would immediately guide me to unplug my most power-hungry devices, which include two PCs equipped with beefy GPUs. A price spike never lasts more than an hour or two, but in that time, I’ll have saved money and removed some of the load from an overtaxed grid.
Illustration by Harry Campbell
Nowadays, almost anything plugged into an outlet could monitor its own power.
Among those essential networks that we never notice until they fail, electricity grids have recently gotten “smart,” at least in places. When I signed up with it, my electric utility replaced a 40-year-old analog meter with a digital meter that maintains Internet connectivity with my provider, allowing it to monitor what I draw from the grid while sending me information to optimize my usage. Tracking usage for each household offers the provider a fine-grained picture of where the energy is going. As that householder, though, I’d prefer even more detail: Where exactly is the power I pay for being used? And could an automated system within my home continuously optimize that usage? Technology solutions for energy monitoring have fallen dramatically in cost: Nowadays, almost anything plugged into an outlet could monitor its own power consumption for about US $1 in components, which translates to perhaps $5 retail. For any electrical appliance that costs more than $100, adding this feature is a no-brainer. Many people would want to manage their power budgets and, as long as their provider is still burning coal or n atural gas to make electricity, reduce their carbon footprints. In principle, smart appliances could form a spot-trading network of their own within the home, queuing for priority access to electricity based on the prevailing price, demand, and the owner’s preferences. An average household should be able to cut its electricity costs and carbon footprint by an appreciable amount simply by delegating to appliances the agency to manage their own operations. Indeed, why not bring all the homes in the neighborhood into this micromarket? During daylight hours, some homeowners could sell excess solar power to neighbors who could use that cheap, local power to run a load of wash or charge an electric vehicle. After sunset, people who were buyers earlier might decide to sell some of the energy stored in those same vehicles back to their neighbors. For more than a century, power engineers have designed electricity grids around centralized generation and distribution. But that arrangement is swiftly changing. Adding intelligence at the edge, as so many utilities are doing, will help transform the grid into a decentralized network of providers, consumers, and algorithms, all working in concert to smooth the flow of electrons to where they are most needed.
MAY 2021 SPECTRUM.IEEE.ORG 23
V E N U S C A L L I N G S I L I C O N C A R B I D E
R A DIO
C I R C U I T S
C AN
T A K E
H EAT
T H E
N E E D E D P H O N E F R O M
T O H O M E
O U R
H E L L I S H S I S T E R
P L A NET
By Alan Mantooth, Carl-Mikael Zetterling & Ana Rusu
Illustration by NoEmotion
MAY 2021 SPECTRUM.IEEE.ORG 25
26 SPECTRUM.IEEE.ORG MAY 2021
464°C
Average surface temperature
92
Megapascals surface pressure
96.5% CO2 3.5% N 0.015% SO2 Atmosphere
5,832 hours
Length of day (243 Earth days)
Venera 13
Longest-surviving lander (at least 127 minutes)
tems might not give them more than an extra 24 hours. The answer is a semiconductor that combines two plentiful elements, carbon and silicon, in a 1:1 ratio—silicon carbide. SiC can withstand extremely high temperatures and still work just fine. Scientists at the NASA Glenn Research Center have already operated SiC circuits for more than a year at 500 °C, demonstrating not only that they can take the heat but can do so over the kinds of lifetimes a Venus lander will need. Silicon carbide is already making its mark in power electronics for solar inverters, electric-vehicle motor-drive electronics, and advanced smart-grid switch gear. But creating SiC circuits that can control a rover on the hellscape of Venus and send data from there to Earth will test this material to its limits. If it succeeds, we’ll get more than just a mobile outpost in one of the least- hospitable spots in the solar system. We’ll also get insight into how to move wireless sensors into places on Earth that they’ve never gone before—on the blades of jet engines and natural-gas turbines, on the heads of deep oil-well drills, and inside a host of high-temperature, high-pressure industrial fabrication processes. The ability to locate electronics in these places has a meaningful chance of lowering the operating and maintenance costs of equipment while improving the performance and safety of both instruments and people in industrial settings. In fact, our team, with members at KTH Royal Institute of Technology, in Stockholm, and at the University of Arkansas, in Fayetteville, believe silicon carbide circuits can take us there and beyond, to applications we have not yet imagined.
S
ILICON CARBIDE IS not a new
material by any means. Largescale production is credited to Edward Goodrich Acheson in 1895. The American chemist was attempting to create artificial diamonds when his experiments produced crystals of SiC. The compound was first successfully used as an electronic material in 1906, when Henry Harrison Chase Dunwoody invented the SiC radio detector. To this day it is regarded as the first commercial semiconductor device. However, large SiC crystals are notoriously difficult to manufacture in
JPL-CALTECH/NASA
T
HERE WERE FEW bright spots in the pandemic summer of 2020. One of the most dazzling was the flight of U.S. astronauts to the International Space Station and their safe return to Earth aboard a commercial spacecraft from SpaceX. This demonstration was significant for many reasons, one of which was that it suggested a future in which NASA, freed from the demands of getting people to low Earth orbit, could aim much farther. Perhaps as far as Venus. Excitement over a possible mission to Venus was stoked by the (now somewhat disputed) discovery of phosphine gas—a possible sign of microbial life—in that planet’s atmosphere. But the second planet from the sun has such an extreme environment that the longest-lasting lander, the Soviet Venera 13, was able to send data for only 2 hours and 7 minutes. The average surface temperature on Venus is 464 °C, the atmosphere is dense with highly corrosive droplets of sulfuric acid, and the atmospheric pressure at the surface is about 90 times that of Earth. Yet scientists think of Venus as our home world’s twin. The size and mass of these two planets are very close, of course. And evidence points to similar early days: For as long as 3 billion years, Venus may have had massive oceans just like we do here on Earth, and therefore, perhaps, there was life. What cataclysmic events led to Venus’s loss of water? Planetary scientists would love to know, because it might inform our own fate as the climate changes. To solve this and Venus’s other riddles, we’ll need several very capable robotic landers. But can we build machines—complete with instrumentation, communications, controllability, and mobility—that can survive such a hostile environment not just for hours but for months to years? We can. Materials technology has advanced enough since the 1960s, when the former Soviet Union began launching its Venera series of landers to Venus, to ensure that the outer hull and mechanics of a future lander will be able to last for months. But what about those tender electronics? Today’s silicon-based systems would not last a day under Venus conditions. (We mean an Earth day, of course. A Venusian day is 243 Earth days.) Even adding active cooling sys-
①
④
⑤
UNIVERSITY OF ARKANSAS
The Vulcan II is a chip with multiple silicon carbide analog and digital circuits for testing at 500 °C. We’ve made more than 40 circuits so far with Vulcan II and its predecessor.
repeatable fashion, and it wasn’t until the late 1990s that engineers invented equipment that enabled the growth of crystals good enough to use to fabricate power transistors. These initial silicon carbide wafers were just 30 millimeters across, but the industry has slowly progressed to 50-, 75-, 100-, 150- and now 200-mm wafer diameters, making the devices more economical. Research and progress has steadily increased over the past 20 years to the point that SiC power semiconductor devices can now be purchased commercially. Silicon carbide boasts some very attractive properties as a semiconducting material. The first of these is a critical electric-field strength almost 10 times that of silicon. That property is basically the point at which a material breaks down and begins conducting electricity uncontrollably, sometimes with explosive results. So if you’ve got a silicon device and a silicon carbide device of the
② ⑥
⑧
⑦
❶ Ring oscillator ❸ RS 485 receiver successive 8-bit adders and ❷ 8-bit ❹ 4-bit approximation analog-tomultipliers digital converters and 4-bit ramp analog-todigital converters
❺ 555 timer
same scale, the one made of silicon carbide can handle 10 times as much voltage. Alternatively, if the two transistors were built to handle the same voltage, the silicon carbide device could be physically much smaller. That size difference translates to a power-consumption advantage. For the same “breakdown voltage” (1,200 volts, say), a SiC transistor has
From the standpoint of scientists hoping to explore other planets, the radio is arguably the most vital system.
③
3-stage operational ❻ amplifier
❼ DC-to-DC converter gate ❽ Integrated drivers
1/200th to 1/400th the “on” resistance of a silicon transistor, and therefore lower power losses. This smaller size also allows for higher switching frequency in a power converter, which would mean smaller, lighter, less costly capacitors and inductors. Silicon carbide’s second amazing attribute is thermal conductivity: As the SiC heats up due to electrical conduction, the heat can be rapidly removed, prolonging the lifetime of a device. In fact, among wide-bandgap semiconductors, the thermal conductivity of SiC is second only to that of diamond. This property lets you connect a high-power silicon carbide transistor to the same size heat sink you’d use on a much lower-power silicon component and still get a fully functional, long-lasting device. A third property, the most relevant to operating on Venus, is SiC’s very low intrinsic concentration of charge carriers at room temperature. The intrinsic
MAY 2021 SPECTRUM.IEEE.ORG 27
Antenna Low-noise amplifier
Critical Component
Mixer
Any Venus lander will need an RF transceiver to communicate with Earth. One of the most essential components of that system is the mixer. On the receiving end, it converts the 59-megahertz carrier signal to a 500-kilohertz frequency that is more easily digitized and processed. On the transmitting side, it does the reverse. The heart of the mixer circuit is a silicon carbide bipolar junction transistor, designed for operation up to 500 °C.
carrier concentration corresponds to how many charge carriers heat makes available to conduct electricity. (Doping a semiconductor with atoms of another element can increase the available charge carriers. But the intrinsic concentration is what’s there without doping.) You might think that a low value here, particularly one that’s lower than silicon’s, would be a bad thing. But that’s not the case if we want to operate at high temperatures. Here’s why. The reason silicon stops working as a semiconductor when tem-
Intrinsic carrier concentration
Variable–gain amplifier
Circulator Low-pass filter
Local oscillator
12-bit digital-to-analog converter
Power amplifier Mixer
Digital processing unit
RF transceiver
peratures climb is not because it melts, or burns, or anything that dramatic. Instead, the transistors start to become flooded with thermally generated charge carriers. Heat gives some electrons enough energy to boil up out of the valence band, where they are bound to atoms, and into the conduction band, leaving behind positively charged holes. The separated electrons and holes can now contribute to conduction. At moderate temperatures, say 250 to 300 °C for silicon, this just makes transistors leak current and become noisy. But at higher tempera-
1.00E+15 1.00E+10 Silicon 1.00E+05
1.00E-05 Silicon carbide 27
80
160
280
500
1000
Temperature, Celsius As temperature increases, so does a semiconductor’s intrinsic carrier concentration—the number of electrons with enough energy to contribute to conduction. After a certain concentration is reached, a transistor is effectively flooded with charge carriers and will not turn off. This happens for most silicon devices at around 250 °C, but silicon carbide devices can still switch at 1,000 °C.
28 SPECTRUM.IEEE.ORG MAY 2021
tures, the intrinsic carrier concentration exceeds whatever contribution doping has provided, and you can no longer turn the transistors off—they become like switches stuck in the “on” position. By contrast, SiC, with its wider bandgap and fewer intrinsic charge carriers, has far more temperature headroom before the onset of “transistor flooding” occurs, enabling it to continue to switch well above 800 °C. Collectively, these properties allow SiC to function at a higher voltage, power, and temperature than silicon can. And, even for temperatures where silicon is able to function, SiC often outperforms it because the devices can be switched at higher frequencies with lower losses. Put it all together and you have devices that are more efficient and rugged, and circuits and systems that are smaller, lighter weight, and capable of surviving in the Venusian environment.
W
1.00E+00
1.00E-10
12-bit analog-to-digital converter
HILE A FUTURE Venus lander will need its share of high-voltage power transistors, most of its circuits—in processors, sensors, and radios—need to be of the low-voltage variety. Those are much less developed in silicon carbide than in silicon, but thanks to a problem of packaging, we’ve made a start. As discrete silicon carbide power devices found commercial usage, engineers recognized the need to reduce electrical parasitics—unwanted resistance, inductance, and capacitance, which wastes power. One way is to better inte-
grate, through advanced packaging, the a gate driver, which directly controls a control, drive, and protection circuits power transistor through its input terwith the power devices. In silicon power minal, or gate. We’ve now completed electronics, these circuits are situated on several versions of the circuit—which printed circuit boards (PCBs). But at the can be packaged with (or even on top of) higher frequencies SiC power transistors the power device—and tested them at can achieve, the PCB’s parasitics may be Venus-like temperatures. This circuit, as too great, leading to excessive noise. well as later versions, enabled very high Packaging or even integrating these cir- fidelity control of the power devices, cuits with the power devices would maximizing efficiency while minimizing remove the noise. But the latter option electromagnetic interference. The bigwould mean making these circuits out of gest challenge was to achieve a design silicon carbide. that could adapt to changing conditions At room temperature, silicon carbide and even account for the effects of aging, is not a natural choice for low-voltage which are bound to occur in the harsh microelectronics for several reasons. Per- conditions on Venus. haps the most important is that the voltage can’t really be all that low, so neither ATE DRIVERS ARE important, is power consumption. Silicon’s small but from the standpoint of scibandgap means you can power microentists hoping to explore other electronics with as little as 1 V. But silicon planets, the radio is arguably carbide’s bandgap is nearly three times as the most vital system. After all, there’s large. Therefore, the minimum voltage no point in sending a package of scienneeded to push current through a transis- tific instruments to another planet if you tor—the threshold voltage—is also can’t get the data back to Earth. greater. We generally use 15 V to supply Compact, rugged radio systems may be our “low-voltage” SiC microelectronics. even more crucial to future planetary misInvestigators around the world have sions, because they could carry data attempted low-voltage microelectronics within the rover itself, replacing some of in SiC for more than 20 years, at first with limited success. During the past 10 years, however, researchers at our universities as well as at Cree, Fraunhofer Institute for Integrated Systems and Device Technology, Purdue University, NASA Glenn, University of Maryland, and Raytheon UK have made some breakthroughs. One of the first key microelectronic circuits built by the Arkansas team was
KTH ROYAL INSTITUTE OF TECHNOLOGY
G
We’ve designed, built, and tested some 40 different circuits for 500°C conditions.
the thousands of point-to-point wires in these machines. Eliminating wires in favor of wireless command and control saves substantial mass, a vital commodity on a 40-million-kilometer trip. So much of our most recent effort has been the design and testing of the components of a silicon carbide–based interplanetary radio transceiver. Silicon carbide would be nobody’s first choice for, say, a 5G radio operating on Earth. For one thing, at room temperature its charge-carrier mobility—part of what sets an upper limit to the frequencies a semiconductor can amplify—is lower than silicon’s. But at Venus surface temperatures, silicon no longer functions at all, so it makes sense to try to adapt silicon carbide to the task. With regard to radio frequencies, silicon carbide does have one thing in its favor. The sparsity of charge carriers means devices made from the material have low parasitic capacitances. In other words, there are few charges around, so those charges are unlikely to interact in ways that sap a device’s performance. The transceiver architecture we’re targeting is called a low-intermediatefrequency heterodyne. (In Greek, hetero means different, and dyne means power.) To unpack what that means, let’s follow an incoming signal through the receiver side of the system. Radio signals from the antenna are boosted by a low-noise amplifier, then fed to a mixer. The mixer combines the received signal with another frequency close to the signal’s carrier frequency. This
MAY 2021 SPECTRUM.IEEE.ORG 29
30 SPECTRUM.IEEE.ORG MAY 2021
Heat flows through a silicon carbide gate-driver chip during testing.
ceramic board instead. Chips attach to this rock-hard board with gold wires, instead of aluminum, which would soon soften. Silver interconnects, some coated in titanium, link the components into a circuit instead of copper traces, which would pull away from the PCB. Inductors are made on the board as spirals of gold. (Yes, these circuits would be pretty expensive.)
and take those to indicate that our circuits and devices could operate for longer. Notably, NASA Glenn Research Center recently reported silicon carbide ICs, with nearly 200 transistors per chip, that operated for a full 60 days in that center’s Venus environment chamber. The chamber subjected the transistors to 9.3 megapascals of pressure, 460 °C RUCIAL AS THE mixer is, a heat, and the planet’s particular caustic future Venus rover will need a atmosphere. Not one of those transistors lot more than that. So far, succumbed, suggesting they could have between the University of soldiered on much longer had more time Arkansas and KTH, we’ve designed, in the chamber been available. built, and tested some 40 different cirThere is still a lot of work to be done. cuits for 500 °C conditions. These We need to focus on integrating the varcircuits include other RF and analog ious circuits that have been developed parts of the transceiver and many of the and on improving the yields of working digital circuits needed for processing circuits. We must still develop more cirdata from the transceiver and future cuits and prove that they can operate planetary-science sensors. Some of together for months or years with the these will be familiar to many engineers, needed stability at Venus surface temsuch as a 555 timer, an 8-bit analog-to- peratures. This last point is particularly digital converter and digital-to-analog important if silicon carbide radios and converter, a phase-locked-loop circuit, other low-power circuits are ever to and a library of Boolean logic circuits. make sense in commercial applications We admit that since these are university- such as jet and natural-gas turbines. produced parts made in small numbers, With enough effort and priority, these long-term testing has not yet been could be years away, not decades. attempted. Our labs have done at most Will silicon carbide circuits be ready a week or two of operation at high tem- for a future Venus mission? You might perature. However, we’re encouraged by more reasonably argue that the mission other groups’ extended experiments won’t be ready without them.
C
UNIVERSITY OF ARKANSAS
mixing creates a signal at two new intermediate frequencies, one higher than the carrier and one lower. The higher frequency is then eliminated by a lowpass filter. The remaining intermediate frequency—which is more suitable to processing—is amplified and then digitized by an analog-to-digital converter, which delivers the resulting bits, representing the received signal, to a digital processing unit. How we actually implemented the RF circuits that performed all these functions was determined by the high- frequency performance of silicon carbide bipolar junction transistor (BJT) technology developed in-house at KTH. That technology resulted in the fundamental RF circuits needed to build a transceiver to send and receive 59-megahertz signals—a balance between the transistor’s high-frequency limits and the constraints of the circuit’s passive components, which get more restrictive at lower frequencies. (This frequency is roughly in the range of the 80 MHz that the Venera landers used. A modern Venus mission is likely to be sending its data first to a satellite orbiting the planet, which could then use NASA’s deep-space frequencies to carry the data home.) One real make-or-break part of the transceiver is the mixer, which down converts the 59-MHz signal to a 500-kilohertz intermediate frequency. The heart of our mixer is a SiC bipolar junction transistor, with both the incoming 59-MHz RF signal and a 59.5‑MHz signal as its inputs. The output, from the transistor’s collector terminal, connects to a network of capacitors and resistors— all designed to withstand 500 °C—that filter out the high frequency, leaving only the 500-kHz intermediate frequency. Compared to the low-frequency analog and digital circuits that come after the mixer, the RF circuits brought challenges at all stages of development, including the absence of accurate models of the transistor, issues with matching impedances to ensure the most signal gets through, and the reliability of resistors, capacitors, inductors, and PCBs. Those PCBs, by the way, look nothing like what you’re used to. The ubiquitous FR-4 circuit boards that underpin everything from handheld gadgets to high-end servers would quickly sag and come apart under Venusian conditions. So we use what’s called a low-temperature cofired
Because financial independence should be for the long term.
Preserve your nest egg and your quality of life.
IEEE Member Long-Term Care Insurance To learn more, call 1-800-588-7421 or visit IEEEinsurance.com/IEEELTC The IEEE Member Group Insurance Program is administered by Mercer Health & Benefits Administration LLC. In CA d/b/a Mercer Health & Benefits Insurance Services LLC AR Insurance License #100102691 • CA Insurance License #0G39709 92617 (5/21) Copyright 2021 Mercer LLC. All rights reserved.
32 SPECTRUM.IEEE.ORG MONTH 2021
Building an AI That Feels AI systems with emotional intelligence could learn faster and be more helpful By Mary Czerwinski, Javier Hernandez & Daniel McDuff
Photo by First Illustration by Lastname Eddie Guy
MAY 2021 SPECTRUM.IEEE.ORG 33
In the past year, have you found yourself under stress? Have you ever wished for help coping? Imagine if, throughout the pandemic, you’d had a virtual therapist powered by an artificial intelligence (AI) system, an entity that empathized with you and gradually got to know your moods and behaviors. Therapy is just one area where we think an AI system that can recognize and interpret emotions could offer great benefits to people. Our team hails from Microsoft’s Human Understanding and Empathy group, where our mission is to imbue technology with emotional intelligence. Why? With that quality, AI can better understand its users, more effectively communicate with them, and improve their interactions with technology. The effort to produce emotionally intelligent AI builds on work in psychology, neuroscience, human-computer interaction, linguistics, electrical engineering, and machine learning. Lately, we’ve been considering how we could improve AI voice assistants such as Alexa and Siri, which many people now use as everyday aides. We anticipate that they’ll soon be deployed in cars, hospitals, stores, schools, and more, where they’ll enable more personalized and meaningful interactions with technology. But to achieve their potential, such voice assistants will require a major boost from the field of affective computing. That term, coined by MIT professor Rosalind W. Picard in a 1997 book by the same name, refers to technology that can sense, understand, and even simulate human emotions. Voice assistants that feature emotional intelligence should be more natural and efficient than those that do not. Consider how such an AI agent could help a person who’s feeling overwhelmed by stress. Currently, the best option might be to see a real human psychologist who, over a series of costly consultations, would discuss the situation and teach relevant stress-management skills. During the sessions, the therapist would continually evaluate the person’s responses and use that information to shape what’s discussed, adapting both content and presentation in an effort to ensure the best outcome. While this treatment is arguably the best existing therapy, and while technology is still far from being able to replicate that experience, it’s not ideal for some. For example, certain people feel uncomfortable discussing their feelings with therapists, and some find the process stigmatizing or time-
34 SPECTRUM.IEEE.ORG MAY 2021
consuming. An AI therapist could provide them with an alternative avenue for support, while also conducting more frequent and personalized assessments. One recent review article found that 1 billion people globally are affected by mental and addictive disorders; a scalable solution such as a virtual counselor could be a huge boon. There’s some evidence that people can feel more engaged and are more willing to disclose sensitive information when they’re talking to a machine. Other research, however, has found that people seeking emotional support from an online platform prefer responses coming from humans to those from a machine, even when the content is the same. Clearly, we need more research in this area. In any case, an AI therapist offers a key advantage: It would always be available. So it could provide crucial support at unexpected moments of crisis or take advantage of those times when a person is in the mood for more analytical talk. It could potentially gather much more information about the person’s behavior than a human therapist could through sporadic sessions, and it could provide reminders to keep the person on track. And as the pandemic has greatly increased the adoption of telehealth methods, people may soon find it quite normal to get guidance from an agent on a computer or phone display. For this kind of virtual therapist to be effective, though, it would require significant emotional intelligence. It would need to sense and understand the user’s preferences and fluctuating emotional states so it could optimize its communication. Ideally, it would also simulate certain emotional responses to promote empathy and better motivate the person. The virtual therapist is not a new invention. The very first example came about in the 1960s, when Joseph Weizenbaum of MIT wrote scripts for his ELIZA natural-language- processing program, which often repeated users’ words back to them in a vastly simplified simulation of psychotherapy. A more serious effort in the 2000s at the University of Southern California’s Institute for Creative Technologies produced SimSensei, a virtual human initially designed to counsel military personnel. Today, the most well-known example may be Woebot, a free chatbot that offers conversations based on cognitive behavioral therapy. But there’s still a long way to go before we’ll see AI systems that truly understand the complexities of human emotion. Our group is doing foundational work that will lead to such sophisticated machines. We’re also exploring what might happen if we build AI systems that are motivated by something approximating human emotions. We argue that such a shift would take modern AI’s already impressive capabilities to the next level.
About 1 billion people globally are affected by mental disorders; a scalable solution such as an AI therapist could be a huge boon. Illustration by Chris Philpot
O
nly a decade ago, affective computing required custom-made hardware and software, which in turn demanded someone with an advanced technical degree to operate. Those early systems usually involved awkwardly large sensors and cumbersome wires, which could easily affect the emotional experience of wearers. Today, high-quality sensors are tiny and wireless, enabling unobtrusive estimates of a person’s emotional state. We can also use mobile phones and wearable devices to study visceral human experiences in real-life settings, where emotions really matter. And instead of short laboratory experiments with small groups of people, we can now study emotions over time and capture data from large populations “in the wild,” as it were.
Earlier studies in affective computing usually measured emotional responses with a single parameter, like heart rate or tone of voice, and were conducted in contrived laboratory settings. Thanks to significant advances in AI—including automated speech recognition, scene and object recognition, and face and body tracking—researchers can do much better today. Using a combination of verbal, visual, and physiological cues, we can better capture subtleties that are indicative of certain emotional states. We’re also building on new psychological models that better explain how and why people express their emotions. For example, psychologists have critiqued the common notion that certain facial expressions always signal certain emotions, arguing that the meaning of expressions like smiles and frowns varies greatly according to context, and also reflects individual and cultural differences. As these models continue to evolve, affective computing must evolve too. Raising inner brow This technology raises a number of societal Raising outer brow issues. First, we must think about the privacy Lowering brow implications of gathering and analyzing peoWidening eyes ple’s visual, verbal, and physiological signals. Raising cheeks One strategy for mitigating privacy concerns Tightening eyes is to reduce the amount of data that needs to Wrinkling nose leave the sensing device, making it more diffiRaising lips cult to identify a person by such data. We must Pulling lips (smile) also ensure that users always know whether Dimpling they’re talking to an AI or a human. AdditionDepressing lips ally, users should clearly understand how their Raising chin data is being used—and know how to opt out Pressing lips or to remain unobserved in a public space that Tightening lips might contain emotion-sensing agents. Parting lips As such agents become more realistic, Opening mouth we’ll also have to grapple with the “uncanny Closing eyes valley” phenomenon, in which people find that somewhat lifelike AI entities are creepier than more obviously synthetic creatures. But before we get to all those deployment chalHeart rate: 88 beats per lenges, we have to make the technology work.
1.
①
②
minute
Voice-activity detection
③
Speech transcription
Positive
“Hello”
Negative Neutral
Voice-feature extraction
To predict someone’s emotional state, it’s best to combine readouts. In this example, software that analyzes facial expressions detects visual cues, tracking the subtle muscle movements that can indicate emotion (1). A physiological monitor detects heart rate (2), and speech-recognition software transcribes a person’s words and extracts features from the audio (3), such as the emotional tone of the speech.
A
s a first step towa rd an AI system that can support people’s mental health and well-being, we created Emma, an emotionally aware phone app. In one 2019 experiment, Emma asked users how they were feeling at random times throughout the day. Half of them then got an empathetic response from Emma that was tailored to their emotional state, while the other half received a neutral response. The result: Those participants who interacted with the empathetic bot more frequently reported a positive mood. In a second experiment with the same cohort, we tested whether we could infer people’s moods from basic mobile-phone data and whether suggesting appropriate wellness activities would boost the spirits of those feeling glum. Using just location (which gave us
MAY 2021 SPECTRUM.IEEE.ORG 35
Our “focus agent” aimed to boost productivity by helping users schedule time to work on important tasks and helping them adhere to their plans. A camera (1) and computer
software (2) kept track of the user’s behavior. The sensing framework (3) detected the number of people in view and the user’s position in front of the computer screen,
estimated the user’s emotional state, and also kept track of the user’s activity within various applications. The agent app (4) controlled the focus agent avatar
that engaged the user in conversation, using an AIpowered conversation bot (5) that drew on a variety of dialogue models to respond to the situation as appropriate.
① CAMERA Facial expressions
AGENT INTERACTIONS • Dialogue
FACE TRACKER
Face count
Head gesture
• Avatar video
• Open windows APP TRACKER
② COMPUTER
③ SENSING
• Keyboard • M ouse activity
FRAMEWORK
the user’s distance from home or work), time of day, and day of the week, we were able to predict reliably where the user’s mood fell within a simple quadrant model of emotions. Depending on whether the user was happy, calm, agitated, or sad, Emma responded in an appropriate tone and recommended simple activities such as taking a deep breath or talking with a friend. We found that users who received Emma’s empathetic urgings were more likely to take the recommended actions and reported greater happiness than users who received the same advice from a neutral bot. We collected other data, too, from the mobile phone: Its built-in accelerometer gave us information about the user’s movements, while metadata from phone calls, text messages, and calendar events told us about the frequency and duration of social contact. Some technical difficulties prevented us from using that data to predict emotion, but we expect that including such information will only make assessments more accurate.
I
n another area of research, we’re trying to help information workers reduce stress and increase productivity. We’ve developed many iterations of productivity support tools, the most recent being our work on “focus agents.” These assistants schedule time on users’ calendars to focus on important tasks. Then they monitor the users’ adherence to their plans, intervene when distractions pop up, remind them to take breaks when appropriate, and help them
36 SPECTRUM.IEEE.ORG MAY 2021
• F irst-time use • Morning • S chedule tasks
EVENTS
• R amp up (task about to start)
• U nlock computer
• Applications
DIALOGUE MODELS
• S chedule task • Open window
• R amp down (task about to end)
↓
• D istraction event
CALENDAR
• End of day
④ AGENT
APP
⑤ CONVERSATION
BOT
reflect on their daily moods and goals. The agents access the users’ calendars and observe their computer activity to see if they’re using applications such as Word that aid their productivity or wandering off to check social media. To see whether emotional intelligence would improve the user experience, we created one focus agent that appeared on the screen as a friendly avatar. This agent used f acial-expression analysis to estimate users’ emotions, and relied on an AI-powered dialogue model to respond in appropriate tones. We compared this avatar agent’s impact with that of an emotionless text-based agent and also with that of an existing Microsoft tool that simply allowed users to schedule time for focused work. We found that both kinds of agents helped information workers stay focused and that people used applications associated with productivity for a larger percentage of their time than did their colleagues using the standard scheduling tool. And overall, users reported feeling the most productive and satisfied with the avatar-based agent. Our agent was adept at predicting a subset of emotions, but there’s still work to be done on recognizing more nuanced states such as focus, boredom, stress, and task fatigue. We’re also refining the timing of the interactions so that they’re seen as helpful and not irritating. We found it interesting that responses to our empathetic, embodied avatar were polarized. Some users felt comforted by the interactions, while others found the avatar to be a distraction from their work. People expressed a wide range of prefer-
ences for how such agents should behave. While we could theoretically design many different types of agents to satisfy many different users, that approach would be an inefficient way to scale up. It would be better to create a single agent that can adapt to a user’s communication preferences, just as humans do in their interactions. For example, many people instinctively match the conversational style of the person they’re speaking with; such “linguistic mimicry” has been shown to increase empathy, rapport, and prosocial behaviors. We developed the first example of an AI agent that performs this same trick, matching its conversational partner’s habits of speech, including pitch, loudness, speech rate, word choice, and statement length. We can imagine integrating such stylistic matching into a focus agent to create a more natural dialogue. We’re always talking with Microsoft’s product teams about our research. We don’t yet know which of our efforts will show up in office workers’ software within the next five years, but we’re confident that future Microsoft products will incorporate emotionally intelligent AI.
A
If an AI agent was motivated by fear, curiosity, or delight, how would that change the technology and its capabilities? feedback along the way. In a stressful situation, such as speeding down the highway during a rainstorm, the person might feel his heart thumping faster in his chest as adrenaline and cortisol course through his body. These changes are part of the person’s fight-or-flight response, which influences decision making. The driver doesn’t have to actually crash into something to feel the difference between a safe maneuver and a risky move. And when he exits the highway and his pulse slows, there’s a clear correlation between the event and the response. We wanted to capture those correlations and create an AI agent that in some sense experiences fear. So we asked people to steer a car through a maze in a simulated environment, measured their physiological responses in both calm and stressful moments, then used that data to train an AI driving agent. We programmed the agent to receive an extrinsic reward for
Negative/unpleasant
Illustrations by Chris Philpot
Positive/pleasant
i systems that can predic t and respond to human emotions are one thing, but what if an AI system could actually experience something akin to human emotions? If an agent was motivated by fear, curiosity, or delight, how would that change the technology and its capabilities? To explore this idea, we trained agents that had the basic emotional drives of fear and happy curiosity. With this work, we’re trying to address a few problems in a field of AI called reinforceGot a minute to enter your emotional ment learning, in which an AI agent learns how state rating right now? to do a task by relentless trial and error. Over millions of attempts, the agent figures out the best actions and strategies to use, and if it sucHigh energy cessfully completes its mission, it earns a reward. Reinforcement learning has been used Stressed Nervous Alert to train AI agents to beat humans at the board game Go, the video game StarCraft II, and a Tense Excited type of poker known as Texas Hold’em. While this type of machine learning works Upset Happy well with games, where winning offers a clear reward, it’s harder to apply in the real world. Sad Content Consider the challenge of training a selfdriving car, for example. If the reward is getLethargic Calm ting safely to the destination, the AI will spend a lot of time crashing into things as it Depressed Fatigued Relaxed Serene tries different strategies, and will only rarely Low energy succeed. That’s the problem of sparse external rewards. It might also take a while for the AI to figure out which specific actions are Submit most important—is it stopping for a red light or speeding up on an empty street? Because You’re the best, thanks! the reward comes only at the end of a long sequence of actions, researchers call this the credit-assignment problem. In one early experiment with Emma, an emotionally aware phone Now think about how a human behaves app, users were asked to rate their emotional state several times while driving. Reaching the destination safely throughout the day, using a quadrant model of emotions. is still the goal, but the person gets a lot of
MAY 2021 SPECTRUM.IEEE.ORG 37
Normalized amplitude
Open space (low stress)
Sharp turn and collision (high stress)
Open space (low stress)
1
0
-1 280
285
290
300
295
305
315
310
320
325
330
Normalized amplitude
1
0
-1 50
100
150
200
250
300
350
400
Time, seconds
We recorded the blood volume per pulse of test subjects while they drove through a virtual maze. In this example, the subject’s blood volume decreases between seconds 285 and 300. During that period, the driver collided with a wall while turning sharply to avoid another obstacle. This data was used to train an AI agent, which was given the objective of minimizing such stressful situations.
exploring a good percentage of the maze, and also an intrinsic reward for minimizing the emotional state associated with dangerous situations. We found that combining these two rewards created agents that learned much faster than one that received only the typical extrinsic reward. These agents also crashed less often. What we found particularly interesting, though, is that an agent motivated primarily by the intrinsic reward didn’t perform very well: If we dialed down the external reward, the agent became so risk averse that it didn’t try very hard to accomplish its objective. During another effort to build intrinsic motivation into an AI agent, we thought about human curiosity and how people are driven to explore because they think they may discover things that make them feel good. In related AI research, other groups have captured something akin to basic curiosity, rewarding agents for seeking novelty as they explore a simulated environment. But we wanted to create a choosier agent that sought out not just novelty but novelty that was likely to make it “happy.” To gather training data for such an agent, we asked people to drive a virtual car within a simulated maze of streets, telling them to explore but giving them no other objectives. As they drove, we used facial-expression analysis to track smiles that flitted across their faces as they navigated successfully through tricky parts or unexpectedly found the exit of the maze. We used that data as the basis for the intrinsic reward function, meaning that the agent was taught to maximize situations that
38 SPECTRUM.IEEE.ORG MAY 2021
would make a human smile. The agent received the external reward by covering as much territory as possible. Again, we found that agents that incorporated intrinsic drive did better than typically trained agents—they drove in the maze for a longer period before crashing into a wall, and they explored more territory. We also found that such agents performed better on related visual-processing tasks, such as estimating depth in a 3D image and segmenting a scene into component parts. We’re at the very beginning of mimicking human emotions in silico, and there will doubtless be philosophical debate over what it means for a machine to be able to imitate the emotional states associated with happiness or fear. But we think such approaches may not only make for more efficient learning, they may also give AI systems the crucial ability to generalize. Today’s AI systems are typically trained to carry out a single task, one that they might get very good at, yet they can’t transfer their painstakingly acquired skills to any other domain. But human beings use their emotions to help navigate new situations every day; that’s what people mean when they talk about using their gut instincts. We want to give AI systems similar abilities. If AI systems are driven by humanlike emotion, might they more closely approximate humanlike intelligence? Perhaps simulated emotions could spur AI systems to achieve much more than they would otherwise. We’re certainly curious to explore this question—in part because we know our discoveries will make us smile.
IEEE MEMBER DISCOUNTS Exclusive offers are another big plus to IEEE Membership. In addition to helping you stay abreast of the latest developments, network with your peers, and gain access to high-quality resources, an IEEE Membership will also give you exclusive offers and discounts on products and services from many trusted brands. Visit www.ieee.org/discounts for Member savings and country-specific insurance. See the Member Discounts Marketplace for new, time-limited deals.
+ +
+ ++ +++ + + + + + + + +
19-FA-032 01/20
+
+
By Michael Koziol
THE CLASH OVER 5G’S FIRST MILE
The wireless industry is divided on Open RAN’s goal to make network components interoperable
MAY 2021 SPECTRUM.IEEE.ORG 41
When all is said and done, 5G will cost almost US $1 trillion to deploy over the next half decade. That enormous expense will be borne mostly by network operators, companies like AT&T, China Mobile, Deutsche Telekom, Vodafone, and dozens more around the world that provide cellular service to their customers. Facing such an immense cost, these operators asked a very reasonable question: How can we make this cheaper and more flexible? Their answer: Make it possible to mix and match network components from different companies, with the goal of fostering more competition and driving down prices. At the same time, they sparked a schism within the industry over how wireless networks should be built. Their opponents— and sometimes begrudging partners—are the handful of telecom-equipment vendors capable of providing the hardware the network operators have been buying and deploying for years. These vendors initially opposed the scheme, called Open RAN, because they believed that if implemented, it would damage—if not destroy—their existing business model. But faced with the collective power of the operators clamoring for a new way to build wireless networks, these vendors have been left with few options, none of them very appealing. Some have responded by trying to set the terms for how Open RAN
42 SPECTRUM.IEEE.ORG MAY 2021
will be developed, while others continue to drag their feet, and risk being left behind. The technology underpinning a generation of wireless like 5G can take a decade or more to go from initial ideas to fully realized hardware. By comparison, Open RAN has emerged practically overnight. In scarcely three years, the idea has gone from little more than a concept to multiple, major deployments around the world. Its supporters believe it will nurture immense innovation and lower the costs of wireless access. Its detractors say it will threaten basic network security and could lead to disaster. Either way, this is a watershed moment in the communications industry, and there’s no turning back. Broadly speaking, a radio access network (RAN) is the framework that links an end device like a cellphone and the larger, wired, core network. A cellular base station, or tower, is the most familiar example of a RAN. Other varieties of base stations, such as the small cells that send and receive signals over short distances in 5G networks, also fit the bill. To function as this link, the RAN performs several steps. When you use your phone to call a friend or family member in a different city, for example, you need to be within range of a cell tower. So the first step is for the cell tower’s antennas to
PREVIOUS PAGES: GEORGE FREY/AFP/GETTY IMAGES
WE’VE ALL BEEN TOLD THAT 5G WIRELESS IS GOING TO DELIVER AMAZING CAPABILITIES AND SERVICES. BUT IT WON’T COME CHEAP.
RAKUTEN
Rakuten Mobile’s receive the phone’s signal. Second, a radio converts the signal standard,” says Eugina Jordan, the vice Open RAN network from analog to digital. Third, a component called the baseband president of marketing at Parallel Wireincludes 4G radios unit processes the signal, corrects errors, and finally transmits less, a New Hampshire–based company from Nokia running software from it into the core network. Within the RAN, these components— developing Open RAN technologies. But another vendor. the antenna, the radio, and the baseband unit—can be, and “those interfaces are not open, because The company has often are, treated as discrete chunks of technology. each vendor creates their own flavor,” deployed one such RAN at the If you separate the radio and the baseband unit from one she adds. Most of these vendor-specific company’s global another, and develop and construct them independently, you tweaks occur in the software and proheadquarters in still need to make sure that they work together. In other words, gramming languages used to connect Tokyo. The Open RAN network also you need their interfaces to be compatible. Withthe radio to the baseband uses servers to out such compatibility, data can be garbled or lost unit. Jordan says that the power the cloudThere is when moving from the radio to the baseband unit, tweaks primarily take the native network. or vice versa. In the worst-case scenario, a radio currently no form of vendors defining and a baseband unit with incompatible interfaces guarantee radio parameters that were will just not work together at all. A functional RAN that a radio intentionally left blank in 3GPP standards for needs to have a common interface between these manufac future development. two components. However, astonishingly, there is tured by one Ultimately, this leads to each vendor constructcurrently no guarantee that a radio manufactured vendor will ing hardware that is too incompatible with the by one vendor will be interoperable with a base- be interop others’ for operators’ comfort. “We see with 3GPP band unit manufactured by another vendor. specification more and more gaps,” says Olivier erable with The specifications for RAN interface standards, Simon, the radio innovation director at Orange, an like all of those for cellular networks, are set by the a baseband operator based in France. Simon says that of the 3rd Generation Partnership Project. Gino Masini, unit manu interfaces specified by 3GPP, “you can see that the chair of 3GPP’s RAN3 working group, says that factured many of them are not really open in the sense that many of 3GPP’s specifications, including those cov- by another they are not enabling multivendor cooperation on ering interfaces, are designed with interoperability vendor. both sides of the interface.” in mind. However, Masini, who is also principal The O-RAN Alliance, of which Simon is an execresearcher for standardization at Ericsson, adds utive committee member, is the largest industry that there is nothing preventing a vendor from “complement- group working on Open RAN specifications. The group formed ing” a standardized interface with additional proprietary tech- in 2018, when five operators—AT&T, China Mobile, Deutsche niques. Many vendors do just that—and Masini says this does Telekom, NTT Docomo, and Orange—joined to spearhead not limit vendor interoperability. more industry development of Open RAN. “I think the realiOthers in the industry don’t agree. “Both Nokia and zation was, we need to create one unified, global operator voice Ericsson are using 3GPP interfaces that are supposed to be to drive this disaggregation and openness,” says Sachin Katti,
MAY 2021 SPECTRUM.IEEE.ORG 43
an associate professor at Stanford University and one of the cochairs of the O-RAN Alliance’s technical steering committee. O-RAN Alliance members hope Open RAN can plug the gaps created by 3GPP’s specifications. They’re quick to say they’re not trying to replace the 3GPP specifications. Instead, they see Open RAN as a necessary tightening of the specifications to prevent big vendors from tacking their proprietary techniques onto the interfaces, thereby locking wireless operators into single-vendor networks. By forcing open interfaces, the wireless industry can arrive at an entirely new way to engineer its networks. And if those open interfaces promote more competition and lower prices, so much the better. As early 5G deployments were underway around the world, in 2019, the wireless industry group GSM Association predicted that operators would spend $1.3 trillion on 5G infrastructure, equipment, and technologies for their networks. RAN construction will consume the lion’s share of those capital expenditures. And much of that spending will go toward the handful of vendors that can still provide complete end-to-end networks.
Inside the Radio Access Network Cellular networks send signals over long distances using a wired or fiber-optic backbone called a core network. The radio access network (RAN) functions as a middleman, connecting an end device like a cellphone to the core network by receiving the phone’s wireless signal with its antenna, converting the signal to digital in the radio unit, and performing tasks like data processing and error correction in the baseband unit. In current 5G systems, the baseband unit splits those tasks between a distributed unit and a centralized unit. Open RAN concepts hope to build on that split to create more flexible, thinly sliced RANs.
Antenna
Cellphone (end device)
Radio unit Server farm
Baseband unit
To core network
Cell tower Distributed unit
Centralized unit
44 SPECTRUM.IEEE.ORG MAY 2021
“This was always the pain point, because RAN is the most expensive part of an operator’s deployment,” says Sridhar Rajagopal, the vice president of technology and strategy at Mavenir, a Texas-based company that provides end-to-end network software. “It takes almost 60, 70 percent of the deployment costs.” By 2025, the GSM Association predicts, operators will be spending as much as 86 percent of their capital budgets on RAN. Not surprisingly, with so much money on the line, operators do everything they can to avoid any fiascoes caused by incompatible hardware. The surest way to avoid such a disaster is to stick with the same vendor from one end of the network to the other, thus avoiding any possibility of mismatched interfaces. Another factor contributing to operator unease is the dwindling number of companies that can provide cutting-edge end-to-end networks. It’s now just three: Ericsson, Nokia, and Huawei. This trio of end-to-end vendors can charge high prices because operators are essentially locked into their systems. Even the arrival of a new generation of wireless doesn’t create a clear opportunity for an operator to switch vendors. New wireless generations maintain backward compatibility, so that, for example, a 5G phone can operate on a 4G network when it’s not within range of any 5G cells. So as operators build out their 5G deployments, they’re mostly sticking with a single vendor’s proprietary tech to ensure a smooth transition. The main alternative is scrapping everything and paying even more for a new deployment from the ground up. There is broad consensus in the wireless industry that Open RAN is making it possible to pick and choose different RAN components from different vendors. This opportunity, called disaggregation, will also remove the stress over whether components will cooperate when plugged together. Whether or not disaggregation is a good thing depends on whom you ask. Operators sure like it. Dish, a television and wireless provider, has been particularly aggressive in embracing Open RAN. Siddhartha Chenumolu, vice president of technology development at Dish, describes his first reaction to the technology: “Hey, there might be something here where it allows us to disaggregate completely,” he says. “I don’t have to rely on Ericsson only to provide radios, or Nokia only.” Dish has committed to using Open RAN for a ground-up deployment of a 5G network in the United States this year. Smaller-scale and more specialized vendors are also optimistic about the boost Open RAN can bring to their businesses. For Software Radio Systems, a maker of advanced software-defined radios, Open RAN makes it easier to focus on developing new software without worrying about losing potential customers intimidated by the task of integrating the tech into their wider networks. Not surprisingly, the big three remaining hardware vendors take different views. In February, Franck Bouétard, the CEO of Ericsson France, called Open RAN an “experimental technology” that was still years away from maturity and could not compete with Ericsson’s products. (Ericsson declined to comment for this article). But some in the industry see the hardware makers as deliberately slowing down the development of Open RAN. “Some of the big vendors, they’re continuously raising one issue
Proposed Open RAN Functional Splits Proponents of Open RAN are exploring several possible “functional splits” to create new, interoperable interfaces in RAN systems, with four possibilities gaining the most traction. Each split assigns the many tasks a RAN undertakes to create a link between the core network and an end device in different ways, based on what different kinds of cellular networks might need. Split 2, for example, creates highly intelligent radio units that handle much of the data processing before the signal is ever transferred. On the other hand, Splits 7.2x and 8 create “dumb” radios that minimize data processing in favor of lower latencies.
Split 2
Centralized unit
Interface
Distributed unit
Split 6
Centralized unit & distributed unit PACKET DATA CONVERGENCE PROTOCOL (PDCP): Compresses addressing and routing information for a message RADIO RESOURCE CONTROL (RRC): Establishes connections and oversees power usage RADIO LINK CONTROL (RLC): Checks for duplicate data and corrects errors
MEDIUM ACCESS CONTROL (MAC): Controls the RAN hardware that interacts with the wired or fiber-optic core-network connection PHYSICAL LAYER (PHY): The hardware responsible for transferring bits from the radio unit to the core network RADIO FREQUENCY LAYER (RF): A subset of the physical layer that creates the wireless signal on the appropriate frequency and transmits it
Radio unit
Split 7.2x
Centralized unit & distributed unit
Radio unit
Split 8
Centralized unit & distributed unit
Radio unit
or another,” says Paul Sutton, a director at Software Radio In its most ambitious version, Open RAN would ystems. “Ericsson is probably in the party that’s fighting back S split the RAN into smaller components beyond the most against Open RAN, because they will probably have the radio and the baseband unit. Proponents of the most to lose.” this level of disaggregation believe it would bring Not every big vendor is pushing back. Nokia, for example, even more vendors into the wireless industry, by allowing sees opportunity. “I think we need to accept the fact that Open companies to hyperspecialize. An operator could contract RAN is going to happen anyway, with or without with a vendor for just the processor that readies us,” says Thomas Barnett, a mobile-network stratthe data received from the core network for wireegy and technology lead at Nokia. “We, at Nokia, By 2025, less transmission, for example. Many in the indusdecided to be proactive in taking a leadership posi- the GSM try also say that this kind of specialization would tion in order to grab a better market-share position.” Association speed technological innovation by making it posJapanese operator Rakuten’s Open RAN deploy- predicts, sible to swap out and deploy a new RAN compoments are using Nokia’s equipment, for example, operators nent without waiting for the entire radio or and Nokia is also working with Deutsche Telekom will be baseband unit to be upgraded. “That’s maybe one to deploy an Open RAN system in Neubrandenburg, spending of the brightest opportunities that Open RAN Germany, later this year. could provide,” says Ted Rappaport, the founding as much as That’s not to say Nokia or other vendors are on director of NYU Wireless, a research center for 86 percent of advanced wireless technologies. the same page as the operators and the specialtheir capital ized vendors like Software Radio Systems. At the The wireless industry’s first efforts with disagbudgets on moment, there’s still plenty of debate. Ericsson gregation were inspired by 5G specifications themand other vendors argue that creating more open RAN. selves. These specifications split the baseband unit, interfaces will inevitably create more points in the which is responsible for processing and transferring network for cyberattacks. Operators and other Open RAN data to or from the core network, into two smaller components. proponents counter that standardized interfaces will make it One component is the distributed unit, which takes over the easier for the industry to identify and fix vulnerabilities. Every- data-processing responsibilities. The other component is the one seems to have a different opinion on how much openness centralized unit, which handles the connection to the core is enough openness, or on just how much the RAN hardware network. The advantage of splitting the baseband unit in this elements should be disaggregated. way is that the centralized unit no longer needs to be located
MAY 2021 SPECTRUM.IEEE.ORG 45
at the cell tower itself. Instead, a single centralized unit can sit storm that might dampen signals. The RIC can reprogram the in a local server farm, maintaining the connection to the core RAN’s software components in order to deliver better service. network for multiple cell towers in the area. “Imagine the possibility where I can really adapt my network, The O-RAN Alliance is working on a handful of different based on the user experience, how the user is feeling in real “functional splits” in the RAN to create more opportunities for time,” says Dish’s Chenumolu. “How great is that?” disaggregation beyond this split between the distributed unit and the centralized unit. Each of these additional splits creates Since its founding in 2018, the O-RAN Alliance a division somewhere amid the many steps between a signal’s has ballooned from its five founding members— arrival from the core network and its transmission to a cellall operators—to more than 260 members. Of the phone. It’s a bit like taking a lunch break: You can take an early big three vendors, only Huawei is not a member, lunch and thus shift many of your responsibilities to the after- citing its belief that Open RAN systems cannot perform as noon, or work for several hours before opting for a later lunch. well as the company’s proprietary systems. Other Open RAN One important split, called Split 7.2x, hands responsibilities groups are growing at a similar pace. The Open RAN Policy such as signal encoding and decoding, as well as modulation, to Coalition, for example, was founded in May 2020 and already the distributed unit. On the other side of the split, the radio is has over 60 members working to coordinate global policy on responsible for some light processing duties like beamforming, Open RAN development and deployment. which establishes the specific direction of a transmission. The In recent months, Rakuten Mobile, a unit of the Japanese radio is also still responsible for converting digital signals to e-commerce giant, and Dish have committed to Open RAN for analog signals and vice versa. extensive new 5G deployments. After a mandate from the British Another split, Split 8, shifts even the responsibility for government to strip all Huawei components from wireless netbeamforming to the distributed unit, leaving the radio respon- works, England-based Vodafone is replacing those components sible only for converting signals. In contrast, Split 2 would push in its own networks with Open RAN equivalents. Because of encoding, decoding, modulation, beamforming, and even more similar mandates, local operators in the United States, such as processing responsibilities to the radio, leaving the Idaho-based Inland Cellular, are doing the same. distributed unit responsible only for compressing These deployments haven’t always gone as data to a smaller number of bits before transferring “Some of the planned. Rakuten, in particular, faced some initial big vendors, the data to the centralized unit. setbacks when its Open RAN network’s perforThe goal in creating open standards for multiple they’re mance didn’t match the performance of a traditional kinds of splits is that operators can then purchase continuously end-to-end system. The operator remains optimisbetter-tailored components for the specific kind of raising one tic, however, and hasn’t given up on it. Many in the network they’re building. For example, an operator issue or industry aren’t concerned about these kinds of might opt for Split 8 for a large-scale deployment another.” issues, arguing that the only way to actually iron out requiring a lot of radios. This split allows the radios the wrinkles in the technology is to deploy it at scale PAUL SUTTON, to be as “dumb,” and therefore cheap, as possible and see what works and what needs improvement. SOFTWARE RADIO because all of the processing happens in the cenThere are also still lingering questions over SYSTEMS tralized unit. where the buck stops. When an operator buys an endIt’s technically possible to put together a disagto-end system from Nokia or Ericsson or Huawei, it gregated RAN with open interfaces using only hardware, but also knows it can depend on that vendor to support the netdefining the components in software has some advantages. work when problems crop up. Not so with Open RAN deploy“Our industry has become really, really hardware-centric,” says ments, where no single vendor is likely to claim responsibility Chih-Lin I, who, along with Stanford’s Katti, is cochair of the for interoperability issues. Larger operators will likely be able to O-RAN Alliance’s technical steering committee. “Every genera- support their own Open RAN networks, but smaller operators tion of our networks basically rely on special-purpose hardware may be reliant on companies like Mavenir, which have positioned with tightly coupled software. So every time we need to have an themselves as system integrators. Critics, however, see that upgrade, or new release, or new fractional release, it takes years.” approach as just creating another kind of end-to-end vendor— In order to move away from a hardware-centric attitude, and adding additional expense—for operators that don’t have the O-RAN Alliance is also encouraging the wireless industry the expertise or resources to support their own networks. to incorporate more software into the RAN. Software-defined In the end, Open RAN’s true test may come when it’s time networks, which replace traditional hardware components to implement the next generation of wireless. “I think 6G will with programmable software equivalents, are more flexible. be built with Open RAN as a prior assumption,” says Rajat Upgrading a virtual component can be as simple as pushing Prakash, the principal engineer of wireless R&D at Qualcomm. out new code to the base station. It remains to be seen how far the movement will go to disThe emphasis on software is also making it possible for aggregate the RAN, to open up new interfaces, or even to bring the industry to consider entirely new technologies, the most new technologies into the mix. What’s important is that the important of which is the RAN Intelligent Controller. The RIC movement has already gained substantial momentum. Even collects data from the RAN components of dozens or hundreds though some corners of the industry still have reservations, of base stations at once and uses machine-learning techniques operators and small-scale vendors have put too much weight to reconfigure network operations in real time. It bases the behind the idea for the movement to fizzle out. Open RAN is modifications on whether particular cell towers are under a here to stay. As it matures, the wireless industry will be open heavy traffic load, for example, or transmitting in a heavy rain- for a new way of doing business.
46 SPECTRUM.IEEE.ORG MAY 2021
We have 30 million reasons to be proud. Thanks to our donors, supporters and volunteers who answered the call of the
Realize the Full Potential of IEEE Campaign, helping impact lives around the world through the power of technology and education.
Illuminate
Educate
Engage
Energize
Realize Your Impact Learn how: ieeefoundation.org/campaign
C O N J U R E R
of
C O M P R E S S I O N From WinZips to cat GIFs, JACOB ZIV’s algorithms have been making data disappear and reappear for decades By Tekla S. Perry
48 SPECTRUM.IEEE.ORG MAY 2021
Photo by Rami Shlush
Lossless data compression seems a bit like a magic trick. Its cousin, lossy compression, is easier to comprehend. Lossy algorithms are used to get music into the popular MP3 format and turn a digital image into a standard JPEG file. They do this by selectively removing bits, taking what scientists know about the way we see and hear to determine which bits we’d least miss. But no one can make the case that the resulting file is a perfect replica of the original. Not so with lossless data compression. Bits do disappear, making the data file dramatically smaller and thus easier to store and transmit. The important difference is that the bits reappear on command. It’s as if the bits are rabbits in a magician’s act, disappearing and then reappearing from inside a hat at the wave of a wand. The world of magic had Houdini, who pioneered tricks that are still performed today. And data compression has Jacob Ziv. In 1977, Ziv, working with Abraham Lempel, published the equivalent of Houdini on Magic: a paper in the IEEE Transactions on Information Theory titled “A Universal Algorithm for S equential Data Compression.” The algorithm described in the paper came to be called LZ77—from the authors’ names, in alphabetical order, and the year. LZ77 wasn’t the first lossless compression algorithm, but it was the first that could work its magic in a single step. The following year, the two researchers issued a refinement, LZ78. That algorithm became the basis for the Unix compress program used in the
early ’80s; WinZip and Gzip, born in the early ’90s; and the GIF and TIFF image formats. Without these algorithms, we’d likely be mailing large data files on discs instead of sending them across the Internet with a click, buying our music on CDs instead of streaming it, and looking at Facebook feeds that don’t have bouncing animated images. Ziv went on to partner with other researchers on other innovations in compression. It is his full body of work, spanning more than half a century, that earned him the 2021 IEEE Medal of Honor “for fundamental contributions to information theory and data compression technology, and for distinguished research leadership.” ziv was born in 1931 to Russian immigrants in Tiberias, a city then in British-ruled Palestine and now part of Israel. Electricity and gadgets— and little else—fascinated him as a child. While practicing violin, for example, he came up with a scheme to turn his music stand into a lamp. He also tried to build a Marconi transmitter from metal player-piano parts. When he plugged the contraption in, the entire house went dark. He never did get that transmitter to work. When the Arab-Israeli War began in 1948, Ziv was in high school. Drafted into the Israel Defense Forces, he served briefly on the front lines until a group of mothers held organized protests, demanding that the youngest soldiers
be sent elsewhere. Ziv’s reassignment took him to the Israeli Air Force, where he trained as a radar technician. When the war ended, he entered Technion— Israel Institute of Technology to study electrical engineering. After completing his master’s degree in 1955, Ziv returned to the defense world, this time joining Israel’s National Defense Research Laboratory (now Rafael Advanced Defense Systems) to develop electronic components for use in missiles and other military systems. The trouble was, Ziv recalls, that none of the engineers in the group, including himself, had more than a basic understanding of electronics. Their electrical engineering education had focused more on power systems. “We had about six people, and we had to teach ourselves,” he says. “We would pick a book and then study together, like religious Jews studying the Hebrew Bible. It wasn’t enough.” The group’s goal was to build a telemetry system using transistors instead of vacuum tubes. They needed not only knowledge, but parts. Ziv contacted Bell Telephone Laboratories and requested a free sample of its transistor; the company sent 100. “That covered our needs for a few months,” he says. “I give myself credit for being the first one in Israel to do something serious with the transistor.” In 1959, Ziv was selected as one of a handful of researchers from Israel’s defense lab to study abroad. That program, he says, transformed the evolution of science in Israel. Its organizers didn’t steer the selected young engineers and scientists into particular fields. Instead, they let them pursue any type of graduate studies in any Western nation.
“In order to run a computer program at the time, you had to use punch cards and I hated them. That is why I didn’t go into real computer science.” —Jacob Ziv
50 SPECTRUM.IEEE.ORG MAY 2021
JACOB ZIV/TECHNION
Ziv planned to continue working in communications, but he was no longer interested in just the hardware. He had recently read Information Theory (Prentice-Hall, 1953), one of the earliest books on the subject, by Stanford Goldman, and he decided to make information theory his focus. And where else would one study information theory but MIT, where Claude Shannon, the field’s pioneer, had started out?
Ziv arrived in Cambridge, Mass., in 1960. His Ph.D. research involved a method of determining how to encode and decode messages sent through a noisy channel, minimizing the probability and error while at the same time keeping the decoding simple. “Information theory is beautiful,” he says. “It tells you what is the best that you can ever achieve, and [it] tells you how to approximate the outcome. So
Jacob Ziv [with glasses], who became chair of Technion’s electrical engineering department in the 1970s, worked earlier on information theory with Moshe Zakai. The two collaborated on a paper describing what became known as the Ziv-Zakai bound.
if you invest the computational effort, you can know you are approaching the best outcome possible.” Ziv contrasts that certainty with the uncertainty of a deep-learning algorithm. It may be clear that the algorithm is working, but nobody really knows whether it is the best result possible. While at MIT, Ziv held a part-time job at U.S. defense contractor Melpar, where he worked on error-correcting software. He found this work less beautiful. “In order to run a computer program at the time, you had to use punch cards,” he recalls. “And I hated them. That is why I didn’t go into real computer science.” Back at the Defense Research Laboratory after two years in the United States, Ziv took charge of the Communications Department. Then in 1970, with several other coworkers, he joined the faculty of Technion. There he met Abraham Lempel. The two discussed trying to improve lossless data compression. The state of the art in lossless data compression at the time was Huffman coding. This approach starts by finding sequences of bits in a data file and then sorting them by the frequency with which they appear. Then the encoder builds a dictionary in which the most common sequences are represented by the smallest number of bits. This is the same idea behind Morse code: The most frequent letter in the English language, e, is represented by a single dot, while rarer letters have more complex combinations of dots and dashes. Huffman coding, while still used today in the MPEG-2 compression format and a lossless form of JPEG, has its drawbacks. It requires two passes through a data file: one to calculate the statistical features of the file, and the second to encode the data. And storing the dictionary along with the encoded data adds to the size of the compressed file. Ziv and Lempel wondered if they could develop a lossless data- compression algorithm that would work on any kind of data, did not require preprocessing, and would achieve the best compression for that data, a target defined by something known as the Shannon entropy. It was
MAY 2021 SPECTRUM.IEEE.ORG 51
52 SPECTRUM.IEEE.ORG MAY 2021
“Information theory is beautiful. It tells you what is the best that you can ever achieve, and [it] tells you how to approximate the outcome.” —Jacob Ziv
sequences that appear more frequently will grow longer branches.” “It turns out,” he says, “that not only was that the optimal [approach], but so simple that it became useful right away.” While ziv and Lempel were developing LZ78, they were both on sabbatical from Technion and working at U.S. companies. They knew their algorithm would be commercially useful, and they wanted to patent it. “I was at Bell Labs,” Ziv recalls, “and so I thought the patent should belong to them. But they said that it’s not possible to get a patent unless it’s a piece of hardware, and they were not interested in trying.” (The U.S. Supreme Court didn’t open the door to direct patent protection for software until the 1980s.) However, Lempel’s employer, Sperry Rand Corp., was willing to try. It got around the restriction on software patents by building hardware that implemented the algorithm and patenting that device. Sperry Rand followed that first patent with a version adapted by researcher Terry Welch, called the LZW algorithm. It was the LZW variant that spread most widely. Ziv regrets not being able to patent LZ78 directly, but, he says, “We enjoyed the fact that [LZW] was very popular. It made us famous, and we also enjoyed the research it led us to.” One concept that followed came to be called Lempel-Ziv complexity, a measure of the number of unique substrings contained in a sequence of bits. The fewer unique substrings, the more a sequence can be compressed. This measure later came to be used to check the security of encryption codes;
JACOB ZIV/TECHNION
unclear if their goal was even possible. that sequence first appeared, along with They decided to find out. the length of the matched sequence. The Ziv says he and Lempel were the number of bits that you need for that “perfect match” to tackle this question. pointer is very small. “I knew all about information theory “It’s basically what they used to do in and statistics, and Abraham was well publishing TV Guide,” Ziv says. “They equipped in Boolean algebra and com- would run a synopsis of each program puter science.” once. If the program appeared more than The two came up with the idea of once, they didn’t republish the synopsis. having the algorithm look for unique They just said, go back to page x.” sequences of bits at the same time that Decoding in this way is even simpler, it’s compressing the data, using point- because the decoder doesn’t have to ideners to refer to previously seen sequences. tify unique sequences. Instead it finds the This approach requires only one pass locations of the sequences by following the through the file, so it’s faster than pointers and then replaces each pointer Huffman coding. with a copy of the relevant sequence. Ziv explains it this way: “You look at The algorithm did everything Ziv and incoming bits to find the longest stretch Lempel had set out to do—it proved that of bits for which there is a match in the universally optimum lossless comprespast. Let’s say that first incoming bit is a sion without preprocessing was possible. 1. Now, since you have only one bit, you “At the time they published their work, have never seen it in the past, so you the fact that the algorithm was crisp and have no choice but to transmit it as is.” elegant and was easily implementable “But then you get another bit,” he with low computational complexity was continues. “Say that’s a 1 as well. So you almost beside the point,” says Tsachy enter into your dictionary 1-1. Say the Weissman, an electrical engineering next bit is a 0. So in your dictionary you professor at Stanford University who now have 1-1 and also 1-0.” specializes in information theory. “It Here’s where the pointer comes in. was more about the theoretical result.” The next time that the stream of bits Eventually, though, researchers recincludes a 1-1 or a 1-0, the software ognized the compression algorithm’s doesn’t transmit those bits. Instead it practical implications, Weissman says. sends a pointer to the location where “The algorithm itself became really useful when our technologies started dealing with larger file sizes beyond 100,000 or even a million characters.” Jacob Ziv [left] and Abraham Lempel “Their story is a story about the power published algorithms for lossless of fundamental theoretical research,” data compression in 1977 and 1978, both in the IEEE Transactions on Weissman adds. “You can establish theInformation Theory. The methods oretical results about what should be became known as LZ77 and LZ78 and achievable—and decades later humanare still in use today. ity benefits from the implementation of algorithms based on those results.” Ziv and Lempel kept working on the technology, trying to get closer to entropy for small data files. That work led to LZ78. Ziv says LZ78 seems similar to LZ77 but is actually very different, because it anticipates the next bit. “Let’s say the first bit is a 1, so you enter in the dictionary two codes, 1-1 and 1-0,” he explains. “You can imagine these two sequences as the first branches of a tree.” “When the second bit comes,” Ziv says, “if it’s a 1, you send the pointer to the first code, the 1-1, and if it’s 0, you point to the other code, 1-0. And then you extend the dictionary by adding two more possibilities to the selected branch of the tree. As you do that repeatedly,
if a code is truly random, it cannot be compressed. Lempel-Ziv complexity has also been used to analyze electroencephalograms—recordings of electrical activity in the brain—to determine the depth of anesthesia, to diagnose depression, and for other purposes. Researchers have even applied it to analyze pop lyrics, to determine trends in repetitiveness. Over his career, Ziv published some 100 peer-reviewed papers. While the 1977 and 1978 papers are the most famous, information theorists that came after Ziv have their own favorites. For Shlomo Shamai, a distinguished professor at Technion, it’s the 1976 paper that introduced the Wyner-Ziv algorithm, a way of characterizing the limits of using supplementary information available to the decoder but not the encoder. That problem e merges, for example, in video applications that take advantage of the fact that the decoder has already deciphered the previous frame and thus it can be used as side information for encoding the next one. For Vincent Poor, a professor of electrical engineering at Princeton University, it’s the 1969 paper describing the Ziv-Zakai bound, a way of knowing whether or not a signal processor is getting the most accurate information possible from a given signal. Ziv also inspired a number of leading data-compression experts through the classes he taught at Technion until 1985. Weissman, a former student, says Ziv “is deeply passionate about the mathematical beauty of compression as a way to quantify information. Taking a course from him in 1999 had a big part in setting me on the path of my own research.” He wasn’t the only one so inspired. “I took a class on information theory from Ziv in 1979, at the beginning of my master’s studies,” says Shamai. “More than 40 years have passed, and I still remember the course. It made me eager to look at these problems, to do research, and to pursue a Ph.D.” In recent years, glaucoma has taken away most of Ziv’s vision. He says that a paper published in IEEE Transactions on Information Theory this January is his last. He is 89. “I started the paper two and a half years ago, when I still had enough vision to use a computer,” he says. “At the end,
Photo by Rami Shlush
Yuval Cassuto, a younger faculty member at Technion, finished the project.” The paper discusses situations in which large information files need to be transmitted quickly to remote databases. As Ziv explains it, such a need may arise when a doctor wants to compare a patient’s DNA sample to past samples from the same patient, to determine if there has been a mutation, or to a library of DNA, to determine if the patient has a genetic disease. Or a researcher studying a new virus may want to compare its DNA sequence to a DNA database of known viruses. “The problem is that the amount of information in a DNA sample is huge,”
Ziv says, “too much to be sent by a network today in a matter of hours or even, sometimes, in days. If you are, say, trying to identify viruses that are changing very quickly in time, that may be too long.” The approach he and Cassuto de scribe involves using known sequences that appear commonly in the database to help compress the new data, without first checking for a specific match between the new data and the known sequences. “I really hope that this research might be used in the future,” Ziv says. If his track record is any indication, Cassuto-Ziv—or perhaps CZ21—will add to his legacy.
MAY 2021 SPECTRUM.IEEE.ORG 53
Bruce Schneier Better code needs big-picture thinking BY DANIEL DERN
T
he security technologist Bruce Schneier has a warning “What you code affects the world now. Gone are the days when programmers could ignore the social context of what they code, when we could say, ‘The users will just figure it all out.’ Today, programs, apps, and algorithms affect society. Facebook’s choices influence democracy. How driverless cars will choose to avoid accidents will affect human lives.” Schneier should know, because synthesizing and explaining the impact of technology is what he does. “I work at the intersection of security, technology, and people, mostly thinking about security and privacy policy…. I don’t have a single job,” says Schneier. “Instead, I do a portfolio of related things.”
54 SPECTRUM.IEEE.ORG MAY 2021
This includes writing books (14 so far); essays and op-eds; his monthly-since-1998 newsletter and his daily-since-2004 blog; teaching cybersecurity policy at the Harvard Kennedy School; being a fellow at the Berkman Klein Center for Internet and Society at Harvard University; being chief of security architecture at Inrupt; speaking at conferences and events (unsurprisingly, he has done a TED talk); and now and then some security consulting. “My latest book, Click Here to Kill Everybody [2018], is about the security of cyberphysical systems. Everything is turning into a computer—cars, appliances, toys, streetlamps, power plants— and these computers can affect the world in a direct physical manner. Computer
MAY 2021
security is now about life and property.” Schneier started out in cryptography in the mid-1990s, becoming a public expert after he was laid off from a tech job at AT&T. “I started writing for computer magazines. I wrote cryptography articles for Dr. Dobb’s Journal. Then I sold my first book to Wiley— Applied Cryptography [1993]—which became a bestseller. The book became a 600-page business card, and I started doing cryptography consulting. From there, I generalized to computer security, then network security, then general security technology...and then to the economics and psychology, sociology, and now, public policy of security.” Schneier does not want to be alone in this work, and encourages others to join him. “We need people who can assess the technologies in social context, how they could impact the real world— and what public policies should address this. To do that, you need to be able to synthesize across technology and policy, and explain this to both technologists and policymakers.” And this greater context needs to be factored in at all stages of the software life cycle, “We need social scientists on our software-development teams.” Does this field sound appealing? “Where you start out almost doesn’t matter. But look outside your silo, look at adjacent or complementary disciplines.” As an example, Schneier points to security economics. “I devote a class session on the economics of security. And another on the psychology of security. If you’re a security engineer and you don’t understand the economic considerations of the problem you’re trying to solve, you are going to get the incentives all wrong. And what you create might never get used.” Becoming a good communicator is essential, stresses Schneier. “Explaining technology across interdisciplinary boundaries requires being able to write, speak, to animate a topic, to analogize and synthesize, to summarize and generalize. These are all critical skills. They’re not specific skills, but they are vitally important.”
AN RONG XU/REDUX
SHARING THE EXPERIENCES OF WORKING ENGINEERS
Full Professor and Assistant / Associate Professor in Computer and Information Science
(Ref. No.: FST/CIS/AAP/03/2021; FST/CIS/FP/03/2021)
The Department of Computer and Information Science (CIS) of the University of Macau (UM) invites applications for the position of Full Professor in Computer Science and Assistant / Associate Professor in areas of Financial Technology and Distributed Computing, Software Engineering and Service Computing, Visualization and AR/VR. We are seeking candidate with a proven record of accomplishment in research and education. CIS was founded in 1990 and currently has 25 academic staff. The Department offers Bachelor’s, Master’s and PhD degree programmes, which cover the main aspects of modern computer science and related technologies. UM is among the top 1% in ESI rankings in both Engineering and Computer Science. In the THE World University Rankings, the Computer Science programme is ranked among the top 200. The candidates must have an earned PhD degree in related areas. Preference will be given to candidates with extensive research and teaching and corresponding academic leadership experience at the tertiary education level for Full Professor. Applicants should visit https://career.admo.um.edu.mo/ for more details, and apply ONLINE.
NOW HIRING Quick Learners & Problems Solvers Preferred
BAW R&D RF Lab Manager Location: Orlando, FL | Qualifications: AS/BS degree in electrical engineering with RF focus, RF test equipment hands-on experience APPLY NOW
BAW R&D Advanced Design Manager Location: Orlando, FL | Qualifications: MS or PhD in electrical engineering with RF focus, tech leadership, strong RF engineering background, EDA tools experience APPLY NOW
qorvo.com/careers © 04-2021 Qorvo US, Inc. | QORVO is a registered trademark of US, Inc.
TAP. CONNECT. NETWORK. SHARE. Connect to IEEE–no matter where you are–with the IEEE App. Stay up-to-date with the latest news Schedule, manage, or join meetups virtually Get geo and interest-based recommendations Read and download your IEEE magazines Create a personalized experience
ARL Distinguished Postdoctoral Fellowships The Army Research Laboratory (ARL) Distinguished Postdoctoral Fellowships provide opportunities to pursue independent research in ARL laboratories. Fellows benefit by working alongside some of the nation’s best scientists and engineers, while enhancing the mission and capabilities of the U.S. Army and the warfighter in times of both peace and war. Fellows must display extraordinary abilities in scientific research and show clear promise of becoming future leaders. Candidates are expected to have already successfully tackled a major scientific or engineering problem or to have provided a new approach or insight, evidenced by a recognized impact in their field. Fellowships are one-year appointments, renewable for up to three based on performance. The award includes a $100,000 annual stipend, health insurance, paid relocation, and a professional travel allowance. Applicants must have completed all requirements for a Ph.D. or Sc.D. degree by October 1, 2021, and may not be more than five years beyond their doctoral degree as of the application deadline. For more information and to apply visit www.nas.edu/arl. Online applications must be submitted by June 15, 2021 at 5 PM Eastern Time.
Locate IEEE members by location, interests, and affiliations
MAY 2021 SPECTRUM.IEEE.ORG 55
HISTORY IN AN OBJECT
BY ALLISON MARSH
On 13 March 1876, Thomas Edison applied for what became U.S. Patent No. 180,857, for an Improvement in Autographic Printing. The “improvement” was an electric pen that worked like a cross between a dentist’s drill and a sewing machine. Powered by two batteries and driven by a motor, it had a needle that made 50 punctures per second to create a perforated stencil. The stencil could then be used in a documentduplication system to print up to 5,000 copies. Although the pen failed to find a market, the basic concept of stencil making with electricity directly inspired both the mimeograph and the tattoo. FOR MORE ON THE HISTORY OF EDISON’S ELECTRIC PEN, SEE spectrum.ieee.org/ pastforward-may2021
56 SPECTRUM.IEEE.ORG MAY 2021
SCIENCE MUSEUM GROUP
Thomas Edison, Tattoo Artist?
From the forefront of innovation to the frontlines of the mission.
Innovation can’t happen without collaboration. So we take time to understand the challenges you face and build the solutions you need to keep the world secure, together. Because the success of the mission relies on teamwork. Learn more at lockheedmartin.com
MATLAB SPEAKS
MACHINE LEARNING With MATLAB® you can use clustering, regression, classification, and deep learning to build predictive models and put them into production.
@2021 The MathWorks, Inc.
mathworks.com/machinelearning