sign in sign up
<<
Home , ACM Interactions, Andrei Broder, Douglas Engelbart, Robert Tarjan, SIGDOC, William Kahan

COMMUNICATIONS CACM.ACM.ORG OF THEACM 11/2017 VOL.60 NO.11

Reconfigurable Cambits Association for Computing Machinery Previous A.M. Recipients

1966 A.J. Perlis 1967 1968 R.W. Hamming 1969 1970 J.H. Wilkinson 1971 John McCarthy 1972 E.W. Dijkstra 1973 1974 1975 1975 Herbert Simon 1976 Michael Rabin 1976 1977 1978 Robert Floyd 1979 Kenneth Iverson 1980 C.A.R Hoare ACM A.M. TURING AWARD 1981 Edgar Codd 1982 1983 NOMINATIONS SOLICITED 1983 1984 Nominations are invited for the 2017 ACM A.M. Turing Award. 1985 Richard Karp 1986 This is ACM’s oldest and most prestigious award and is given 1986 to recognize contributions of a technical nature which are of 1987 1988 lasting and major technical importance to the computing field. 1989 The award is accompanied by a prize of $1,000,000. 1990 Fernando Corbató 1991 Financial support for the award is provided by Google Inc. 1992 1993 Nomination information and the online submission form 1993 Richard Stearns 1994 are available on: 1994 http://amturing.acm.org/call_for_nominations.cfm 1995 1996 1997 Additional information on the Turing Laureates 1998 James Gray is available on: 1999 Frederick Brooks http://amturing.acm.org/byyear.cfm 2000 2001 Ole-Johan Dahl 2001 The deadline for nominations/endorsements is 2002 January 15, 2018. 2002 Ronald Rivest 2002 2003 For additional information on ACM’s award program 2004 Vinton Cerf 2004 Robert Kahn please visit: www.acm.org/awards/ 2005 2006 Frances E. Allen 2007 Edmund Clarke 2007 E. Allen Emerson 2007 2008 2009 Charles P. Thacker 2010 Leslie G. Valiant 2011 2012 2012 2013 2014 2015 2015 2016 Sir Tim Berners-Lee

CACM_2017_TuringSubmit_FP.indd 1 10/4/17 11:54 AM Introducing ACM Transactions on Human-Robot Interaction

Now accepting submissions to ACM THRI

In January 2018, the Journal of Human-Robot Interaction (JHRI) will become an ACM publication and be rebranded as the ACM Transactions on Human-Robot Interaction (THRI).

Founded in 2012, the Journal of HRI has been serving as the premier peer-reviewed interdisciplinary journal in the eld.

Since that time, the human-robot interaction eld has experienced substantial growth. Research ndings at the intersection of robotics, human-computer interaction, arti cial intelligence, haptics, and natural language processing have been responsible for important discoveries and breakthrough technologies across many industries.

THRI now joins the ACM portfolio of highly respected journals. It will continue to be open access, fostering the widest possible readership of HRI research and information. All issues will be available on the ACM Digital Library.

Editors-in-Chief Odest Chadwicke Jenkins of the University of Michigan and Selma Šabanović of Indiana University plan to expand the scope of the publication, adding a new section on mechanical HRI to the existing sections on computational, social/behavioral, and design-related scholarship in HRI.

The inaugural issue of the rebranded ACM Transactions on Human-Robot Interaction is planned for March 2018.

To submit, go to https://mc.manuscriptcentral.com/thri COMMUNICATIONS OF THE ACM

Departments News Viewpoints

5 From the Co-Chairs of the ACM 20 Legally Speaking Student Research Competition Disgorging Profits in Highlights of the ACM Design Patent Cases Student Research Competition Does the recent U.S. Supreme Court By Laurie Williams and Doug Baldwin decision in the Apple v. Samsung case represent a quagmire? 6 Cerf’s Up By Pamela Samuelson Heidelberg Laureate Forum By Vinton G. Cerf 23 Computing Ethics Engaging the Ethics 7 Vardi’s Insights of Data Science in Practice Would Turing Have Won Seeking more common ground the Turing Award? between data scientists and By Moshe Y. Vardi their critics. 17 By Solon Barocas and danah boyd 8 Letters to the Editor They See What You See 12 A Block on the Old Chip 26 Education Block copolymers may Keeping the Machinery in 10 BLOG@CACM help transistors shrink Computing Education Opportunities for Women, to tinier dimensions. Incorporating intellectual and Minorities in Information Retrieval By Neil Savage developmental frameworks into Mei Kobayashi describes activities to a Scottish school curriculum. support diversity and inclusion at the 15 Censoring Sensors By Richard Connor, Quintin Cutts, annual meeting of the ACM Special Amid growing outcry over and Judy Robertson Interest Group on Information controversial online videos, Retrieval in Tokyo this summer. tech firms grapple with how best 29 Viewpoint to police online advertising. Pay What You Want as a Pricing 27 Calendar By Alex Wright Model for Open Access Publishing? Analyzing the “Pay What You Want” 100 Careers 17 Overcoming Disabilities business model for Brain-computer interfaces hold open access publishing. the promise of fully featured By Martin Spann, Lucas Stich, Last Byte replacements for body parts that and Klaus M. Schmidt don’t work or are missing. 112 Future Tense By Esther Shein 32 Viewpoint Butterfly Effect Social Agents: Bridging But, like the weather, what can Simulation and Engineering anyone do about it? Seeking better integration By Seth Shostak of two research communities. By Virginia Dignum

Association for Computing Machinery Advancing Computing as a Science & Profession PHOTO BY BSIP/UIG VIA GETTY BY IMAGES PHOTO

2 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 11/2017 VOL. 60 NO. 11

Practice Contributed Articles Review Articles

44 54 68

36 Hootsuite: In Pursuit 54 Cambits: A Reconfigurable 68 Healthcare Robotics of Reactive Systems Camera System Healthcare robotics can provide A discussion with Edward Steel, Multiple computational cameras health and wellness support to Yanik Berube, Jonas Bonér, can be assembled from a common billions of people. Ken Britton, and Terry Coatta set of imaging components. By Laurel D. Riek By Makoto Odamaki and Shree K. Nayar 44 Breadth and Depth

We all wear many hats, but make Watch the author discuss her work in this exclusive sure you have one that fits well. Watch the authors discuss Communications video. their work in this exclusive By Kate Matsudaira https://cacm.acm.org/ Communications video. videos/healthcare-robotics https://cacm.acm.org/ 46 Is There a Single Method for videos/cambits the of Things? Research Highlights Essence can keep software 62 User Reviews of Top Mobile Apps development for the IoT in Apple and Google App Stores 80 Technical Perspective from becoming unwieldy. The varying review dynamics seen in Solving Imperfect Information Games By Ivar Jacobson, Ian Spence, different app stores can help guide By David Silver and Pan-Wei Ng future app development strategies. By Stuart McIlroy, Weiyi Shang, 81 Heads-Up Limit Hold’em Articles’ development led by Nasir Ali, and Ahmed E. Hassan Poker Is Solved queue.acm.org By Michael Bowling, Neil Burch, Michael Johanson, and Oskari Tammelin

89 Technical Perspective About the Cover: Exploring a Kingdom This month’s covers depict by Geodesic Measures a simple set of blocks that can be used to build By Marc Alexa a variety of cameras with very different functionalities. Four 90 The Heat Method for different Communications Distance Computation covers are circulating worldwide, each depicting By Keenan Crane, Clarisse Weischedel, a different “Cambit.” and Max Wardetzky Our thanks to Shree Nayar and Anne Fleming of Columbia University for hosting this photoshoot.

L-TO-R IMAGES BY: LIGHTSPRING; ALEXANDER BERG; LAURA LEZZA/GETTY BERG; LAURA IMAGES ALEXANDER LIGHTSPRING; BY: IMAGES L-TO-R Covers by Alexander Berg.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 3 COMMUNICATIONS OF THE ACM Trusted insights for computing’s leading professionals.

Communications of the ACM is the leading monthly print and online magazine for the computing and information technology fields. Communications is recognized as the most trusted and knowledgeable source of industry information for today’s computing professional. Communications brings its readership in-depth coverage of emerging areas of , new trends in information technology, and practical applications. Industry leaders use Communications as a platform to present and debate various technology implications, public policies, engineering challenges, and market trends. The prestige and unmatched reputation that Communications of the ACM enjoys today is built upon a 50-year commitment to high-quality editorial content and a steadfast dedication to advancing the arts, sciences, and applications of information technology.

ACM, the world’s largest educational STAFF EDITORIAL BOARD ACM Copyright Notice and scientific computing society, delivers DIRECTOR OF PUBLICATIONS EDITOR-IN-CHIEF Copyright © 2017 by Association for resources that advance computing as a Scott E. Delman Andrew A. Chien Computing Machinery, Inc. (ACM). science and profession. ACM provides the [email protected] [email protected] Permission to make digital or hard copies computing field’s premier Digital Library of part or all of this work for personal and serves its members and the computing Deputy to the Editor-in-Chief or classroom use is granted without Executive Editor profession with leading-edge publications, Lihan Chen fee provided that copies are not made Diane Crawford conferences, and career resources. [email protected] or distributed for profit or commercial Managing Editor advantage and that copies bear this Thomas E. Lambert Executive Director and CEO SENIOR EDITOR notice and full citation on the first Senior Editor Bobby Schnabel Moshe Y. Vardi page. Copyright for components of this Andrew Rosenbloom Deputy Executive Director and COO work owned by others than ACM must Senior Editor/News Patricia Ryan NEWS be honored. Abstracting with credit is Lawrence M. Fisher Director, Office of Information Systems Co-Chairs permitted. To copy otherwise, to republish, Web Editor Wayne Graves William Pulleyblank and Marc Snir to post on servers, or to redistribute to David Roman Director, Office of Financial Services Board Members lists, requires prior specific permission Rights and Permissions Darren Ramdin Monica Divitini; Mei Kobayashi; and/or fee. Request permission to publish Deborah Cotton Director, Office of SIG Services ; Rajeev Rastogi; from [email protected] or fax Editorial Assistant Donna Cappo François Sillion (212) 869-0481. Jade Morris Director, Office of Publications Scott E. Delman VIEWPOINTS For other copying of articles that carry a Art Director Co-Chairs code at the bottom of the first or last page Tim Finin; Susanne E. Hambrusch; or screen display, copying is permitted ACM COUNCIL Andrij Borys John Leslie King; Paul Rosenbloom provided that the per-copy fee indicated President Associate Art Director Board Members in the code is paid through the Copyright Vicki L. Hanson Margaret Gray Assistant Art Director Stefan Bechtold; Michael L. Best; Clearance Center; www.copyright.com. Vice-President Judith Bishop; Mark Guzdial; Cherri M. Pancake Mia Angelica Balaquiot Richard Ladner; Carl Landwehr; Subscriptions Secretary/Treasurer Production Manager Beng Chin Ooi; Loren Terveen; An annual subscription cost is included Elizabeth Churchill Bernadette Shade Marshall Van Alstyne; Jeannette Wing in ACM member dues of $99 ($40 of Past President Advertising Sales Account Manager which is allocated to a subscription to Alexander L. Wolf Ilia Rodriguez Communications); for students, cost Chair, SGB Board PRACTICE is included in $42 dues ($20 of which Jeanna Matthews Columnists Chair is allocated to a Communications Co-Chairs, Publications Board David Anderson; Phillip G. Armour; Stephen Bourne and Theo Schlossnagle subscription). A nonmember annual Jack Davidson and Joseph Konstan Michael Cusumano; Peter J. Denning; Board Members subscription is $269. Members-at-Large Mark Guzdial; Thomas Haigh; ; Samy Bahra; Peter Bailis; Gabriele Anderst-Kotis; Susan Dumais; Leah Hoffmann; Mari Sako; Terry Coatta; Stuart Feldman; Nicole Forsgren; ACM Media Advertising Policy Elizabeth D. Mynatt; Pamela Samuelson; Pamela Samuelson; Marshall Van Alstyne Camille Fournier; Benjamin Fried; Communications of the ACM and other Eugene H. Spafford ; Tom Killalea; Tom Limoncelli; ACM Media publications accept advertising SGB Council Representatives Kate Matsudaira; Marshall Kirk McKusick; CONTACT POINTS in both print and electronic formats. All Paul Beame; Jenna Neefe Matthews; Erik Meijer; George Neville-Neil; Copyright permission advertising in ACM Media publications is Barbara Boucher Owens Jim Waldo; Meredith Whittaker [email protected] at the discretion of ACM and is intended to provide financial support for the various BOARD CHAIRS Calendar items [email protected] CONTRIBUTED ARTICLES activities and services for ACM members. Education Board Change of address Co-Chairs Current advertising rates can be found Mehran Sahami and Jane Chu Prey [email protected] James Larus and Gail Murphy by visiting http://www.acm-media.org or Practitioners Board Letters to the Editor Board Members by contacting ACM Media Sales at Terry Coatta and Stephen Ibaraki [email protected] William Aiello; Robert Austin; (212) 626-0686. Elisa Bertino; Gilles Brassard; Kim Bruce; REGIONAL COUNCIL CHAIRS WEBSITE Alan Bundy; Peter Buneman; Carl Gutwin; Single Copies ACM Europe Council http://cacm.acm.org Yannis Ioannidis; Gal A. Kaminka; Single copies of Communications of the Dame Professor Wendy Hall Karl Levitt; Igor Markov; Bernhard Nebel; ACM are available for purchase. Please ACM India Council Lionel M. Ni; Adrian Perrig;Sriram Rajamani; contact [email protected]. AUTHOR GUIDELINES Srinivas Padmanabhuni Marie-Christine Rousset; Krishan Sabnani; http://cacm.acm.org/about- ACM China Council Ron Shamir; Josep Torrellas; COMMUNICATIONS OF THE ACM communications/author-center Jiaguang Sun Michael Vitale; Hannes Werthner; (ISSN 0001-0782) is published monthly Reinhard Wilhelm by ACM Media, 2 Penn Plaza, Suite 701, PUBLICATIONS BOARD ACM ADVERTISING DEPARTMENT New York, NY 10121-0701. Periodicals Co-Chairs 2 Penn Plaza, Suite 701, New York, NY RESEARCH HIGHLIGHTS postage paid at New York, NY 10001, Jack Davidson; Joseph Konstan 10121-0701 Co-Chairs and other mailing offices. Board Members T (212) 626-0686 Azer Bestavros and Gregory Morrisett Phoebe Ayers; Karin K. Breitman; F (212) 869-0481 Board Members POSTMASTER Terry J. Coatta; Anne Condon; Nikil Dutt; Martin Abadi; Amr El Abbadi; Sanjeev Arora; Please send address changes to Roch Guerrin; Chris Hankin; Carol Hutchins; Michael Backes; Maria-Florina Balcan; Communications of the ACM Yannis Ioannidis; Michael L. Nelson; Advertising Sales Account Manager ; Doug Burger; Stuart K. Card; 2 Penn Plaza, Suite 701 M. Tamer Ozsu; Eugene H. Spafford; Ilia Rodriguez Jeff Chase; Jon Crowcroft; Alexei Efros; New York, NY 10121-0701 USA Stephen N. Spencer; Alex Wade; [email protected] Alon Halevy; Sven Koenig; Steve Marschner; Keith Webster; Julie R. Williamson Tim Roughgarden; Guy Steele, Jr.; Printed in the U.S.A. Media Kit [email protected] Margaret H. Wright; Nicholai Zeldovich; ACM U.S. Public Policy Office Andreas Zeller 1701 Pennsylvania Ave NW, Suite 300, Washington, DC 20006 USA WEB T (202) 659-9711; F (202) 667-1066 Association for Computing Machinery Chair (ACM) James Landay

E R E C Computer Science Teachers Association 2 Penn Plaza, Suite 701 Board Members S Y A C E L L E Deborah Seehorn, New York, NY 10121-0701 USA Marti Hearst; Jason I. Hong; P

T E H Interim Executive Director T (212) 869-7440; F (212) 869-0481 Jeff Johnson; Wendy E. MacKay N I I S Z M A G A

4 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 from the co-chairs of the acm student research competition

DOI:10.1145/3145811 Laurie Williams and Doug Baldwin Highlights of the ACM Student Research Competition

INCE 2003, ACM in con- sentation of their work followed by a with me in my future academic and pro- junction with Microsoft five-minute question-and-answer ses- fessional endeavors.” have sponsored research sion to official SRC judges and any — Michele Hu, Cornell Tech competitions for under- conference attendees who attend the ASSETS 2015 graduate and graduate open session. At least five judges are Sstudents in computing. The com- assigned to the semifinals. “It was a great opportunity to be able petitions provide a vehicle for these 5. The conference’s top three final- to present at ACM SRC the work devel- students to present their original re- ists in each category are chosen based oped during a study abroad experience search before a panel of judges and on these presentations. in the USA. Working with students from attendees at well-known ACM-spon- 6. The first-place winners from other places with different backgrounds sored and co-sponsored conferences. each conference are invited to com- was an incredible experience. That was The students have the opportunity pete in the Grand Finals. These stu- my first time presenting at a conference to experience a research conference, dents submit an updated 4,000-word and it felt great to expose our research, get feedback on their research, meet paper on their research that is judged discuss it, and get invaluable feedback. academic and industrial researchers by a panel of experts. It was an amazing chance to train my and other students, and appreciate The winner of the Grand Finals research pitch and share ideas with oth- the practical applications of their re- and their advisors are invited to the ers.” search. Student competitors also have Annual ACM Awards Banquet, where — Clarissa Tuxen, the opportunity to sharpen commu- they are recognized for their accom- Fluminense Federal University nication, visual, organizational, and plishments and can witness other Grace Hopper 2016 presentation skills in preparation for luminaries in the computing field re- the SRC. Participation by undergradu- ceive prestigious society awards, such To learn more about the SRC, visit ates may be literally life-changing if as the ACM A.M. Turing award. The http://src.acm.org/. they alter their career path to pursue first-, second-, and third-place win- We acknowledge the dedication of graduate studies and research careers ners of each conference competition the many volunteers who make the after experiencing a conference and receive cash prizes, medals, ACM stu- SRC work. Each participating confer- competition. dent memberships, and recognition ence has a student research compe- The following process is used to se- among their peers and professors— tition chair and program committee lect the SRC winners: recognition that strengthens their ré- who review the papers submitted by 1. Each student submits an 800- sumés. All students who are selected the students. This committee and ad- word abstract of his or her research. to participate in the conference-level ditional volunteers judge the posters The abstract is evaluated by a mini- competition receive $500 travel fund- and presentations at the conference mum of three reviewers. Feedback ing for the conference. venue to select the winners. Evelyne on the abstract is provided to the stu- In 2016–2017, competitions took Viegas at Microsoft supports the SRC dents. place at 24 participating conferenc- program; she replaces Judith Bishop 2. The students submitting the es sponsored by the following ACM who was a dedicated Microsoft volun- highest-evaluated abstracts are invit- SIGs: SIGACCESS, SIGARCH, SIGCHI, teer for many years. Finally, Nanette ed to attend the conference and pres- SIGCOMM, SIGCSE, SIGDA, SIGDOC, Hernandez of ACM handles all logisti- ent their work. Typically, 10 graduate SIGGRAPH, SIGHPC, SIGMIS, SIG- cal aspects of running the SRC, such students and 10 undergraduate stu- MOBILE, SIGPLAN, SIGACT, SIGSAC, as sending materials to conference dents are invited to compete at the SIGSOFT and SIGSPATIAL. More than sites around the world, and handles conference. 330 students participated in these the interactions with students and 3. At the conference, the students competitions. conference volunteers. present their work in front of a post- Students find the SRC highly re- er. Any conference attendee may warding. Representative comments Laurie Williams, a professor of computer science in the College of Engineering at North Carolina State come and ask a student about their include the following: University, Raleigh, and Doug Baldwin, a professor research. A minimum of five official “Participating in the SRC was an of at the State University of New York at Geneseo, serve as co-chairs of the ACM Student SRC evaluators assesses each poster. amazing opportunity. It was my first Research Competition. 4. The top five graduate and top time attending any conference, and it five undergraduate students advance really showed me how to pitch my re- to the semifinals. In the semifinals, search project, and interact with other each student makes a 10-minute pre- researchers. I will carry this experience Copyright held by authors.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 5 cerf’s up

DOI:10.1145/3148147 Vinton G. Cerf Heidelberg Laureate Forum It is fall in Heidelberg and the leaves Once again, a boat trip on the Neckar River afforded several hours of free dis- on the trees are already turning. This is the cussion among laureates and students fifth year of the Heidelberg Laureate Forum amidst the beautiful fall colors in the trees along the journey. A Bavarian (http://www.heidelberg-laureate-forum.org/) night and a visit to the amazing Speyer Museumd also provided ample interac- and it continues to be a highlight of the tion of transparency of machine learning tive opportunities in addition to which year for me and for about 250 others who in which it is made clear where the deep the Forum added office hours, stu- participate. This year, computer science learning takes place and what has been dent working groups, and discussion was heavily represented. There were learned remains unanswered, as does an groups to stimulate student/laureate fewer , but they made easy way to explain how the system has interactions. A final visit to the 17th- up for smaller numbers by their extraor- reached its conclusions. Mathematical century castle overlooking Heidelberge dinary qualifications. A new cohort of and computational models of biological rounded out the week. laureates was added this year: recipients processes found their way onto the agen- By chance (or perhaps by careful plan- of the ACM Prize for Computing.a da, reinforcing the notion that we might ning), the celebration of the linking of Quantum computing and machine someday understand better how these Heidelberg to Palo Alto, CA, as sister cit- learning were major foci of attention processes work by modeling them and ies was hosted in the Town Hall of Hei- and the Hot Topic sessions drew on predicting some of their behaviors. delberg. A number of HLF participants speakers in the research and private Visits to local research and business live or have lived in the Palo Alto area sectors beyond the normal cohort of centers were arranged, such as the Euro- and represented the city along with its laureates. I took several lessons away pean Microbiology Laboratory (EMBL), mayor and other dignitaries. Of course, from the quantum computing discus- European Media Laboratory, German the Mayor of Heidelberg and the City sions: serious progress is being made in Cancer Research Center, Heidelberg Council were in attendance. The stained multi-qubit hardware development; the Institute for Theoretical Studies, the glass walls of the ceremonial room of somewhat inaptly named “Quantum Max Planck Institute for Astronomy, the Town Hall had dates in the 1300s— Supremacy” challengeb is being met by NEC, the Mathematics Center of the something that is fairly mind-boggling more than one organization; and pro- University of Heidelberg, and the In- for Americans whose sense of national gramming languages for quantum com- terdisciplinary Center for Scientific history tends to start in the late 1700s! puting are being developed.c Computing and to SAP headquarters. I This event continues to represent The other hot topic was machine learn- spent at least two hours with a research- a remarkable gathering of brilliant ing and and impres- er at EMBL discussing in depth intra- and energetic minds and ACM and sive results were outlined for image recog- and inter-cellular communication. the other participating organizations nition, speech understanding, machine owe a great debt to the Klaus Tschira translation, and self-supervised learning. Foundation and the Heidelberg Laure- Multilayer neural networks are produc- Quantum computing ate Foundation for their adept and ef- ing interesting results such as the abil- ficient support for this signature meet- ity to turn an ordinary photograph into and machine learning ing. We remember gratefully Klaus one that inherits the style of various well- were major foci Tschira, a founder of the German SAP known artists or schools of art. The ques- company, who passed away far too of attention at this soon in 2015. a Formerly known as the ACM-Infosys Founda- year’s Heidelberg d https://speyer.technik-museum.de/en/ tion Award in the Computing Sciences. e https://en.wikipedia.org/wiki/Heidelberg_Castle b That is, quantum computers in the 49–50+ qu- Laureate Forum. bit range that can theoretically perform func- tions beyond the capacity of current or project- Vinton G. Cerf is vice president and Chief Internet Evangelist ed conventional classical computing systems. at Google. He served as ACM president from 2012–2014. c https://en.wikipedia.org/wiki/Quantum_ programming#Quantum_computing_language Copyright held by author.

6 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 vardi’s insights

DOI:DOI:10.1145/3144590 Moshe Y. Vardi

When it comes to historical inter- Would Turing Have Won pretation, the same facts may lead different people to different inter- the Turing Award? pretations, but one should pay atten- tion to the facts! In August 2017, Leo N 2017, WE celebrated 50 years tional Lecture after Turing. Today, Tur- Corry published an article in Commu- of the ACM A.M. Turing Award, ing is widely regarded as one of the most nications on “Turing’s Pre-War Ana- known simply as the Turing outstanding scientists of the 20th cen- log Computers: The Fatherhood of Award. The list of Turing Award tury, but that was not the case in 1966. the Modern Computer Revisited” winners (http://amturing.acm. The question, therefore, can be posed as (https://goo.gl/M7jCaj) in which he Iorg), starting from in 1966, follows: Had Turing been alive in 1966 carefully examined the purported “for his influence in the area of ad- (he died in 1954), would he have been se- connection between the “Univer- vanced computer programming tech- lected for ACM’s first National Lecture? sal Turing Machine,” as introduced niques and compiler construction,” A debate about Turing’s accom- in Turing’s 1936 paper and the design to Sir Tim Berners-Lee in 2016, “for in- plishments has been going on for and implementation in the mid-1940s venting the , the first quite a while. In 1997, in an after-din- of the first stored-program computers. Web browser, and the fundamental ner speech in Cambridge, U.K., Mau- He concluded “There is no straightfor- protocols and algorithms allowing the rice Wilkes, the 1967 Turing Award ward, let alone deterministic, historical Web to scale,” offers a bird-eye view of winner (for designing and building path leading from Turing’s 1936 ideas the highlights of computing science the EDSAC, the first stored-program on the Universal Machine to the first and technology over the past 50 years. computer in 1949), offered some bit- stored-program electronic computers of Justifiably, the Turing Award is often ing comments about Turing: “How- the mid-1940s.” accompanied by the tagline “The No- ever, on a technical level, of course I But the debate about how much bel Prize in Computing.” How did this did not go along with his ideas about credit Turing should get for the idea prestigious award come to be? computer architecture, and I thought of the stored-program computer di- The early history of the Turing Award that the programming system that he minishes, in my opinion, from Tur- is somewhat murky. The minutes of introduced at Manchester University ing’s actual contributions. The Tur- meetings of ACM Council from the mid- was bizarre in the extreme. … Turing’s ing Machine model offered a robust 1960s shed some, but not complete light work was of course a great contribu- definition of computability that has on this history. The Turing Award was tion to the world of mathematics, but been studied, refined, and debated not originally created as a “big prize,” there is a question of exactly how it is since 1936, giving rise in the 1960s but rather a lecture given at the annual related to the world of computing.” to computational complexity the- ACM meeting. In August 1965, ACM (See Wilkes’s complete comments at ory, a gem of theoretical computer Council considered and tabled a pro- https://goo.gl/XkjM7n.) science. Turing’s philosophical ex- posal that “the National ACM Lecture The controversy about Turing’s ac- amination in 1950 of the possibility be named the Allen [sic] M. Turing Lec- complishments flared again over the of machine intelligence is lucid and ture.” In December 1965, ACM Council last few years. In a 2013 Communica- incisive today as it was then. Finally, adopted the motion that “A.M. Turing tions’ editorial (https://goo.gl/SpkhKw) we learned in the 1970s about Tur- be the name of the National Lectureship I argued that “The claims that Turing ing’s critical contributions to comput- series.” In a 1966 meeting, ACM Coun- invented the stored-program computer, ing-aided code breaking. cil voted to name Alan Perlis as first lec- which typically refers to the uniform Would Turing have won the Turing turer. The minutes shed no light on why handling of programs and data, are sim- Award? My answer is, he should have! the lectureship was named after Alan ply ahistorical.” In response to this edi- Follow me on Facebook, Google+, Turing. The historical record is also not torial, Copeland et al. argued in the 2017 and Twitter. clear on how a lectureship turned into a Turing Guide (https://goo.gl/DjC8uk) major award. Perhaps there is a lesson that “Vardi is ignoring the fact that Moshe Y. Vardi ([email protected]) is the Karen Ostrum George Distinguished Service Professor in Computational here for ACM to keep better minutes of some inventions belong equally to the Engineering and Director of the Ken Kennedy Institute for its Council’s meetings! realm of mathematics and engineer- Information Technology at Rice University, Houston, TX. He is the former Editor-in-Chief of Communications. From today’s perspective, however, ing. The Universal Turing Machine was we can wonder whether ACM Council one such, and this is part of its bril- was justified in 1966 in naming its Na- liance.” So who is right? Copyright held by author.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 7 letters to the editor

DOI:10.1145/3145540 They See What You See

NDREW CONWAY’S AND Pe- countries but so far has had only lim- framing its context for a wider societal ter Eckersley’s Viewpoint ited effect because these countries are discussion we consider essential. “When Does Law En- not sufficiently powerful individually Andrew Conway, Melbourne, Australia, forcement’s Demand to to enforce sanctions against large mul- and Peter Eckersley, San Francisco, CA Read Your Data Become tinational foreign-domiciled IT provid- Aa Demand to Read Your Mind?” (Sept. ers. However, if Burr-Feinstein does in- 2017) was an important contribution deed become law, then these countries Bring ‘Diseased’ to the ongoing debate over electronic might be more able to pursue mass Software Up to Code backdoors, whereby a backdoor is a surveillance domestically, as IT com- We agree with Vinton G. Cerf’s advice means for accessing and exfiltrating panies could lose much of the legal in his Cerf’s Up column “Take Two user information not specifically au- grounds they would need to resist. Aspirin and Call Me in the Morning” thorized in advance by users. Here, I Conway and Eckersley also did not (Sept. 2017) that we all practice better would like to outline several key as- mention a near-term technology that “cyber-hygiene” but must quarrel with pects of that debate that also need to might be used to implement highly the continued use of public health as be addressed. secure backdoors in IoT devices by a metaphor for cyber security. If we Although Conway and Eckersley requiring that each device have a dif- as computing professionals intend to did discuss the possibility that law ferent public key that could enable improve the cybersecurity of our criti- enforcement could gain access to our government security services to take cal infrastructures, rather than merely most private thoughts, they did not over the device.1 Even if hackers pen- tolerate their current “diseased” state, mention a crucial near-term technol- etrated the security of a government- we must think differently. We thus ogy through which this exfiltration installed virtual machine for a device, propose a return to an older metaphor could happen. Within the next 10 they would gain no lasting advantage for software, likening its structures years, “hologlasses,” or holographic hacking additional devices. to physical structures and its archi- glasses, are projected by Apple, Face- Finally and most important, no tecture to the architecture of physical book, Google, Microsoft, and Sam- mention was made of a technology buildings. Such thinking suggests we sung, along with numerous startups, proposal1 that could ameliorate some consider how to build software that to become almost as common as cell- of the negative effects of mass surveil- will not fall over when attacked or phones are today, as reflected in the lance, whereby citizens’ most sensitive build it from weak materials unable to scale of their investment in its devel- information is stored on their own de- bear expected stress. opment. A backdoor in hologlasses vices, provided personal IoT devices Software we rely on for critical func- could enable a “we see and hear what include protection against self-incrim- tions (such as controlling medical de- you see and hear” capability that ination. By storing sensitive informa- vices, delivering electrical power to would provide extraordinary insight tion on these devices, that information households, and guiding automobiles) into what users are thinking, as well could be protected from the kind of ef- must conform to an appropriate set of as how they are behaving online and forts Conway and Eckersley identified. constraints, just as physical structures even in the physical world. conform to building codes before they Also not mentioned was a legislative Reference can be occupied. A third party must be proposal that could facilitate manda- 1. Hewitt, C. Islets protect sensitive IoT information: able to certify conformance to these Verifiably ending use of sensitive IoT information for tory backdoors for Internet of Things mass surveillance can foster (international) commerce constraints, just as building inspectors devices. In 2016, Senators Richard Burr and law enforcement. Social Science Research Network certify buildings. WP 2836282; https://papers.ssrn.com/sol3/papers. (R., North Carolina) and Dianne Fein- cfm?abstract_id=2836282 These codes are best developed stein (D., ) introduced leg- by those who build the systems, not islation—The Compliance with Court Carl Hewitt, Palo Alto, CA by government, though governments Orders Act—in the U.S. Senate to man- might use them once they are in place. date providers of information products An industry-consensus building code, and services also provide unencrypted Authors Respond: with third-party assessment of con- information on IoT devices to the gov- We generally agree with Hewitt who offers formance, can help the marketplace ernment pursuant to court order. The definite specific instances of the general reward those who build systems with result could be “Nothing is Beyond Our issues we covered. Virtual reality, in fewer vulnerabilities. Reach,” or no information is beyond particular, is, as he implies, a valuable window Over the past few years, with sup- the reach of law enforcement, likewise into the mind that also involves important port from the IEEE’s Cybersecurity pursuant to court order. Similar legis- technical and legislative dynamics. What to Initiative and the National Science lation has been adopted in Australia, do about it is a complex question we did not Foundation, workshops have been France, Germany, the U.K., and other attempt to answer in our Viewpoint beyond held to begin to develop such build-

8 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 letters to the editor ing codes for medical-device software have reliable ways of producing the kind plicitly tied the computing elements for and for power-system software.1,2 In of code Landwehr et al. and I likely agree the EDVAC to McCulloch’s and Pitts’s addition to these draft codes, related we need and want, we may also want to model of a neuron, saying: promising developments include argue that infected or vulnerable devices “Following W. Pitts and W.S. Mac- Consumers Reports’ collaboration ought not to be naively tolerated and Culloch [1943] … we ignore the more with the Cyber Independent Testing that owners (and suppliers) bear at least complicated aspects of neuron func- Laboratory (http://cyber-itl.org/) to some responsibility for observing diligent tioning … It is easily seen, that these develop methods for publicly rating software hygiene. simplified neuron functions can be software products, and UL’s (http:// Vinton G. Cerf, Mountain View, CA imitated by telegraph relays or by vac- www.ul.com) development and use of uum tubes … We propose to use them a standard for certifying cybersecurity accordingly for the purpose described assurance of products. A Neuron Net Is there: as the constituent elements Treating software security as a pub- Also a Turing Machine of the devices, for the duration of lic health problem is not likely to lead Leo Corry’s article “Turing’s Pre-War the preliminary discussion … The past the decades-old ideas of aftermar- Analog Computers: The Fatherhood element which we will discuss, to ket vaccines, antivirus, and quaran- of the Modern Computer Revisited” be called an E-element … which re- tine. Providing evidence that software (Aug. 2017) described the Turing ma- ceives the excitatory and inhibitory is at least free of specified classes of chine as a purely mathematical notion. stimuli, and emits its own stimuli vulnerabilities covered by an appropri- While Corry’s argument was persua- along a line attached to it. … In all ate building code can yield a more ef- sive, there is indeed a direct connec- this we are following the procedure fective market incentive for companies tion from Turing’s construction of the of W. Pitts and W.J. MacCulloch.”2 to produce the cyberinfrastructures we Turing machine to the Electronic Dis- Since McCulloch and Pitts had all need—and that are up to code. crete Variable Automatic Computer shown that neuron nets are universal (EDVAC) in the 1940s via the McCull- computing machines in the sense of the References och-Pitts model of the brain. For ex- Church-Turing , the same connec- 1. Haigh, T. and Landwehr, C. A Building Code for Medical Device Software Security. Technical Report. IEEE ample, during the discussion portion tion to universal computing machines Computer Society, Mar. 2015; https://www.computer. of a 1951 talk by John von Neumann, would apply to the EDVAC. org/cms/CYBSI/docs/BCMDSS.pdf 2. Landwehr, C.E. and Valdes, A. Building Code for Power neuroscientist Warren S. McCulloch System Software Security. Technical Report. IEEE described the influence of Turing’s Computer Society, Mar. 2017; https://www.computer. References org/cms/CYBSI/docs/BCPSSS.pdf original 1936 paper, saying: 1. von Neumann, J. The general and logical theory of automata. Chapter in Cerebral Mechanisms in “… I came, from a major interest Behavior, L.A. Jeffress, Ed. The Hixon Symposium, Robert K. Cunningham, Lexington, MA, in philosophy and mathematics into John Wiley & Sons, New York, 1951, 1–41. Tom Haigh, Minneapolis, MN, 2. von Neumann, J. First Draft of a Report on the EDVAC psychology with the problem of how (June 30, 1945). Reprinted as a chapter in Papers Carl Landwehr, New Buffalo, MI, a thing like mathematics could ever of John von Neumann on Computing and Computer Theory, W. Aspray and A. Burks, Eds. MIT Press, and Alfonso Valdes, Urbana, IL arise—what sort of a thing it was … Cambridge, MA, 1987, 17–82. The attempt to construct a theory in a field like [neurophysiology], so that it Brad Barber, Arlington, MA Author Responds: can be put to any verification, is tough It is always a pleasure to hear from Carl … it was not until I saw Turing’s paper Landwehr with whom I have had a long that I began to get going the right way Author Responds: acquaintance and for whom I have great around, and with [logician Walter] My entire argument was about Alan respect. An interesting challenge with his Pitts’ help formulated the required Turing’s own views prior to the war, building/architecture metaphor relates logical calculus. What we thought we not about his influence on later to the way software is often constructed were doing (and I think we succeeded developments. But a more general point these days by incorporating (vast) libraries fairly well) was treating the brain as a I wanted to make was that scientific and of code reflecting, perhaps, uncertain Turing machine.”1 technological ideas develop historically provenance. There is also the uncertainty McCulloch’s and Pitts’s 1943 paper and that what happens later sometimes of software interactions across the emphasized the equivalence of artificial misleads us when we try to understand network that may never have been tested neuron nets to Turing machines, say- what happened earlier on. This may until a chance encounter leads to a breach. ing: also be the case with McCulloch’s very None of this invalidates the building- “It is easily shown: first, that every interesting, retrospective testimony, which code metaphor but might make it more [neuron] net, … can compute only such by all means deserves a critical eye. difficult to establish that the ensemble numbers as can a Turing machine … Leo Corry, Tel Aviv, Israel meets the desired code standards and This is of interest as affording a psy- properties. I am, in fact, very interested chological justification of the Turing in the development of programming aids definition of computability and its Communications welcomes your opinion. To submit a Letter to the Editor, please limit yourself to 500 words or less, and that will do a much better job of assessing equivalents, Church’s λ-definability send to [email protected]. source code against desirable properties and Kleene’s primitive recursiveness.” of attack resistance and identifying Moreover, in the first draft of the de- potential sources of weakness. Until we sign for the EDVAC, von Neumann ex- ©2017 ACM 0001-0782/17/11

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 9 The Communications Web site, http://cacm.acm.org, features more than a dozen bloggers in the BLOG@CACM community. In each issue of Communications, we’ll publish selected posts or excerpts.

Follow us on Twitter at http://twitter.com/blogCACM

DOI:10.1145/3137627 http://cacm.acm.org/blogs/blog-cacm Opportunities for Women, Minorities in Information Retrieval Mei Kobayashi describes activities to support diversity and inclusion at the annual meeting of the ACM Special Interest Group on Information Retrieval in Tokyo this summer.

Mei Kobayashi just before the welcome party. A week ˲˲ Maria Maistro, a Ph.D. student at the SIGIR 2017: Diversity before the conference, I received an e- University of Padua, Italy, who works on and Inclusion mail from the secretary of the session, IR evaluation with a focus on query log http://bit.ly/2fFB1Uh Maram Hasanain, a graduate student analysis to integrate user search behavior August 13, 2017 in computer science (CS) at Qatar Uni- into the evaluation process. Maria is a co- versity, asking if I would like to prepare author of On Including the User Dynamic in Hello and welcome! a one-minute introduction of myself Learning to Rank (http://dl.acm.org/cita- Diversity was a central theme in for the session. I was so overwhelmed tion.cfm?id=3080714) at SIGIR 2017. the ACM SIGIR 2017 held in Shinjuku by her beautifully written e-mail, ˲˲ Nazli Goharian, professor of com- Ward in Tokyo, Japan. Upon arrival, and the excitement of a first-time con- puter science at Georgetown Univer- all registrants were given a beautiful tact with someone from Qatar, that I sity, who works on health search and keychain and card as commemora- immediately accepted her invitation. mining, including clinical and radio- tive gifts from the local organizers to The session started with one-min- logical reports, social media posts for celebrate the 40th anniversary of the ute presentations by: mental health and adverse drug reac- conference series: ˲˲ Vanessa Murdock, Principal Ap- tions, summarization, and decision “… the 40th Anniversary Logo…fea- plied Researcher at Cortana Research, support systems. She is a co-author of tures Mt. Fuji, a view of Shinjuku sky- Microsoft, USA. Contextualizing Citations for Scientific scrapers, including the Tokyo Metropoli- ˲˲ Grace Hui Yang, associate profes- Summarization using Word Embeddings tan Government (Office), as seen from sor at Georgetown University. and Domain Knowledge (https://arxiv. Keio Plaza the conference hotel, and fire- ˲˲ Sahar Asadi of Spotify (digital mu- org/abs/1705.08063) at SIGIR 2017. works celebrating the 40th anniversary. sic service), USA. ˲˲ Tingting Dong, researcher at Sys- The colorfulness of the fireworks and the ˲˲ Mei Kobayashi, an applied math- tem Platform Research Laboratories, circles within and enclosing the logo rep- ematician turned big data algorithms NEC, Japan, who works on diversifica- resent diversity and inclusion.” specialist. She is the first female man- tion and summarization of video search SIGIR 2017 featured a session on ager in Customer Service of NTT Com- results to provide well-organized and Women in IR (Information Retrieval) munications, Tokyo. intuitive views for users. organized by Laura Dietz of the Univer- ˲˲ Zehong Tan, Senior Software Engi- ˲˲ Harumi Murakami, professor of sity of New Hampshire on the first day, neer on the Search Team for eBay, USA. the Graduate School for Creative Cit-

10 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 blog@cacm ies and Vice Director of Media Center, into one main message and present Osaka City University. it succinctly, in easily digestible form ˲˲ Xiaolu Lu, graduate student at Roy- The diversity among for the audience. Maistro noted that al Melbourne Institute of Technology. the presenters the short talks at the beginning of She is a co-author of: Can Deep Effective- poster sessions were also effective, ness Metrics be Evaluated Using Shallow dispelled any notions and they provided seeds for breaking Judgment Pools? (http://dl.acm.org/cita- of a fixed template for the ice during the remainder of the tion.cfm?id=3080793) at SIGIR 2017. conference. For myself, it was a lot of ˲˲ Zhuyun Dai, graduate student at success or stereotype fun meeting male and female audi- Carnegie Mellon. She is a co-author of of a woman in ence members during coffee breaks two papers at SIGIR 2017: Learning to and evening parties who came up and Rank Resources (http://bit.ly/2vMswxS ) the sciences. introduced themselves. and End-to-End Neural Ad-hoc Ranking A big thanks to Dietz and Hasainan with Kernel Pooling (https://arxiv.org/ for putting together this session, and abs/1706.06613). to the SIGIR Organizing Committee ˲˲ Maram Hasanain, a Ph.D. student the end. Among all of the presen- Members who did not schedule com- in computer science at the College of tations, Hasanain’s received the peting parallel technical sessions so Engineering, Qatar University. greatest round of applause by far, all could attend Women in IR without Murdock kicked off the lightning proving to all the possibility of suc- fear of missing out on learning about talks and got people chuckling with cessful remote e-participation! new technical work. Dietz had a busy the closing line, “We’re hiring!” For bet- The second part of the session was week: she co-organized a full-day ter or for worse, it became a common a longer presentation by Hannah Bast workshop (one of eight workshops at theme among presenters from indus- of University of Freiburg, on identi- SIGIR 2017) with Edgar Meij (Bloom- try. Yang made everyone feel fantastic fication of discrimination and ste- berg) and Chenyan Xiong (Carnegie with her empowering statement, “I just reotyping in the workplace. She pre- Mellon University) on Knowledge became associate professor!” The audi- sented types of follow-up actions an Graphs and Semantics for Text Retrieval ence broke out in a big, round of ap- individual can take, such as, pointing and Analysis. And as the first Student plause. I pointed out my chemistry and out specific inappropriate behavior(s) Affairs Chair for SIGIR, she organized applied mathematics background and or statistics to decision-makers, and a student buddy program for first- stated, “Researchers in IR come from proposing concrete methods to cor- time attendees of SIGIR, and an in- very diverse backgrounds, not just NLP.” rect the situation. Although some in expensive and enjoyable get-together As it turned out, there was another the audience may have attended simi- featuring a karaoke stage at a nearby in the crowd: Maistro. lar sessions at other conferences, for food court for students at the begin- And Murakami, a SIGIR 2017 Commit- some first-time attendees, the session ning of the conference. tee Member, said her interest in work- was an eye-opener. Royal Sequeira, a Final Note: To promote awareness ing in IR is from a psychology perspec- graduate student at the University of of and respect for diversity—including tive, her original area of expertise. Waterloo in Canada, sent me e-mail geographic, religious, and dietary, as Side note: In addition to her work as saying, “Women in IR is one of my best well as gender diversity—this year’s a faculty member at Osaka City Univer- experiences at SIGIR this year. While it SIGIR Conference featured a Diver- sity, Murakami served as one of three has made me proud about the women sity and Inclusion Luncheon with de- SIGIR 2017 Social & Publicity Chairs, in our community, it has also led me to licious and healthy vegetarian and the other two being Yukino Baba of introspect on several aspects of equity Halal dishes. Since it took place on Kyoto University, and Falk Scholer of and diversity.” The session was educa- the third day of the conference, huge RMIT. During the SIGIR Business tional for me as well. I was shocked to boats of fresh vegetables and fruit Meeting, chaired by Diane Kelly of learn that women were given the right were welcomed by all participants. the University of Tennessee, Knox- to vote in all Swiss Cantons in 1991. All meals during the conference—in- ville, attendees learned that SIGIR On a more positive note, it was ex- cluding the main banquet at the his- 2017 had been expected to run a sub- citing to hear about the work of wom- toric Hotel Chinzanso—were clearly stantial deficit due to its expensive en from so many different countries, marked vegetarian, Halal, etc. They venue, Tokyo. However, the outstand- backgrounds, and age groups. The were buffet or boxed to enable every- ing work of the Social & Publicity Chairs diversity among the presenters dis- one to sit together, chat and mingle. brought a record-breaking number of pelled any notions of a fixed template To the Conference Organizers, par- registrants (more than 200% of initial for success or stereotype of a woman ticularly the Social Chairs for their estimates) and drove conference cof- in the sciences. All of us need to en- thoughtfulness and impeccable orga- fers solidly into the black. sure that anyone with the interest, nization skills—Cheers! The closing one-minute presen- passion, and will to work hard can suc- tation by Hasanain was a short vid- ceed. The consensus among attend- Mei Kobayashi is manager, Data Science/Text Analysis at eo presentation. It was so impres- ees I talked with was very short pre- NTT Communications. sive and heart-warming that I felt sentations make a lasting impression; as though I knew her personally by speakers have to distill their thoughts © 2017 ACM 0001-0782/17/11 $15.00

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 11 news

Science | DOI:10.1145/3137674 Neil Savage A Block on the Old Chip N Block copolymers may help transistors shrink to tinier dimensions. OR DECADES, COMPUTERS have academia are trying to perfect a proc- called block copolymers arrange them- grown more powerful because ess that would let chemicals arrange selves into a desired pattern. chipmakers have been able to themselves into tiny lines to serve as a Block copolymers consist of two make ever-smaller transistors, pattern for the circuits. The lines and different materials that do not like to allowing them to cram more spaces in most chips now are on the or- mix, like oil and water, but they are Fonto a single chip. That steady march der of 40 nanometers (nm) wide, but bonded together, so they cannot fully has always depended on optics—short- are expected to drop to less than 10 nm separate. When liquid block copoly- er wavelengths of light allowed chip- within a few years. mers are heated, they form a structure makers to draw smaller lines for circuit “Ten nanometers is really where in which each polymer has the least paths, which then can be closer togeth- they’re going to be forced to imple- possible contact with the other. They er. It has become increasingly harder, ment some new technology,” says might, for instance, form alternating however, to reach the high resolutions Christopher Ober, professor of materi- stripes of each polymer, or create a needed for ever-tinier features. als engineering at in checkerboard pattern, depending on The answer, or at least part of it, may Ithaca, NY. The technology Ober and how they were designed. lie not with optics at all, but with chem- others are interested in is directed self- In chip manufacturing, that would istry. Researchers in both industry and assembly (DSA), in which materials mean coating the silicon wafer with a

Lithographically-defined substrate Self-Assembly of diblock copolymers Directed Self-Assembly

Self-assembly of block copolymers on a flat substrate normally create fingerprint-like patterns having nanodomains with no long-range order (left). To fabricate well-aligned nanostructures over a large area, the self-assembly of the block copolymers can be induced in a

confined space on the sub-micrometer scale. UNIVERSITY NATIONAL SEOUL LABORATORY, AND SOFT NANOMATERIALS OF POLYMERIC COURTESY IMAGES

12 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 news neutral chemical to act as a base, then polymers can perform the same self- adding the block copolymer on top, assembly as PS-PMMA, but at a smaller and heating the whole thing. The poly- Because self- size. Ellison’s group has created struc- mers would form alternating lines— assembly is a tures just four to five nm wide. think of them as red stripes and blue The trouble with high-chi copoly- stripes—and one set, say the red, would thermodynamic mers is that it can be difficult to get be washed away. The remaining blue process, it is them to stand up straight. To form par- stripes would act as a pattern for the allel strips, they need to orient them- same etching process used in current controlled by selves perpendicular to the surface photolithography to inscribe circuits differences in surface they are on. Often, however, they will into silicon. flop over onto their sides, destroying Photolithography also relies on energy, and the top the pattern. chemicals—photoresists that are pat- “surface” is open air. Because self-assembly is a thermo- terned through exposure to light, creat- dynamic process, it is controlled by ing a pattern that is used to define which differences in surface energy, and the parts of the wafer to etch. Many modern top “surface” is open air. Some re- designs, however, require multiple searchers, such as Ellison and his col- masks, driving up costs. The machines Over time, Ronse says, researchers league C. Grant Willson, a professor for focusing the light beams to ever- managed to reduce the level of de- of chemical engineering at the Uni- smaller sizes are also expensive, so DSA fects, while at the same time, EUV was versity of Texas at Austin, have been could, the thinking goes, create smaller beginning to look more promising. working on topcoats, films of chemi- features and do it more cheaply. The combination made it seem less cals that change the surface energy A few years ago, DSA was seen as the likely that DSA would provide the low- on top of the copolymer so that it next big thing, and many chipmakers cost alternative the industry sought. aligns itself correctly. were investing heavily in it, says Kurt “After a couple of years of develop- Paul Nealey, a professor of molecular Ronse, director of the advanced lithog- ment, the window of opportunity for engineering at the University of Chicago raphy research program at the Inter- 28nm pitch was starting to close,” and one of the pioneers of DSA, says university Microelectronics Center Ronse says. there are three or four methods to con- (imec) in Leuven, Belgium. The focus Though the hype has faded, he says, trol the orientation of high-chi materi- was on a block copolymer consisting of work continues. Some of the compa- als. Nealey and Karen Gleason, a profes- polystyrene and polymethyl methacry- nies imec works with have ended their sor of chemical engineering at the late, or PS-PMMA, which was targeted DSA research, while others are still go- Massachusetts Institute of Technology, for features with a 28nm pitch, in ing, though they have directed some of Cambridge, MA, published a paper in which lines and spaces are each 14nm their chemistry work toward develop- Nature Nanotechnology in March in wide. At the time, the anticipated next ing photoresists for EUV. “They have which they reported using a technique photolithography technology was ex- not stopped DSA development, but by called initiated chemical vapor deposi- treme ultraviolet (EUV), but progress putting less effort in it, it definitely tion to create a topcoat that forced the toward making it commercial seemed slows down,” Ronse says. block copolymer to line up the way they to be stalled. While research into DSA is not as ag- wanted it. The approach allowed them gressive as it once was, it has not to create features only 9.3nm wide. Too Many Defects stopped, says Christopher Ellison, as- Another approach exposes the co- The hype over DSA did not last, says sociate professor of chemical engi- polymer to vapor from a solvent to Ronse. “Very quickly it turned out that neering at the University of Minnesota. change the balance of energy. A tech- one of the big issues was going to be “I don’t think it’s going away,” he says. nique explored by IBM mixes in an ad- defects,” he says. Sometimes the co- “It’s not growing at a rapid rate.” ditive to the copolymer that changes polymer would leave a gap, or the Part of the reason for that slowdown how it responds to surface energy at stripes would get too close to each may be that it is not the optical physics the air interface. other or even cross, a phenomenon that the chipmaking industry has al- However they are controlled, block called “line edge roughness.” The un- ways relied upon. “It makes it challeng- copolymers cannot do everything on dercoating might have a pinhole or a ing for industry to accept because it’s their own. Some lithography must be slight bump that would throw the so radically different,” Ellison says. used to inscribe guiding lines for them process off. to follow—that is the “directed” part of “It uses thermodynamics to form Surface Tension DSA. Because the material forms lines the patterns,” says Charles Black, di- While PS-PMMA may seem less attrac- much thinner than those inscribed, rector of the Center for Functional tive than it once did because its resolu- they require less-advanced photolithog- Nanomaterials at Brookhaven Nation- tion limit is 11 nm, chemists are pursu- raphy processes. At some resolutions, al Laboratory in Upton, NY. “It usually ing other block copolymers that could the guiding pattern could be created by ends up with a defect here or a defect go even smaller. Some materials have a the current state-of-the-art process, there, and microelectronics is really in- high chi, a measure of how incompati- 193nm immersion lithography, which tolerant of that.” ble the two polymers are. High-chi co- uses a liquid to focus the light beam to a

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 13 news

smaller area. It could appeal to industry of shapes and forms you can produce,” to use DSA with immersion lithography he says. “People have got very clever at Further Reading equipment that is already paid for, rath- taking these limited set of shapes and Self-Assembly: Lego Blocks in Nature er than needing to invest in new equip- learning how to arrange them.” Paul Nealey, University of Chicago ment for EUV. The extra expense of DSA Researchers will continue to push https://vimeo.com/138700499 is minimal, Ober says; “It’s just an extra ahead with DSA, Ronse says, seeking Suh, H.S., Kim, D.H., Moni, P., Xiong, S., Ocola, bottle of photoresist.” the right blend of polymers, working L.E., Zaluzec, N., Gleason, K., and Nealey, P.F. In fact, EUV and DSA may turn out on ways to keep the chemicals consis- Sub-10-nm patterning via directed to be complementary technologies, tent from one batch to the next, and self-assembly of block copolymer films with a vapour-phase deposited topcoat, each doing things that are difficult for finding ways to bring down the number Nature Nanotechnology 12, 2017 the other. The first use for DSA, the re- of defects to acceptable levels. http://www.nature.com/nnano/journal/v12/ searchers say, will not be for drawing At the same time, the industry will n6/full/nnano.2017.34.html lines at all, but for controlling the size push forward with EUV. The technology Sinturel, C., Bates, F.S., and Hillmyer, M.A. of the holes through which different has failed to materialize despite a cou- High χ−Low N Block Polymers: How Far layers of transistors are connected. ple decades of development, but it may Can We Go?, ACS Macro Letters 4, 2015 Conventional lithography makes those have turned a corner; “Although it Jiang, J., Jacobs, A., holes, or vias, too large. They could be looked like it was never going to hap- Thompson, M. O., and Ober, C. K. filled with a block copolymer, which pen, now it looks really, really close,” Laser spike annealing of DSA photoresists, J Photopolymer Science and Technology 28, 2015. would assemble to form a narrower Ober says. https://www.jstage.jst.go.jp/article/ channel through which connections While it is not certain what mix of photopolymer/28/5/28_631/_article could be made. technologies will succeed, the re- Lane, A.P., Maher, M.J., Willson, C.G., searchers are not worried about reach- and Ellison, C.J. Design Challenges ing the limits of semiconductor tech- Photopatterning of Block Copolymer Thin One shortcoming of DSA is that it can- nology any time soon. “The demise of Films, ACS Macro Letters 5, 2016 not inscribe sharp 90-degree turns, photolithography has been predicted http://pubs.acs.org/doi/abs/10.1021/ acsmacrolett.6b00075 while photolithography can. Any of forever,” Black says. “Probably for 20 these technologies will impose some years people have been saying ‘the end Neil Savage is a science and technology writer based in constraints on chip designers, but chip is here,’ and those photolithography Lowell, MA. architects have learned to deal with the guys have been coming up with ways to existing ones, Ober says. “You have a set get around it.” © 2017 ACM 0001-0782/17/11 $15.00

Milestones Computer Science Awards, Appointments

ACM, IEEE POSTHUMOUSLY award. ACM and IEEE Computer Conference series, the ACM IEEE- Smith’s research is in the NAME THACKER Society co-sponsor the award, CS George Michael Memorial general area of parallel and high AWARD RECIPIENT which was initiated in 1979. Fellowships honor exceptional performance computing, with The late Charles P. “Chuck” It recognizes contributions to Ph.D. students throughout the a special focus on developing Thacker has been named computer and digital systems world whose research focus algorithms for sparse tensor recipient of the ACM–IEEE architecture and comes with areas are in high performance factorization, which facilitates the CS Eckert-Mauchly Award a $5,000 prize. The award was computing, networking, storage, analysis of unstructured and high for fundamental networking named for John Presper Eckert and large-scale data analysis. dimensional data. Smith has made and distributed computing and John William Mauchly, ACM, the IEEE Computer Society, several fundamental contributions contributions including Ethernet, who collaborated on the and the SC Conference support that already have advanced the the , and development design and construction of the this award. state of the art on sparse tensor of the first tablet computers. Electronic Numerical Integrator Shaden Smith of the factorization algorithms. Often hailed as an “engineer’s and Computer (ENIAC), the University of Minnesota, and You’s research interests engineer,” Thacker made pioneering large-scale electronic Yang You of the University include scalable algorithms, fundamental contributions computing machine, which was of California, Berkeley, were parallel computing, distributed across the full breadth of completed in 1947. chosen to receive 2017 ACM/ systems, and machine computer development, from IEEE-CS George Michael learning. You has made several analog circuit and power supply TWO RECEIVE GEORGE Memorial HPC Fellowships. fundamental contributions that design to logic design, processor MICHAEL MEMORIAL Smith is being recognized reduce the communications and network architecture, HPC FELLOWSHIPS for his work on efficient and between levels of a memory system software, languages, and ACM, IEEE, and the SC Conference parallel large-scale sparse hierarchy or between processors applications. He passed away have announced the 2017 tensor factorization for machine over a network. June 12 at the age of 74, after a recipients of the ACM/IEEE learning applications. You is The Fellowships include brief illness. George Michael Memorial HPC being recognized for his work a $5,000 honorarium and The ACM–IEEE CS Eckert- Fellowships. on designing accurate, fast, travel expenses to attend SC17 Mauchly Award is known as Endowed in memory of and scalable machine learning in Denver, CO, Nov. 12–17, the computer architecture George Michael, one of the algorithms on distributed where the Fellowships will be community’s most prestigious founding fathers of the SC systems. formally presented.

14 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 news

Society | DOI:10.1145/3137764 Alex Wright Censoring Sensors Amid growing outcry over controversial online videos, tech firms grapple with how best to police online advertising.

OLLOWING THE WAVE of U.K. terror attacks in the spring of 2017, prime minister Theresa May called on technology companies like Facebook and FYouTube to create better tools for screening out controversial content— especially digital video—that directly promotes terrorism. Meanwhile, in the U.S., major adver- tisers including AT&T, Verizon, and Wal- Mart have pulled ad campaigns from their own devices in terms of develop- treasonous in another. Navigating this YouTube after discovering their content ing ethics,” says Sarah T. Roberts, an shifting terrain of international laws, had been appearing in proximity to assistant professor in the Department regulations, and advertiser sensibili- videos espousing terrorism, anti-Semi- of Information Studies of the Universi- ties—while continuing to provide as tism, and other forms of hate speech. ty of California, Los Angeles. “The pub- open a forum as possible to grow their In response to these controversies, lic is starting to ask questions about audiences—presents companies like Google expanded its advertising rules the power these companies have.” Google and Facebook with a complex, to take a more aggressive stance While certain types of content multidimensional challenge. against hate speech, and released a clearly violate the law (child pornogra- In an attempt to give the public suite of tools allowing advertisers to phy, for example), far more material some visibility into its internal dia- block their ads from appearing on falls into a vast grey area ranging from logue around these questions, Face- certain sites. The company also de- the mildly insensitive or tasteless to book launched a series of blog posts ployed new teams of human monitors outright hate speech. Much of this called “Hard Questions” that pro- to review videos for objectionable content falls well within the free vides a window into the company’s content. In a similar vein, Facebook speech protections of the First Amend- current thinking. announced that it would add 3,000 ment in the U.S., yet remains unpalat- “How aggressively should social me- new employees to screen videos for able to advertisers who fear being as- dia companies monitor and remove inappropriate content. sociated with anything that could controversial posts and images from Yet these kinds of manual fixes won’t alienate potential consumers. their platforms?,” Facebook vice presi- solve the problem, especially as video be- In the absence of government- dent of Public Policy and Communica- comes more and more ubiquitous enforced laws and regulations, however, tions Eric Schrage wrote in an intro- thanks to smartphones, car cameras, what ethical obligations do these com- ductory post. “Who gets to decide and other embedded capture devices. panies have toward safeguarding the what’s controversial, especially in a Ultimately, monitoring user-generat- public digital commons that provides global community with a multitude of ed content at scale will demand both the foundation for their businesses? cultural norms?” computational and policy solutions. “It’s hard to argue with a straight The company currently employs a At the heart of this problem lies a con- face about freedom of speech online combination of image , lan- flict that is deeply embedded in the his- when that speech is commoditized and guage understanding, and human tory of the Internet itself: between the highly lucrative,” says Roberts. “This is monitoring to identify and remove network’s capacity for unleashing cre- a particular version of free speech that Facebook groups that promote terror- ative self-expression, and platform pro- is deeply influenced by the politics and ist activity around the world. viders’ business need to give advertisers’ ethos of .” “Although our use of AI against ter- control over their messages. Caught be- Laws and regulations around free- rorism is fairly recent, it’s already tween these opposing forces, compa- dom of expression also vary widely from changing the ways we keep potential nies like Google and Facebook make country to country. For example, most terrorist propaganda and accounts off judgment calls about what kinds of con- European Union countries place much Facebook,” wrote Facebook’s head of tent to deem acceptable—decisions that tighter restrictions around hate speech Global Policy Management Monika are often obscured from public view. than the U.S. does, and a political video Bickert in a recent blog post. “These platforms and the firms that commentary that’s considered satirical Policy matters aside, the sheer vast- control them have largely been left to in one country might be considered ness of the Internet—with over one

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 15 news

billion videos on YouTube (and more other video-capable devices through- than 400 hours of new content being out our homes and offices. uploaded every hour)—introduces fur- Even if it were Whereas today, machine learning ther complexities in terms of both possible to identify happens primarily over the network— process and technology. with supercomputers in datacenters Many researchers think the long-term visual elements analyzing big datasets stored in the solution to monitoring online content across all kinds of cloud—eventually some of those pro- will inevitably involve artificial intelli- cesses will migrate toward edge-layer gence. Recent advances in big data, ma- video files with devices. Over time, the task of identi- chine learning, and embedded Graphics 100% accuracy, fying objectionable content may be- Processing Units (GPUs) are starting to come more diffuse, as computer vi- pave the way for more scalable approach- that alone wouldn’t sion algorithms increasingly come es to computer vision, allowing neural solve the problem pre-coded on chips embedded on networks to identify emerging patterns these devices. “The more algorithms in user-generated content that may de- of screening for move to the source of data capture, mand further human scrutiny. objectionable the more challenging it will be to cope “With machine learning, we can with real-world factors,” says Awad. now understand a lot more about material. Ultimately, the future of monitoring what’s going in a video or an image,” digital content may have less to do with says Reza Zadeh, an adjunct professor policy-making and brute-force pro- at and founder and cessing at the platform provider level, CEO of Matroid, a Palo Alto, CA-based and more to do with algorithmic filters computer vision software start-up that director of TRECVID, a U.S. National In- making their way into the devices all is developing tools for video analysis. stitute of Standards and Technology around us—for better or worse. Built atop TensorFlow (Google’s (NIST)-sponsored project that evaluates As Zadeh puts it: “Computers are open-source library for machine intelli- video search engines and explores new opening their eyes.” gence), Matroid uses a video player cou- approaches to content-based video re- pled with a so-called detector program trieval. “The major challenges occur Further Reading to identify similar images in a given vid- when dealing with user videos in the eo stream. For example, a Matroid de- wild that are not professionally edited.” Abu-El-Haija, S., Kothari, N., Lee, J., tector could look for images of Donald Even if it were possible to identify Natsev, P., Toderici, G., Varadarajan, B., and Vijayanarasimhan, S. Trump across five hours of video—like a visual elements across all kinds of video YouTube-8M: A Large-Scale Video few weeks’ worth of network news files with 100% accuracy, that alone Classification Benchmark. arxiv.org/ broadcasts, or large volumes of You- wouldn’t solve the problem of screening abs/1609.08675 Tube videos—and pinpoint the spots for potentially objectionable material. Bickert, M. where those images appear. It can also Much of the “content” of a video involves Hard Questions: How we Combat Terrorism, easily detect images containing gore or spoken words, after all, or other contex- Facebook blog post, violence, nudity and other forms of tual cues that won’t be readily apparent https://newsroom.fb.com/news/2017/06/ how-we-counter-terrorism/ NSFW (not safe for work) content, and from simply identifying an image. It’s look for “more like this” elsewhere in notoriously difficult for computers to Hegde, V., Zadeh, R. FusionNet: 3D Object Classification other video streams. distinguish news from satire, for exam- Using Multiple Data Representations, The company offers a self-service ple—or an editorial opinion piece about 3D Deep Learning Workshop at tool for non-technical users to train the terrorism from a call to arms. NIPS 2016. Barcelona, Spain, 2016. system to spot particular images, as well In order to automate the process of http://3ddl.cs.princeton.edu/2016/papers/ as a more advanced version geared to- screening video content at scale, re- Hegde_Zadeh.pdf ward machine learning engineers that searchers will likely need to apply natu- Real, S., Shlens, J., Mazzocchi, S., enables them to edit the neural network ral language search techniques to be- Pan, X., and Vanhoucke, V. YouTube-Bounding Boxes: A Large architecture, explore histograms, and gin parsing videos for deeper levels of High-Precision Human-Annotated Data Set ultimately create their own detectors for nuance. “The gap between what the for Object Detection in Video. (preprint) others to use. videos demonstrate and what an auto- Accepted at the Conference on Computer While deep learning approaches are mated system would generate for a nat- Vision and Pattern Recognition (CVPR) 2017. yielding advances in analyzing videos ural language description is still very arXiv:1702.00824. and other image-based content, the big and challenging,” says Awad. Venugopalan, S., Xu, H., Donahue, J., wide variety in the type and quality of Looking ahead, Zadeh also sees Rohrbach, M., Mooney, R., and Saenko, K. Translating Videos to Natural Language video across different capture devices plenty of opportunity on the hardware Using Deep Recurrent Neural Networks. poses additional obstacles. front, with semiconductor makers de- arXiv:1412.4729 [cs.CV] “Applying machine learning tech- vising computer vision-capable chips

niques to analyzing video content works that can work on devices like next-gen- Alex Wright is a writer and researcher based in Brooklyn, NY. reasonably well when the right condi- eration smartphones and cameras, tions exist,” says George Awad, project self-driving cars, and a wide range of © 2017 ACM 0001-0782/17/11 $15.00

16 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 news

Technology | DOI:10.1145/3137838 Esther Shein Overcoming Disabilities Brain-computer interfaces hold the promise of fully featured replacements for body parts that don’t work or are missing.

N THE MOVIE Star Wars: The Empire Strikes Back, Luke Skywalker is given a mechanical hand that moves and perform functions as well as his real hand. Konrad IKording, an avid Star Wars fan, has no doubt that advances in brain-machine interfaces (BMIs) will make this bit of science fiction a reality; he just doesn’t know when. “We have applications for one chan- nel and a few channels,” says Kording, a neuroscientist and professor of Physical Medicine and Rehabilitation, Physiolo- gy, and Biomedical Engineering at Northwestern University in Evanston, IL. “The question is, what are the BMI applications with hundreds of thou- sands of channels, and no one knows that at the moment.” The channels he’s referring to are electrical wires or opti- cal connectors that can be attached to the brain and can be controlled and measured. Michel Fornasier, one of the presenters of the Cybathlon, uses his bionic hand prosthesis to “The blind will see, and the amputat- demonstrate one of the Cybathlon disciplines. ed will move with limb replacements as well as you and me,’’ Kording predicts. eral years to study and facilitate BCI ap- channel of input from the user to the A BMI is also sometimes referred to plications in the public, private, and ed- computer, in addition to the keyboard, as a brain-computer interface (BCI), a ucational sectors. mouse, etc.,’’ Jacob says. mind-machine interface (MMI), or Most brain-computer interface re- The research organization Brain- direct neural interface (DNI). The bot- search focuses on reading signals from Gate (http://www.braingate.org) has tom line is such systems facilitate a di- the brain to the computer, usually EEG been working to develop and test BCI rect communication pathway between (electroencephalogram) or fNIRS (func- devices to restore the communication, an enhanced or wired brain and an ex- tional near-infrared spectroscopy), mobility, and overall independence of ternal device. notes Robert Jacob, a computer science people with a spinal cord injury, brain- BMIs are nothing new. Kording professor at Tufts University, Medford, stem stroke, or Amyotrophic lateral points to one of the earliest simple MA. EEGs can be invasive, as when a pa- sclerosis (ALS). The Intelligence Ad- brain machines, the cochlear implant, tient’s head is cut open to insert elec- vanced Research Projects Activity (IAR- which was invented in 1961. Deep brain trodes, whereas fNIRS is always non-in- PA), part of the U.S. Office of the Direc- stimulation, a surgical procedure to im- vasive, he says. “There are other ways of tor of National Intelligence, sponsors plant a hair-thin electrode wire in the reading information from the brain to research programs using multidisci- part of the brain responsible for abnor- the computer, like fMRI or MEG [mag- plinary approaches “to advance under- mal movement, is being used in treat- netoencephalography], but these are standing of cognition and computa- ing diseases like Parkinson’s and epi- generally too clumsy for reasonable BCI tion in the brain.” lepsy. “What they all have in common,’’ applications,’’ Jacob says. In 2013, then-President Barak Obama he says, “is that [they are] low-dimen- Most of this work used to be targeted announced $100 million in funding for sional brain machine interfaces,’’ at helping physically disabled people, the Brain Research through Advancing meaning there are not many wires or but now, it is “spreading into more Innovative (BRAIN) channels going into the brain. mainstream applications, where the initiative to help neuroscientists under-

PHOTO BY BSIP/UIG VIA GETTY BY IMAGES PHOTO Research has been under way for sev- brain information provides an extra stand the origins of cognition, percep-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 17 news

tion, and other brain activities. The counted for over 50% of the market in ing a balance between the and thought was the research could lead to 2014, followed by communication and accuracy of both hardware and software new and more effective treatments for control, entertainment & gaming, and before they are likely to be adopted by conditions like autism and mood disor- smart home control, according to the mainstream users. “For instance, heart ders, as well as people suffering from Grand View report. rate monitors, another type of biosig- brain injuries. As part of its mission, the Portland, OR-based Allied Market Re- nals system, could only become more BRAIN initiative will develop and deploy search is forecasting the BCI market will widespread with the existence of practi- neurotechnologies to further under- reach $1.46 billion by 2020, with a CAGR cal and decently accurate wearables like standing of the links between brain of 11.5% over that period. chestbands and watches, alongside the function and behavior. The key factors driving this sector’s development of a number of fitness This prompted federal agencies like growth are increasing focus on integrat- apps,’’ she says. the Defense Advanced Research Projects ing the technology into various health- Already, systems allowing for gestur- Agency (DARPA) to get involved, with the care and military applications, as well as ing and voice input have created new formation of the on developing communication technol- and intuitive ways for people to interact System Design (NESD) program to en- ogy for the disabled and geriatric popu- with computers without the need for the hance research capabilities in neuro- lations, according to market research more traditional combination of key- technology and provide an interface for firm Future Market Insights, Valley Cot- boards and mice, Liu notes. new therapies. tage, NY. The firm anticipates there will “Brain-computer systems push this a So it was only a matter of time before also be a growing focus on utilizing BCI step further, with the goal of more di- BCI began receiving greater commercial to control Internet of Things (IoT) devic- rectly using the brain to convey our in- attention from the likes of billionaire en- es, smart home applications, and vari- tentions, rather than having an extra, trepreneurs like Elon Musk and Bryan ous virtual reality applications. physical step translating those inten- Johnson. Musk, CEO of Tesla and Fannie Liu, a Ph.D. student in the tions to text, speech, or gestures.” This, SpaceX, is backing , a BCI ven- Human-Computer Interaction Institute she says, could make our interactions ture to create devices that can be im- at Carnegie Mellon University, agrees easier, faster, and ideally more natural. planted in human brains to eventually with the research, but stresses that use Not only can these systems support improve memory and interface with of BCIs right now is fairly limited. human-computer interaction, but po- computer systems. Johnson, who found- “Industries have been focusing on tentially human-human interaction as ed online payments company Braintree gaming and medical applications, par- well, Liu says. Her work is investigating (which was acquired by eBay in 2013 for ticularly for disabled or paralyzed indi- whether “we could potentially use BCI to $800 million), has invested $100 million viduals,” she says, “though research in better understand and communicate into startup venture Kernel, whose goal this field has shown potential in a variety with each other,” by clarifying our un- is to enhance human intelligence with of areas, such as communication, thera- derlying thoughts and feelings. brain implants that can link people’s py, and controlling computer tools and In addition to Neuralink, a slew of thoughts with computers. devices.” other startups are also hoping to capi- “Once one or two billionaires do Liu, who is working closely with her talize on the BCI market. Among them something, everyone wants to be in- advisor, Geoff Kaufman, an assistant is NeuroPace, a medical device compa- volved,’’ observes Kording. “If I was a professor at the Institute, adds that ny aimed at improving the lives of pa- successful business person with the while several applications are possible, tients suffering from epilepsy, which chance to build something really, really BCI systems still face challenges in strik- the company says affects approximately cool and be part of something that’s re- 1% of the population worldwide. Neuro- ally, really exciting, what’s more interest- Pace has developed a medical device ing than the brain?” BCI systems called the RNS System, which was ap- That interest has spiked growth in proved by the FDA in November 2013 the global BCI market, which was valued face challenges and which uses responsive neurostimu- at $723.64 million in 2014 and is expect- in balancing the lation (RNS) technology to monitor and ed to grow at a compound annual growth control seizures in patients with drug- rate (CAGR) of over 10% through 2022, usability and resistant epilepsy, according to CEO according to San Francisco-based Grand accuracy of hardware Frank Fischer. RNS is an approach to View Research. treating medically uncontrolled partial “Rising occurrence of target disor- and software before onset seizures, and works by continu- ders and neurodegenerative conditions they are likely ously monitoring brain activity, detect- are expected to propel market growth ing abnormal patterns, and in response, over the forecast period,” the firm notes. to be adopted by delivering imperceptible electrical puls- “R&D efforts have led to pioneering in mainstream users. es to normalize the activity before an in- the engineering of headset develop- dividual experiences a seizure. ment, which provides treatment for im- The RNS System can be used on indi- paired cognitive function.” viduals 18 and older to monitor and re- Healthcare applications make up by spond to brain activity in real time, pre- far the largest vertical segment, and ac- venting epileptic seizures at their source,

18 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 news

like the devices being developed by Musk and others, these are the higher- ACM Liu is investigating dimensional, more complex BCI sys- whether “we could tems, he says. Member In addition to heat concerns, if the potentially use BCI devices are not encased properly, they to better understand could potentially corrode and damage News the brain, Kording adds. AT THE INTERSECTION and communicate “So there are lots of problems, but OF PROGRAMMING with each other” lots of opportunities,’’ and neurotech- LANGUAGES AND nology is helping keep technology so in- MACHINE LEARNING by clarifying our “My goal is teresting, he says. “In some way, neuro- to make underlying thoughts technology is the final frontier, because programmers if we understand how the brain works, more and feelings. productive, and we can understand a lot of the things enable them to that ultimately affect us and make us be write programs humans.” It is also probably the hardest more easily,” says Ganesan area to figure out, Kording says. Ramalingam, a principal researcher for Microsoft Fischer says. The system costs $37,000, “Understanding any other part of sci- Research India. Ramalingam’s which includes the implantable device, ence is probably easier than under- expertise is in static program leads, and remote monitor. standing the brain and arguably, less analysis, or the identification Both Kording and Liu caution that it exciting.” of bugs and errors in code without having to execute will take years before BCIs become more the program being analyzed. widespread, with Liu observing that ex- “When the programs don’t Further Reading isting systems require a lot of setup. “Us- work as they are expected to, you want to know what they are ers need to be equipped either invasively Jankowska, E. Spinal control of motor outputs by intrinsic doing,” he explains. through surgical implantation or nonin- and externally induced electric field Ramalingam earned his vasively with several sensors or hardware potentials, Journal of Neurophysiology. 24 undergraduate degree in that can take a while to calibrate.” As a May 2017. DOI: 10.1152/jn.00169.2017 computer science from the http://bit.ly/2sk6FIY Indian Institute of Technology, result, right now, both options “are not in Madras, and earned his practical for everyday use,” she says. Stevenson, I.H. and Kording, K.P. Ph.D. in the discipline in Although more usable consumer- How Advances in neural recording affect 1993 from the University of grade devices have come out in the last data analysis. Nature . 26 Wisconsin–Madison. He then January 2011. took a position at IBM’s T.J. decade, they tend to be less accurate, she http://go.nature.com/2rFboTQ Watson Research Center in observes. “Future iterations of BCI New York, where he stayed until would likely continue to be noninvasive, Marblestone, A.H., Zamft, B.M., Maguire, Y.G., 2006, when he joined Microsoft Shapiro, M.G., Cybulski, T.R., Glaser, J.I., Research India. but with new methods that can enable Amodei, D., Stranges, P.B., Kalhor, R., Currently, Ramalingam the recording and processing of stron- Dalrymple, D.A., Seo, D., Alon, E., Maharbiz, is primarily interested in ger and better signals from the brain.” M.M., Carmena, J.M., Rabaey, J.M., Boyden, E.S., programming languages and Major hurdles like privacy concerns Church, G.M., and Kording, K.P. their interaction with machine Physical principles for scalable neural learning, which he thinks has also need to be addressed. Liu says her recording. NCBI. 21 Oct., 2013. become more important thanks research has found people are reluctant http://bit.ly/2tbf1CQ to rapid advances in this area to allow systems to tap into their brains over the past decade. “We Bansal, A.K., Truccolo, W., Vargas-Irwin, C.E., are still in the early stages of and understand their intentions, as they Donoghue, J.P. machine learning, and it is quite may feel their inner thoughts can be Decoding 3D reach and grasp from hybrid challenging to build applications read and revealed to anyone with access signals in motor and premotor cortices: in this area,” he adds. to the system. “BCI systems need to pro- spikes, multiunit activity, and local field Ramalingam feels machine potentials. Journal of Neurophysiology. 1 learning is going to be an tect users’ privacy if they are ever to be- March, 2012, Vol. 107, no. 5. increasingly important field, come prevalent,” she says. http://bit.ly/2tw7uy3 with a much wider spectrum How to keep implants in the brain, of programmers wanting Pohlmeyer, E.A., Oby, E.R., Perreault, E.J., Solla, to incorporate it into their particularly if a device involves wiring, S.A., Kilgore, K.L., Kirsch, R.F., and Miller, L.E. programs. “We need to also needs to be determined because the Toward the Restoration of Hand Use to a make it easier for non-expert brain can reject them, adds Kording. “At Paralyzed Monkey: Brain-Controlled programmers to incorporate the moment, we don’t have ways of inter- Functional Electrical Stimulation of Forearm machine learning into their programs, and that these facing with the brain in ways where de- Muscles. PLOS Journals. 15 June, 2009. http://bit.ly/2sa32XJ machine learning programs vices will last a lifetime.” interact correctly with the rest of Scientists and researchers also need the programs these non-expert Esther Shein is a freelance technology and business to deal with issues involving the heat programmers write,” he says. “I writer based in the Boston area. think there be lots of interesting BCI devices give off, and how to analyze challenges here.” and control the data they produce. Un- © 2017 ACM 0001-0782/17/11 $15.00 —John Delaney

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 19 viewpoints

DOI:10.1145/3144170 Pamela Samuelson VLegally Speaking Disgorging Profits in Design Patent Cases Does the recent U.S. Supreme Court decision in the Apple v. Samsung case represent a quagmire?

S REPORTED IN my March On remand, Apple is still insisting Samsung’s victory notwithstanding, 2017 column, the U.S. Su- that the relevant article is the whole it is premature to assume the risk of preme Court struck down smartphone. (For the sake of brevity, excessive awards in design patent cas- a $399 million award I will use the term “article” instead of es has subsided. The Supreme Court against Samsung for in- repeating “article of manufacture” or did not rule that the relevant article Afringing three Apple design patents. adopting AOM as an acronym.) Sam- would necessarily and always be an Samsung’s win concerned an impor- sung, however, contends the relevant individual component of a multicom- tant but narrow issue. The Court ruled articles are the relatively small compo- ponent device, only that it might be a that Apple is entitled to be awarded nents that embody the three designs component. How the relevant article Samsung’s profits from sale of the at issue (that is, a rectangular flat face and profits-from-that-article issues are article(s) of manufacture to which the with rounded corners, a rectangu- resolved in Apple v. Samsung will have protected designs were applied. How- lar flat face with bezel, and a colorful significant implications for future de- ever, lower courts erred in ruling that screen with 16 icons). Apple will be sign patent cases involving multicom- the relevant article of manufacture was entitled to a much more modest award ponent devices. necessarily the whole smartphone; it than $399 million if Samsung prevails could instead be one or more compo- on this issue. Recap on Design nents of the smartphones. Samsung was not the only tech- Patent Disgorgement The Apple v. Samsung case has been nology company relieved by the Su- Disgorgement of infringer profits as sent back to the trial court to deter- preme Court’s ruling. Facebook, a remedy in design patent cases came mine, first, to what relevant article(s) Google, eBay, Hewlett-Packard, and about because Congress got upset of manufacture were the patented Dell, along with several high-technol- about a set of cases decided the 1880s. designs applied, and second, what ogy industry associations, filed am- It reacted to Supreme Court rulings part of the $399 million total profits icus curiae briefs in support of Sam- that the owner of a patent for a car- Samsung made from sales of the in- sung’s appeal. The briefs argued that pet design was entitled to only nomi- fringing smartphones is attributable when design patent infringement oc- nal damages (at the time, six cents, to the relevant article(s) of manufac- curs as to only one or a small number now worth about $1.50) against two ture. The Court offered no guidance of components, it would be improper infringers. The Court rejected a plea about how lower courts should make to disgorge total profits from the sale for disgorgement, saying that the pat- either assessment. of multicomponent devices. entee had failed to prove what portion

20 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 V viewpoints

of the profits were attributable to the infringement more than an appor- erators because they like the design infringing designs. tionment remedy would. of its interior light. However, if the Soon thereafter, Congress effective- light and switch mechanism is sold ly overturned the Court’s ruling by cre- What’s the Relevant Article? as a component to be incorporated ating a new remedy for design patent In many, and perhaps most, design into other refrigerator models or oth- infringement. Under it, design paten- patent cases, the article embodying a er products with doors, it should be tees could elect, as an alternative to an patented design will be the product as possible to estimate what profits to award of actual damages (for example, a whole. The design of a carpet, wallpa- disgorge against an infringer of this lost profits or a reasonable royalty), a per, or chair, for instance, will often be design. It would be patently unfair, disgorgement of the defendant’s “to- the main selling point for the product, though, to disgorge profits from in- tal profits” on the sale of any articles of even though the quality of the materi- fringer’s sales of refrigerators when manufacture embodying the protected als used in the manufacturing process, only the interior light embodies the design, or at least $250 (approximately the skill of the craftsmen, and other patented design. equivalent to $6,100 today). characteristics of the product or pro- Identifying the relevant article for The legislative history made clear ducer may contribute to the value of disgorgement purposes is more diffi- that courts should not try to deter- the product. If, however, the design is cult, however, in cases involving com- mine how much of the defendant’s what actually drives demand for the plex multicomponent products, such profits were attributable to the in- product, it seems fair that infringers as smartphones, when there is no sep- fringing design, as opposed to the should have to disgorge all profits. arate market for components that may non-aesthetic attributes of the arti- A patented design may, of course, embody a patented design. What evi- cle. Instead, they should award total also be embodied in a component dence or factors should a judge or jury profits from the sale of the article of of a multicomponent product. Prof- consider in determining the relevant manufacture embodying the design. its disgorgement should be relative- article in such cases? Congress recognized that this might ly straightforward when there is a The place to start a relevant ar- overcompensate some design paten- separate market for the component. ticle inquiry should be looking at the tees, but that was better than under- Consider, for instance, a patented design patent, which is supposed to compensating them. Moreover, the design for an interior light that turns identify the article intended to embody potential availability of a total profits on when a refrigerator door opens. the design. The patent must include

IMAGE BY ALICIA KUBISTA/ANDRIJ BORYS ASSOCIATES BORYS ALICIA KUBISTA/ANDRIJ BY IMAGE award would also deter design patent Consumers are unlikely to buy refrig- a drawing of the patented design as it

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 21 viewpoints

would appear in the article. The patent uct and service features or attributes. may also have a textual description of By asking many people to express a the intended article. In addition, one One might consider preference between a few dozen pairs should examine the patentee’s and al- how much total profit of designs that differ on features and leged infringer’s products to discern attributes, product marketers can use how the patented design was embod- Samsung would have clever mathematical techniques to es- ied in the litigants’ products. These made from sales timate the importance of each feature types of evidence may sometimes suf- in isolation. This technique has been fice to identify the relevant article as to of smartphones used in some patent infringement cas- which disgorgement should occur. if it had not infringed es, and would seem well suited for reso- In her argument before the Su- lution of cases such as Apple v. Samsung preme Court, Samsung’s lawyer rec- Apple’s design. in which design patents may cover only ommended these steps for the relevant one or a small number of components article inquiry. She also suggested that of multicomponent products. market studies might be useful to un- Alternatively, a profits disgorge- derstand what consumers perceive the er reports about Apple smartphones, ment assessment might be built on a article embodying the design to be. which typically discuss features that counterfactual scenario. As applied in Another indicator might be the costs consumers find most desirable. Apple v. Samsung, one might consider incurred in developing the component how much total profit Samsung would embodying the design. What Profits to Disgorge? have made from sales of smartphones The Solicitor General of the United Once the relevant article inquiry is re- if it had not infringed Apple’s design States, in a brief supporting Samsung’s solved, the next question is what “total patents and compare this estimate to appeal and during oral argument, pro- profits” did the infringer make from the total profit Samsung actually made posed consideration of four factors in sale of that article. With a multicompo- on infringing smartphones. The differ- making the relevant article determina- nent product whose components are ence between the two profit scenarios tion: first, the scope of the patented de- sold only as a package, “total profits” would be the amount that Samsung sign; second, the prominence of the de- on one or more component-article(s) should have to pay Apple under this sign in the challenged product; third, will almost certainly be some percent- model of the disgorgement remedy. whether the design is conceptually age of the profits made from sales of This approach contrasts with a more distinct from the product as a whole; the infringing products. In Apple v. compensatory approach that would and fourth, the physical relationship Samsung, $399 million was determined focus on how much total profit Apple between the patented design and the as the total profit Samsung made sell- made from sales of its smartphones rest of the product. The Supreme Court ing the infringing smartphones. So if and how much profit it would have did not endorse use of these factors. the relevant article is not the smart- made if Samsung had not infringed. In support of its claim that the phone as a whole, how should one de- Economic experts could create models smartphone as a whole should still be cide what portion of those profits are for undertaking these assessments. considered the relevant article, Apple attributable to the components held to can be expected to argue that the pat- be the relevant article(s)? Conclusion ented designs are inseparable from the Expert witnesses are likely to play Samsung won an important victory products embodying them, and that a significant role in assessing atotal- for itself and for other high technol- consumer demand for Apple products profits-attributable-to-the-relevant- ogy companies in challenging the total is due to its well-integrated designs. article award in design patent cases. profits award in the Apple case. Howev- Would Samsung have sold so many Experts hired by the patentee and by er, uncertainty exists about how courts millions of smartphones if it had not the alleged infringer are, of course, un- or juries should go about determining misappropriated the cool look of Apple likely to agree on the bottom line. How- the relevant embodiment of a patented designs? Apple thinks not. ever, their assessments, as set forth in design to serve as the “article” on which Samsung will argue that the flat face reports and testimony, will generally the infringer’s “total profits” should be of the smartphone and the flat face with set the bounds within which the trier disgorged. Because high-tech compa- bezel are two minor components of the of fact, whether a judge or a jury, will nies are utilizing design patents much exterior design of its smartphones. make the award. Juries, in particular, more now than in the past, they have The colorful 16-icon design is similar- are likely to consider the relative culpa- reason to worry about the legal quag- ly one small component of the screen bility of the infringer in making such mire hovering over disgorgement of displays of which smartphones are ca- awards. It is consistent with principles profits awards in design patent cases pable. These should be the three “ar- of unjust enrichment for them to do so. involving multicomponent products. ticles” to which the Apple designs have Conjoint analysis may be a useful been applied. Samsung will point to economic technique to contribute to Pamela Samuelson ([email protected]) is the Richard M. Sherman Distinguished Professor of Law and the very large number of components a design patent profits disgorgement Information at the University of California, Berkeley, and a in smartphones to put the infringing analysis. It is often used to analyze how member of the ACM Council. components into proper perspective. consumers conceive, integrate, value, Samsung could also point to consum- and trade off different clusters of prod- Copyright held by author.

22 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints

VDOI:10.1145/3144172 Solon Barocas and danah boyd Computing Ethics Engaging the Ethics of Data Science in Practice Seeking more common ground between data scientists and their critics.

RITICAL COMMENTARY ON data science has converged on a worrisome idea: that data scientists do not recog- nize their power and, thus, Cwield it carelessly. These criticisms channel legitimate concerns about data science into doubts about the eth- ical awareness of its practitioners. For these critics, carelessness and indiffer- ence explains much of the problem— to which only they can offer a solution. Such a critique is not new. In the 1990s, Science and Technology Studies (STS) scholars challenged efforts by AI researchers to replicate human behav- iors and organizational functions in software (for example, Collins3). The scholarship from the time was damn- ing: expert systems routinely failed, critical researchers argued, because developers had impoverished under- standings of the social worlds into which they intended to introduce their understandings of the relevant tech- trying to make machines learn some- tools.6 At the end of the decade, how- niques, they also expressed frustration thing useful, valuable, and reliable. ever, Mark Ackerman reframed this as that critics failed to account for the For example, dealing with dirty and in- a social-technical gap between “what careful thinking and critical reflection complete data is as much a moral as a we know we must support socially and that data scientists already do as part practical concern. It requires making a what we can support technically.”1 He of their everyday work. This was more series of small decisions that are often argued that AI’s deficiencies did not than resentment at being subject to fraught, forcing reflection at each step. reflect a lack of care on the part of re- outside judgment by non-experts. In- How was this data collected? Does it searchers, but a profound challenge of stead, these data scientists felt that capture the entire population and full dealing with the full complexity of the easy criticisms overlooked the kinds of range of behavior that is of interest? social world. Yet here we are again. routine deliberative activities that out- The same is true for validating a model Our interviews with data scientists siders seem to have in mind when they and settling on an acceptable error rate. give us reason to think we can avoid talk about ethics. What must a data scientist do to prove this repetition. While practitioners to herself that a model will indeed per- were quick to point out that common Ethics in Practice form well when deployed? How do data criticisms of data science tend to lack Data scientists engage in countless acts scientists decide that a reported error

IMAGE BY LIU ZISHAN BY IMAGE technical specificity or rest on faulty of implicit ethical deliberation while rate is tolerable—and defendable? Eth-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 23 viewpoints

ical considerations also emerge while dealt with independently or on top of making more fundamental decisions technical practice—without noticing regarding the choice of learning algo- that ethical deliberation is embedded rithm, where practitioners frequent- in the everyday work of data scientists. ly struggle to find an approach that Even when attempting to address maximizes the resulting models’ per- ethical issues more explicitly, practi- formance while also providing some tioners face difficult trade-offs. One degree of interpretability. When is the interviewee described a dilemma in ability to meaningfully interrogate a choosing whether or not to “know” the model sufficiently important to justify gender of the individuals in his mod- some cost in performance? What kinds el—with that information, he could of decisions—and real-world effects— check whether his model might exhibit Call for drive data scientists to develop a model some kind of gender bias; without it, he that they can explain, even if its deci- could claim that this sensitive attribute Nominations sions might be less accurate as a result? did not figure into the model. Other These are difficult decisions, for researchers who are concerned about for ACM which data scientists must employ care- gender biases in data have attempted to General Election fully cultivated judgment. Yet, many build technical interventions to address data scientists do not use the language them,5 but such an approach requires of ethics to talk about these practices. trading off privacy in order to construct They may speak of trade-offs, but they a viable fairness remedy, a decision that The ACM Nominating primarily talk about what it takes to presents its own challenges.4 be good at what they do. Pressed about Committee is preparing “ethics” directly, many data scientists Where Ethics Is Not Enough to nominate candidates say “this is not my area,” even though Critics are right to emphasize the se- for the officers of ACM: they draw on a wide range of values to riousness of the implications of data President, work through difficult tensions. science. And, as Cathy O’Neil has Broad critiques of data science prac- pointed out in The Weapons of Math Vice-President, tices cannot account for the diversity of Destruction,8 data science is being Secretary/Treasurer; practices, concerns, or efforts among deployed by powerful organizations and two data scientists. Instead, they often pre- to achieve goals that can magnify in- sume ignorance or corrupt intentions. equality and undermine democratic Members at Large. All too often, the data scientists we have decision-making. She calls on data encountered are quite sympathetic scientists to recognize how they are Suggestions for candidates to the sentiment behind the critiques being used—and to push back against are solicited. Names should be they hear, but feel maligned and mis- misuse of their skills. sent by November 5, 2017 understood, unacknowledged for their Unfortunately, certain problems to the Nominating Committee Chair, efforts, and frustrated by vague recom- may stem from genuine value con- c/o Pat Ryan, mendations that are not actionable. flicts, not simply a lack of attention to Chief Operating Officer, Outsiders’ use of the term “ethics” sug- the values at stake. Over the past year, ACM, 2 Penn Plaza, Suite 701, gests that normative concerns must be a debate has unfolded over the use of New York, NY 10121-0701, USA. data science in criminal justice, where courts rely on risk scores to make de- With each recommendation, please include background Broad critiques cisions about who should be released information and names of individuals from prison while awaiting trial. The the Nominating Committee of data science stakes are high: those given bail are can contact for additional practices cannot more likely to keep their jobs, house, information if necessary. children, and spouse; those who are account for the not are more likely to plead guilty, even Alexander L. Wolf is the Chair when they are innocent. of the Nominating Committee, diversity of practices, A group of data scientists work- and the members are concerns, or ing with ProPublica established that Karin Breitman, Judith Gal-Ezer, black defendants in Broward County, Rashmi Mohan, and Satoshi Matsuoka. efforts among data FL, who did not reoffend were twice scientists. as likely to be mislabeled as posing a high risk of recidivism than white de- fendants.2 They argued that the system exhibited a clear racial bias because er- rors imposed a far greater cost on black defendants, who were more likely to

24 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints be wrongly incarcerated, while white which such learning can occur. Many defendants were more likely to be set data scientists feel as though critics free but nevertheless recidivate. North- Technical actors only talk at them. When we asked one pointe (now Equivant), the company are often far more informant why he did not try to talk behind the risk assessment, countered back, he explained that social scien- that its tool was equally accurate in sophisticated tists and humanists were taught to predicting recidivism for black and than critics at debate and that he was not. Critics get white defendants. Since then, com- rewarded for speaking out publicly, he puter scientists and statisticians have understanding the said, garnering rewards for writing es- debated the different qualities that an limits of their analysis. says addressed to a general audience. intuitive sense of fairness might imply: This was not his skillset nor recognized that a risk score is equally accurate in as productive by his peers. predicting the likelihood of recidivism The gaps between data scientists for members of different racial groups; and critics are wide, but critique di- that members of different groups have vorced from practice only increases the same chance of being wrongly pre- standing the limits of their analysis. In them. Data scientists, as the ones clos- dicted to recidivate; or that failure to many ways, the work of data scientists est to the work, are often the best posi- predict recidivism happens at the same is a qualitative practice: they are called tioned to address ethical concerns, but rate across groups. While each of these upon to parse an amorphous problem, they often need help from those who expectations of a fair score might seem wrangle a messy collection of data, are willing to take time to understand like complementary requirements, re- and make it amenable to systematic what they are doing and the challenges cent work has established that satisfy- analysis. To do this work well, they of their practice. We must work col- ing all three at the same time would be must constantly struggle to under- lectively to make the deliberation that impossible in most situations; meet- stand the contours and the limitations is already a crucial part of data science ing two will mean failing to comply of both the data and their analysis. visible. Doing so will reveal far more with the third.4,7 Even if Northpointe Practitioners want their analysis to be common ground between data sci- had been more sensitive to disparities accurate and they are deeply troubled entists and their critics and provide a in the false positive and false negative by the limits of tests of validity, the meaningful foundation from which to rates, the appropriate way to handle problems with reproducibility, and the articulate shared values. such a situation may not have been ob- shortcomings of their methods. vious. Favoring certain fairness proper- Many data scientists are also deeply References 1. Ackerman, M.S. The intellectual challenge of CSCW: ties over others could just as well have disturbed by those who are coming The gap between social requirements and technical feasibility. Human-Computer Interaction 15, (2–3), reflected a difference in values, rather into the field without rigorous train- 2000, 179–203. than a failure to recognize the values ing and those who are playing into the 2. Angwin, J. et al. Machine bias. ProPublica. (May 23, 2016); https://www.propublica.org/article/machine- at stake. One thing is for certain: this hype by promising analyses that are bias-risk-assessments-in-criminal-sentencing use of data science has prompted a not technically or socially responsible. 3. Collins, H.M. Artificial Experts: Social Knowledge and Intelligent Machines. MIT Press, 1993. vigorous debate, making clear that our In this way, they should serve as allies 4. Corbett-Davies, S. et al. A computer program used normative commitments are not well with critics. Both see a need for nuanc- for bail and sentencing decisions was labeled biased against blacks. It’s actually not that clear. Washington articulated, that fuzzy values will be es within the field. Unfortunately, uni- Post (Oct. 17, 2016); https://www.washingtonpost. difficult to resolve computationally, versalizing critiques may undermine com/news/monkey-cage/wp/2016/10/17/can-an- algorithm-be-racist-our-analysis-is-more-cautious- and that existing ethical frameworks critics’ opportunities to work with data than-propublicas/ may not deliver clear answers to data scientists to meaningfully address 5. Feldman, M. et al. Certifying and removing disparate impact. In Proceedings of the 21st ACM SIGKDD science challenges. some of the most urgent problems. International Conference on Knowledge Discovery and Of course, even if data scientists Data Mining, (2015), 259–268. 6. Hess, D.J. Editor’s Introduction, Studying Those Who Toward a Constructive take care in their work and seek to Study Us: An Anthropologist in the World of Artificial Collaboration engage critics, they may not be well Intelligence. Stanford University Press, 2001. 7. Kleinberg, J., Mullainathan, S. and Raghavan, S. The critical writing on data science has prepared to consider the full range of Inherent Trade-Offs in the Fair Determination of Risk Scores. Arxiv.org. 2016; https://arxiv.org/ taken the paradoxical position of in- ethical issues that such work raises. In abs/1609.05807 sisting that normative issues pervade truth, few people are. Our research sug- 8. O’Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. all work with data while leaving unad- gests the informal networks that data Crown, 2016. dressed the issue of data scientists’ scientists rely on are fallible, incom- 9. Žliobaite', I. and Custers, B. Using sensitive personal data may be necessary for avoiding discrimination in ethical agency. Critics need to consid- plete, and insufficient, and that this data-driven decision models. Artificial Intelligence and er how data scientists learn to think is often frustrating for data scientists Law 24, 2 (Feb. 2016), 183–201. about and handle these trade-offs, themselves. Solon Barocas ([email protected]) is an Assistant while practicing data scientists need In order to bridge the socio-techni- Professor of Information Science at Cornell University. to be more forthcoming about all of cal gap that Ackerman warned about danah boyd ([email protected]) is a Principal the small choices that shape their deci- 20 years ago, data scientists and critics Researcher at Microsoft Research and the Founder/ sions and systems. need to learn to appreciate each oth- President of Data & Society. Technical actors are often far more er’s knowledge, practices, and limits. sophisticated than critics at under- Unfortunately, there are few places in Copyright held by authors.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 25 viewpoints

VDOI:10.1145/3144174 Richard Connor, Quintin Cutts, and Judy Robertson Education Keeping the Machinery in Computing Education Incorporating intellectual and developmental frameworks into a Scottish school curriculum.

E DO NOT think there can be “computer sci- ence” without a com- puter. Some efforts at deep thinking about Wcomputing education seem to sidestep the fact that there is technology at the core of this subject, and an important technology at that. Computer science practitioners are concerned with mak- ing and using these powerful, gener- al-purpose engines. To achieve this, computational thinking is essential, however, so is a deep understanding of machines and languages, and how these are used to create artifacts. In our opinion, efforts to make computer science entirely about “computational thinking” in the absence of “comput- ers” are mistaken. As academics, we were invited to help develop a new curriculum for computer science in Scottish schools covering ages 3–15. We proposed a single coherent discipline of com- puter science running from this early curriculum for the average five-year man worlds. Hence, we wanted a cur- start through to tertiary education old, as we have not taught them CS be- riculum that properly grounds their and beyond, similar to disciplines fore, but we are unconvinced of the co- understanding of that non-human such as mathematics. Pupils take herence of many current curricula: an world and gives every child the oppor- time to develop deep principles in underlying intellectual and develop- tunity, should they wish, of a future those disciplines, and with appro- mental framework seems to be miss- career in our area. Our school systems priate support the majority of pupils ing, and such a framework is our prin- have these aspirations in teaching make good progress. From our back- cipal offering to the curriculum. about the natural world—why not the ground in CS education research, we We understand both the desperate digital world also? saw an opportunity for all children to calls from industry to meet the labor In March 2017, the new curriculum learn valuable foundations in com- market demands of the digital econ- was formally adopted at government puting as well, no matter how far they omy, and the extraordinary environ- level, and its delivery has started. A progressed ultimately. ment that will be our children’s, with teachers’ guide is here—http://www.

Nobody knows exactly the right CS ever more blurring of digital and hu- teachcs.scot—and we encourage inter- OF OZOBOT COURTESY IMAGE

26 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints viewpoints

ested readers to look at the full guide. All curriculum design requires Calendar compromise. We have balanced: our All curriculum initial vision of a curriculum that cap- design requires of Events tures the essence of computation at the heart of the digital revolution; the compromise. November 1–3 practical realization that only a small IMC ‘17: Internet Measurement amount of resources are available for Conference, London, U.K. V teachers’ professional development; Co-Sponsored: ACM/SIG, the requirement to reuse a varied body Contact: Steve Uhlig, of existing early-years computing edu- A deep understanding of computer Email: [email protected] cational material; and the desire from science requires the following three November 6–10 government to direct computing edu- aspects, our three-point framework, CIKM ‘17: ACM Conference on cation down a narrow agenda to fill a which can be neatly separated as the Information and Knowledge perceived skills shortage. understanding of: Management, Singapore, Singapore, ˲˲ Nonetheless, we have kept in view Domains that can be modeled by Co-Sponsored: ACM/SIG, throughout our overarching frame- computational mechanisms; Contact: Marianne Winslett, work consisting of three main points ˲˲ The computational mechanisms Email: [email protected] that we think is the real contribution themselves; and November 7–10 ˲˲ of the curriculum, and the three points How to use the computational SIGSPATIAL’17: 25th ACM are the focus of this column. We will mechanisms to model aspects of the SIGSPATIAL International describe the essence of our proposed domains. Conference on Advances three-point underpinning, developing It is our belief that a computer sci- in Geographic Information Systems, three essential strands of learning, and entist is habitually and implicitly aware Los Angeles, CA, the way these have been eventually in- of these, and indeed is expert at quickly Contact: Erik Hoel, terpreted in the adopted curriculum. assimilating new instances of them. Email: [email protected] We believe this is a core skill with many November 8–10 Computational Foundations applications to a modern process- and MiG ‘17: Motion in Games, We aimed to identify a core framework information-driven world. Barcelona, Spain, for the discipline that is equally rel- Computational thinking, as well Sponsored: ACM/SIG, Contact: Nuria Pelechano, evant to a child, a university student, as the learning delivered via the Un- Email: [email protected] or a software engineer. The essence of plugged approach, are, we believe, computation is clear: the Church-Tur- largely captured within the first aspect. November 8–10 ing thesis. Some kind of computation- The skill of programming, as taught VRST ‘17: 23rd ACM Symposium on Virtual Reality Software and al mechanism—whether the Scratch even at university level, is mostly with- Technology programming environment, a Turing in the third. The second all-important Gothenburg, Sweden, Machine, or the Lambda Calculus— aspect seems to be often neglected, at Co-Sponsored: ACM/SIG, can be used to model any tail-recursive least until the later stages of a com- Contact: Morten Fjeld, Email: [email protected] numeric function … and therefore puting degree. It has long been a wry anything that can be computed … and observation of the authors that, while November 12–17 furthermore all such mechanisms are “programming” is taught right from SC17: The International somehow equivalent. the start of university computing cours- Conference for High Performance Computing, To be of interest, such mecha- es, more “advanced” topics such as Networking, Storage and nisms should be restricted to those programming language syntax and se- Analysis, that can perform some kind of model- mantics are typically taught much later Denver, CO, ing function over another domain or on. This begs the question: How can Contact: Bernd Mohr, Email: [email protected] world. That is, they can be set up in one learn to program in the absence such a way that their operation, when of such knowledge? Research shows November 13–16 viewed in the context of the other do- that concentrating on explaining how ICCAD ‘17: IEEE/ACM main, can be seen as simulating some programs work, rather than writing International Conference on Computer-Aided Design, aspect of that domain. Hence, a pro- them, helps students early on to learn Irvine, CA, gramming language can be used to programming. Could it be that we nor- Contact: Sri Parameswaran, model an aspect of the real world; a mally teach “by example” only, rather Email: [email protected] processor can be set up with appro- than ever properly defining the domain priate machine code to model a com- in which the modeling is performed, or putation expressed in a programming even the domain being modeled? language; a lambda calculus expres- sion, under the application of reduc- Our Curriculum tion, can provide the result of some The resulting curriculum is formally recursive function. structured around these three as-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 27 viewpoints

pects. Here we outline how they are could make lists, tables, family trees, presented to non-computer scien- Venn diagrams, and data models to tists—see the detail at http://www. Although solutions capture key information from the teachcs.scot. The vocabulary and con- can be created problems they are working on. cepts used are accessible to those who Learners will use their skills in lan- need to read them; the difficulty of in many ways, guage to create descriptions of pro- this should not be underestimated, it it is expected cesses that can be used by other peo- is hard for an academic computer sci- ple. For example, a computer program entist to communicate with a teacher that all learners is a great way to describe a process. of early years computing. will experience Learners will understand how to Each of our three main aspects per- read, write, and translate between sists through the five defined levels of creating solutions different representations such as be- the curriculum, from ages 3–15; the on computers. tween English statements, planning text here is mostly aimed at teachers of representations, and actual computer the lower levels. code. For example, developing skills Understanding the world through in writing code could be scaffolded computational thinking. The first as- by studying worked examples or by pect looks at the underlying theory in giving learners jumbled lines of code the academic discipline of Comput- and asking them to put the lines into ing Science. Theoretical concepts of an appropriate order. Computing Science include the char- cal and textual, used to represent pro- Although solutions can be cre- acteristics of information processes, cesses and information. Some of these ated in many ways, it is expected that identifying information, classifying representations are used by people and all learners will experience creating and seeing patterns. others by machines, for example, a ver- solutions on computers. This shows This aspect is about understand- bal description, a sequence of blocks learners that computers implement ing the nature and characteristics of in a visual programming language exactly what they—the learners—have processes and information. These can such as Scratch, or as a series of 1s and written, which is often not what they be taught through Unplugged activi- 0s in binary. intended, as well as giving them prac- ties (fun active learning tasks related In this aspect, learners will learn tice in debugging. to computing science topics but car- how to ‘read’ program code (before ried out without a computer) and with writing it in the next aspect) and de- Reflections structured discussions with learners. scribe its behavior in terms of the We have presented a curriculum that There is a focus on recognizing compu- processes they have learned about in explicitly connects computational tational thinking when it is applied in the first aspect, processes that will thinking with the more mechanical the real world such as in school rules, be carried out by the underlying ma- aspects of computing, with particular finding the shortest or fastest route chinery when the program runs. For concentration on the explicit modeling between school and home, or the way example, learners could read a section of computational domains by com- objects are stored in collections. of code and predict what will happen putational mechanism. Not everyone Learners will be able to identify steps when it runs or if lines of code change needs to become a software engineer and patterns in a process, for example order. Learners will learn and explore or ; the curriculum seeing repeated steps in a dance or different representations of informa- provides valuable learning at all lev- lines of a song. In later stages, learners tion and how these are stored and els, including the essential founda- will begin to reason about properties manipulated in the computing system tions for those who wish to study the of processes, for example considering under study. subject further. While our curriculum whether tasks could be carried out at Designing, building, and testing is informed by previous educational the same time, whether the output of a computing solutions. The third as- computing research, we emphasize process is predictable, and how to com- pect is about taking the concepts and quite different learning outcomes via pare the efficiency of two processes. understanding from the first two as- our three-point framework. Learners will identify information, pects and applying them. Learners

classify it, and see patterns. For exam- will create solutions, perhaps by de- Richard Connor ([email protected]) is a ple, learners might classify and group signing, building, and testing solu- Professor of Computer Science at the University of objects where there is a clear distinction tions on a computer or by writing a Strathclyde, Scotland.

between types or where objects might computational process down on pa- Quintin Cutts ([email protected]) is a belong to more than one category. per. In doing so, they will learn about Professor of Computer Science Education and Director of the Centre for Computing Science Education at the Understanding and analyzing com- modeling process and information University of Glasgow, Scotland. puting technology. This aspect aims to from the real world in programs, and give learners insight into the hidden what makes a good model to represent Judy Robertson ([email protected]) is a mechanisms of computers and the or solve a particular problem. Professor of Digital Learning at the University of Edinburgh, Scotland. programs that run on them. It explores Learners will create representations the different kinds of language, graphi- of information. For example, learners Copyright held by authors.

28 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints

VDOI:10.1145/3140822 Martin Spann, Lucas Stich, and Klaus M. Schmidt Viewpoint Pay What You Want as a Pricing Model for Open Access Publishing? Analyzing the “Pay What You Want” business model for open access publishing.

HE OPEN ACCESS publishing movement has received strong support from scien- tists, lawmakers, and fund- ing institutions. Many pub- lishers are reacting to this demand by T 7 offering open access journals. Howev- er, there is an ongoing debate on how open access publishing models should be financed.2,4 Most open access jour- nals that rely on the so-called “gold open access model”8—which makes the research output immediately avail- able from the publisher—let authors of accepted papers pay article processing charges (APCs) of several hundred to several thousand U.S. dollars.1 Howev- er, APCs are often criticized for poten- tially excluding researchers with limit- ed funds.3 As one potential solution to this problem and also to gain a better un- derstanding of the role of APCs in the scientific community, some publishers are starting to use Pay What You Want (PWYW) as a pricing model for gold open access publishing. PWYW is a pricing model where sellers delegate Sciences, and Thieme Publishers have covering the fields of surgery and engi- the full pricing power to buyers. So far, started to experiment with the PWYW neering, and The Surgery Journal PWYW has mainly been applied in ser- model for APCs of open access jour- (Thieme Publishers), an open access vice industries (for example, restau- nals. More specifically, Cogent OA em- journal for surgeons and trainee sur- rants, theaters), but also for the sale of powers authors to decide how much geons of all disciplines, have started to digital products like software (such as they want to pay for their open access delegate pricing power to their contrib- http://humblebundle.com). More re- publication in 15 broad journals cover- uting authors. cently, several publishers of open ac- ing different domains of academic re- Although neoclassical economic cess journals like Cogent OA (belong- search. Likewise, SICOT-J (edp scienc- theory predicts that buyers (that is, au-

IMAGE BY RT IMAGES/SHUTTERSTOCK RT BY IMAGE ing to the Taylor & Francis Group), edp es), a multidisciplinary journal thors in OA publishing) pay nothing if

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 29 viewpoints

Distribution of PWYW payments according to payment form.

State Payment Specify Discount $2,000

Outlier $1,500 $1,000 Payment in US$

Advertise with ACM! $500

Reach $0 and thought leaders The box indicates the lower and upper quartile and the line within depicts the median of the payments. The whiskers extend to include all data points within the 1.5 interquartile working at the range (IQR) of the upper/lower quartile and stop at the largest/smallest such value. cutting edge of computing they are not forced to do so, empirical cluded from publishing because they research on PWYW consistently finds can adjust their PWYW payments to and information that many buyers pay positive prices, of- their available means. technology through ten exceeding the marginal costs of the For publishers, PWYW achieves en- product (for example, Gneezy et al5). In dogenous price discrimination and ACM’s magazines, the case of open access publishing, au- higher market penetration because no websites thors may be willing to voluntarily pay author is excluded by APCs that he can- APCs for reasons of fairness and reci- not afford. This can be especially im- and newsletters. procity—because they want to com- portant for the introduction of a new pensate the publisher for his costs or to scientific journal. Furthermore, PWYW reciprocate his generosity, as well as can initiate a debate about funding ◊◆◊◆◊ for strategic considerations, that is, be- and affordability of APCs and might in- cause authors understand the journal crease the acceptance of open access will not be sustainable if the produc- APCs in the scientific community. The Request a media kit tion costs are not covered.9 Further- obvious risk, however, is free riding of with specifications more, with their payment authors may authors who do have the funds to pay signal to others (and potentially to for APCs but choose to pay nothing or and pricing: themselves) the value they attach to only very little. A more specific concern their publication.6 At the same time, in the context of open access is that Ilia Rodriguez authors with limited funds are not ex- some authors may be fundamentally convinced that research articles should +1 212-626-0686 be “free” to both readers and authors [email protected] With their payment instead of just “open” and for this rea- son refuse a voluntary payment. authors may Some insights on the performance signal to others of PWYW for open access APCs at the individual level are provided by initial (and potentially data from the peer-reviewed journal to themselves) The Surgery Journal launched by Thieme Publishers in June of 2015 that the value they attach exclusively uses PWYW pricing. At this to their publication. peer-reviewed journal, authors are prompted after acceptance of their arti- cle to state the APCs. After an article is accepted authors are directed randomly to one of two

30 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints possible forms (that is, experimental scientists to open access publishing conditions) to make their PWYW pay- and to raise the visibility and the posi- ments. On both forms, all authors are For publishers, using tive perception of their journals. informed about the recommended the PWYW pricing However, there are some points that “regular price” of the publication fee of we cannot address in our study. First, it $1,600. On one form they have to state model can be a is unclear to what extent the payments their PWYW payment directly. On the strategy to attract we have seen in The Surgery Journal de- other form they have to specify a dis- pend on the journal’s field (for exam- count from the regular price (a 100% even more scientists ple, medicine). It is likely that financial discount was possible).10 to open access resources vary across different scien- So far, authors of 27 papers made tific disciplines and thus affect volun- their payment decisions with a mean publishing. tarily paid APCs. Consequently, the payment of $480 across conditions. Only PWYW model may be better suited for four (15%) authors paid zero, and eight some publications and fields than for (30%) authors paid less than $100. Two others. In this regard we would like to authors paid the recommended “regular refer to the experiment at Cogent OA, price” for the publication fee of $1,600 even more than regularly asked. From where the comparison of payments and one author even paid 50% more discussions with the publisher, we across the 15 journals can allow for than the recommendation ($2,200). Al- learned the submissions exceeded cautious conclusions in this direction. together, five (19%) authors paid $1,000 their expectations compared to similar In general, we believe we need more ex- or more. In the condition where authors new open access journal introductions periments on PWYW as a pricing mod- had to specify their discount from the in this field. Also, the share of about el for open access publishing and our recommended publication fee, mean 40% of publications from developing contribution can be seen as a starting payments are substantially higher ($616 countries where authors might other- point in this direction. vs. $333; see the accompanying figure wise not be able to pay the regular APCs for the distribution of PWYW payments can be regarded as rather high. It must References 1. Beaudouin-Lafon, M. Open access to scientific according to payment form). However, be noted that many publishers have in- publications. Commun. ACM 53, 2 (Feb. 2010), 32–34. 2. Boisvert, R.F. and Davidson, J.W. Positioning ACM for an the standard deviation of the payment troduced schemes that give authors a open access future. Commun. ACM 56, 2 (Feb. 2013), 5. amounts is high ($588.7), which might significant discount or the option to 3. Celec, P. Open access and those lacking funds. Science 303, 5663, (2004), 1467. be attributed to heterogeneity in social completely waive APCs if the authors 4. Cerf, V.G. Open access. Commun. ACM 56, 4 (Apr. 2013), 7. preferences, available funds as well as cannot afford them. However, such ar- 5. Gneezy, A. et al. Shared social responsibility: A field experiment in pay-what-you-want pricing and cultural norms. The 27 publications rangements often require the disclo- charitable giving. Science 329, 5989 (2010), 325–327. came from a total of 12 different coun- sure of the financial situation of the 6. Gneezy, A. et al. Pay-what-you-want, identity, and self-signaling in markets. Proceedings of the National tries, which differ in terms of develop- authors that can be uncomfortable for Academy of Sciences of the of America ment. 11 publications came from devel- them. In contrast, the advantage of 109, 19 (2012), 7236–7240. 7. Malakoff, D. Scientific publishing. Opening the books oping countries, 16 from developed PWYW is that it allows authors to pay a on open access. Science 302, 5645 (2003), 550–554. countries (thereof six from the U.S.). reduced amount and thus allows them 8. Mann, F. et al. Open access publishing in science. Commun. ACM 52, 3 (Mar. 2009), 135. Payments from developed countries (M to save their face as part of the regular 9. Schmidt, K.M. et al. Pay what you want as a marketing = $728.9, SD = $654.4) were significantly publication and payment process. strategy in monopolistic and competitive markets. Management Science 61, 6 (2015), 1217–1236. higher than those from developing Clearly, PWYW is credible only if the 10. Schröder, M. et al. Pay-what-you-want or mark-off- countries (M = $118.2, SD = $124.6) at journal takes the acceptance decision your-own-price—A framing effect in customer- selected pricing. Journal of Behavioral and the 1%-level using a Mann-Whitney U first and only then asks the authors how Experimental Economics 57, (2015), 200–204. test. These findings are directionally much they are willing to pay for the pub- consistent with preliminary and aggre- lication. One important aspect that will Martin Spann ([email protected]) is a Professor at the Institute of Electronic Commerce and Digital Markets, gated findings provided by Cogent OA. influence the sustainability of this pric- LMU Munich, Germany. They report that 55% of authors decid- ing model for scientific publications is Lucas Stich ([email protected]) is an Assistant Professor at ed to pay something and some authors whether funding bodies such as the NSF the Institute of Electronic Commerce and Digital Markets, LMU Munich, Germany. are paying even more than the recom- permit authors to make positive pay- Klaus M. Schmidt ([email protected]) is a Professor mended APC (http://bit.ly/2wM5X8l). ments voluntarily (within reasonable at the Department of Economics, LMU Munich, Germany. Although these constitute early em- limits) despite the common require- pirical results and the sample size is ment to use funds economically. See the EC Workshop on OA publishing: http://bit.ly/1MWD0tF limited, we can derive a few implica- We believe the high motivation of tions. The observed payments from our scientific authors to achieve a wide dis- The authors gratefully acknowledge data provision from Thieme Publishers and financial support by the German experiment together with the prelimi- semination of their work and an inter- Science Foundation (grants SCHM 1196/5-1 and SP nary results from other experiments est to support and help sustain the 702/2-1, and CRC TRR 190) and are also grateful for many helpful comments from the European Commission (that is, Cogent OA) indicate that journal that publishes their research Workshop on “Alternative Open Access Publishing Models” in Brussels on Oct. 12, 2015 and the Open Access Days PWYW may work in the context of open facilitate the applicability of PWYW in 2016 in Munich. There are no financial interests of the access publishing. The results suggest open access publishing. For publish- authors involved. that a substantial fraction of authors ers, using the PWYW pricing model do pay APCs voluntarily, in some cases can be a strategy to attract even more Copyright held by authors.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 31 viewpoints

VDOI:10.1145/3148265 Virginia Dignum Viewpoint Social Agents: Bridging Simulation and Engineering Seeking better integration of two research communities.

HE USE OF the agent para- digm to understand and design complex systems oc- cupies an important and growing role in different ar- Teas of social and natural sciences and technology. Application areas where the agent paradigm delivers appropriate solutions include online trading,16 di- saster management,10 and policy mak- ing.11 However, the two main agent ap- proaches, Multi-Agent Systems (MAS) and Agent-Based Modeling (ABM) dif- fer considerably in methodology, appli- cations, and aims. MAS focus on solv- ing specific complex problems using autonomous heterogeneous agents, while ABM is used to capture the dy- namics of a (social or technical) system for analytical purposes. ABM is a form of computational modeling whereby a population of individual agents is given simple rules to govern their behavior agents at the macro level, whereas MAS setting, indirectly affects the environ- such that global properties of the whole are operational systems, acting and af- ment. And, the design of MAS is often can be analyzed.9 The terminology of fecting its (physical) environment, with geared to analytic insights and simu- ABM tends to be used more often in the a focus on solving specific practical or lations toward the understanding of social sciences, whereas MAS is more engineering problems, and emphasiz- how configurations of agents behave used in engineering and technology. ing agent architectures with sophisti- in different circumstances. Currently, Although there is considerable overlap cated reasoning and decision process- applications of MAS are broader than between the two approaches, histori- es. This has lead to the development of pure distribution problems, including cally the differences between ABM and two research communities proceeding interactive virtual characters, where MAS are often more salient than their on nearly independent tracks. the focus is on the cognitive, affective, similarities. For example, it is often re- However, this division is not as and emotional characteristics of the marked that a main difference between black and white as it may seem. In fact, system, and game-theoretic models, ABM and MAS is that ABM models are much ABM work goes beyond descrip- focusing on the design of incentive descriptive aiming at explanatory in- tive simulations of a situation, and, as mechanisms that guarantee a given

sight into the collective behavior of input for decision making and policy strategic behavior. PONGVANIT PHANU D. BY IMAGE

32 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 viewpoints viewpoints

Social abilities are central both in tions, and the ability to create, structure ABM, where agents represent humans and ‘rationalize’ the environment to fit and their interactions, and in MAS, Unfortunately, ones expectations and abilities (lead- that enable game-theoretic analyses of from a modeling ing, for example, to the design of orga- decision strategies, or provide interac- nizations, institutions, and norms). tive virtual agents in varied situations. perspective, Following an increasing number of It is precisely in this area where the real human behavior researchers in both ABM and MAS that need for integration of ABM and MAS in recent years have come to similar is neither simple 7,13,18,19 V is undoubtedly the most necessary. In conclusions, we claim that new social simulation, the benefits of com- nor rational. models of preference and belief forma- bining MAS and ABM have been advo- tion are needed that show how behavior cated for many years, and are the focus derives from identities, emotions, moti- of the long-lasting workshop series on vation, values, and practices.6 Multi-Agent Based Simulation (MABS).2 The endeavor required to construct ABM has increasingly and successfully such agent models that are socially real- been used for social simulations,3 but whereas MAS focuses on solving specif- istic requires the effort and the capabil- it is in the MAS area that fundamental ic problems using independent agents, ities of both the MAS and ABM commu- research on agent architectures imple- through the formalization of the com- nities, bringing together formalization menting psychological traits and social plex goal-oriented processes, such as and computational efficiency, and concepts such as norms, commitments, the Beliefs-Desires and Intentions (BDI) planning techniques as in MAS, with emotions, identity, and social order, has model proposed by Bratman20 or game- the ABM expertise on empirical valida- been most prominent.4,5 Bridging these theoretic approaches. tion and on adapting and integrating somewhat parallel tracks requires a new The main advantages of such ra- social sciences theories into a unified grounding for agent architectures. tionality assumptions are their parsi- set of assumptions,1 furthering the mony and applicability to a very broad fundamental understanding of social Questioning Rationality range of situations and environments, deliberation processes, and develop- Traditionally, one of the most salient and their ability to generate falsifiable, ing techniques to make these acces- aspects shared by both ABM and MAS and sometimes empirically confirmed, sible for simulation platforms. This approaches is the premise of rational- hypotheses about actions in these en- Viewpoint is therefore an appeal to join ity. This is derived from the traditional vironments. This gives conventional the strengths of both communities to- definition of agents as autonomous, rational choice approaches a combina- ward sociality-based agents. proactive, and interactive entities where tion of generality and predictive power Without claiming a readily available each agent has bounded (incomplete) not found in other approaches. In fact, solution, we propose the concept of resources to solve a given problem; rationality approaches are the basis of sociality as the leading principle of there is no global system control; data most theoretical models in the social agency, as an alternative for rational- is decentralized; and computation is sciences, including economics, politi- ity. Following the aforementioned asynchronous.21 Agent rationality can cal science, or social choice theories. description of rational behavior, the be summarized as follows: Unfortunately, from a modeling per- main characteristics of sociality- ˲˲ Agents hold consistent beliefs; spective, real human behavior is neither based reasoning are: ˲˲ Agents have preferences, or priori- simple nor rational, but derives from a ˲˲ Ability to hold and deal with incon- ties, on outcomes of actions; and complex mix of mental, physical, emo- sistent beliefs for the sake of coherence ˲˲ Agents optimize actions based on tional, and social aspects. Realistic appli- with identity and cultural background. those preferences and beliefs. cations must consider situations in That is, beliefs originate from other This view on rationally entails that which not all alternatives, consequenc- sources than observation, including agents are expected, and designed, es, and event probabilities can be fore- ideology or culture. to act rationally in the sense that they seen. This renders rational choice ap- ˲˲ Ability to fulfill several roles, and choose the best means available to proaches unable to accurately model and pursue seemingly incompatible achieve a given end, and maintain con- predict a wide range of human behaviors. goals concurrently, for example, si- sistency between what is wanted and multaneously aiming for comfort what is chosen.14 Even though mul- Toward Social Agents and environmental friendliness, or tiple alternatives have been proposed, Human sociability refers to the nature, for riches and philanthropy. in both the ABM and MAS approaches, quantity, and quality of interactions ˲˲ Preferences are not only a cause individual agents are still typically char- with others, including both pro-social, for action but also a result of action. acterized as bounded rational, acting or cooperative, behaviors, and conflict, Moreover, preferences change signifi- toward their own perceived interests. competitive, or dominating behaviors. cantly over time and their ordering is The main difference is that agent be- Sociability is also the ability to influence influenced by the different roles being haviors in ABM are used to capture the others, by changing their behaviors, fulfilled simultaneously, which requires dynamics of a system for analytical pur- goals, and beliefs, the emotional reac- the need to deal with misalignment and poses, grounded whenever possible on tion to others and to the environment, incompatible orderings. existing data about system outcomes, and how actions are affected by emo- ˲˲ Action decisions are not only geared

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 33 viewpoints

to the optimization of own wealth, but search path linking theory, model, and often motivated by altruism, fairness, implementation, and suggested possi- justice, or by an attempt to prevent re- Sociality-based ble theories and techniques to develop gret at a later stage. agents are also sociality-based agents. These incorpo- ˲˲ Understand when there is no need rate expertise from both ABM and MAS to further maximize utility beyond some fundamental to the and require integration of both areas in reasonably achievable threshold. new generations of order to succeed. We welcome the dis- ˲˲ Understand how identity, cul- cussion of these issues toward a novel ture, and values influence action, and intelligent devices. area of research on social agents, which use this knowledge to decide about take sociability as the basis for agent de- reputation and trust about who and liberation and enable interaction. how to interact. The first step toward sociality-based References 1. Chai, S. Choosing an Identity: A General Model agents is a thorough understanding of of Preference and Belief Formation. University of social practices as a kind of shortcuts Michigan Press, 2001. these principles, and open discussion 2. Conte, R., Gilbert, N., and Sichman, J. MAS and social across disciplines on the grounds and for deliberation.15,17 Where it concerns simulation: A suitable commitment. In J. Sichman, R. Conte, and N. Gilbert, Eds, Multi-Agent Systems requirements for sociality from differ- utility, satisficing can be more suitable and Agent-Based Simulation, volume 1534 of ent perspectives. This discussion will approach than maximizing.12 This also Lecture Notes in Computer Science,. Springer Berlin Heidelberg, 1998, 1–9 be fundamental to the development of allows us to integrate agents of varied 3. Davidsson, P. Agent based social simulation: A formal models and agent architectures richness levels, for example, using computer science view. Journal of Artificial Societies and Social Simulation 5, 1 (2002). that make sociality-based behavior pos- rich cognitive models to zoom-in the 4. Dias, J., S. Mascarenhas, and A. Paiva. Fatima sible and verifiable. behavior of salient agents in a simula- modular: Towards an agent architecture with a generic appraisal framework. In Proceedings of the Moreover, it is necessary to iden- tion, whereas other agents just follow International Workshop on Standards for Emotion tify and formalize which mechanisms, simple rules. This approach can coun- Modeling, 2011. 5. Dignum, F., Dignum, V., and Jonker, C.M. Towards other than imitation, can describe ter the obvious criticism that sociality- agents for policy making. In MABS IX, Springer, 2009, how agents can adapt to pressures based agents will become too complex 141–153. 6. Dignum, F. et al. A conceptual architecture for social in the environment to behave in a so- for use in computational simulations. deliberation in multi-agent organizations. Multiagent cially acceptable, resource-sustainable Sociality-based agents are also fun- and Grid Systems 11, 3 (2015), 147–166. 7. Dignum, F., Prada, R., and Hofstede, G.J. From autistic fashion. Resulting models support the damental to the new generations of in- to social agents. In AAMAS 2014, May 2014. 8. Dignum, V. Mind as a service: Building socially understanding or predicting human telligent devices, and interactive char- intelligent agents. In V. Dignum, P. Noriega, M. Sensoy, behavior, including rich models of acters in smart environments. These and J. Sichman, Eds, COIN XI: Revised Selected Papers, Springer International Publishing, 2016, 19–133. emotions, identities, culture, values, artifacts not only must build (partial) 9. Epstein, J.M. and Axtell, R. Growing Artificial Societies: norms, and many other socio-cogni- social models about the humans they Social Science from the Bottom Up. The Brookings Institution, Washington, D.C., 1996. tive characteristics. Such models of interact with, but also need to take so- 10. Fiedrich, F. and Burghardt, P. Agent-based systems social reality are also needed to study cial roles in a mixed human/digital real- for disaster management. Commun. ACM 50, 3 (Mar. 2007), 41–42. the complex influences on behavior ity. An interesting challenge would be to 11. Ghorbani, A. Enhancing abm into an inevitable tool of different socio-cognitive character- use the same technologies in real time for policy analysis. Policy and Complex Systems 1, 1 (2014). istics and their relationships. The in- mixed human/artificial interactions, 12. Gigerenzer, G. Moral satisficing: Rethinking moral tegration of psychological models of and criticisms could also be on the fea- behavior as bounded rationality. Topics in 2, 3 (2010), 528–554. motivation and cognition, sociological sibility to use these architectures (or 13. Kaminka, G. Curing robot autism: A challenge. In theories of value and identity forma- controlled reductions/simplifications) Proceedings of the AAMAS 2013, May 2013), 801–804. 14. Lindenberg, S. Social rationality versus rational tion, and philosophical theories of co- in real time or near real time. egoism. In Handbook of Sociological Theory, Springer, herence and higher-order rationality, 2001, 635–668. 15. Reckwitz, A. Toward a theory of social practices. together with different formal meth- Moving Forward European Journal of Social Theory, 5, 2 (2002), 243–263. 16. Rogers, A. et al. The effects of proxy bidding and ods, quickly yields intractable models. The intent of this Viewpoint has been minimum bid increments within ebay auctions. ACM However, it is important to identify to appeal for a collaborative research Trans. Web 1, 2 (Aug. 2007). 17. Shove, E., Pantzar, M., and Watson, M. The Dynamics of what is the model being developed for. effort toward fundamental formal theo- Social Practice. Sage, 2012. In fact, richer models are not always ries and models that increase our un- 18. Silverman, B. et al. Rich socio-cognitive agents for immersive training environments: The case of nonkin the most appropriate ones. derstanding of the principles behind village. Journal of Autonomous Agents and Multi- Once these characteristics are well human deliberation (such as the ones Agent Systems 24, 2 (Mar. 2012); 312–343. 19. Vercouter, L. et al. An experience on reputation models understood, then simplified mod- listed discussed here), before deciding interoperability based on a functional ontology. els can be developed to suit different on which modeling techniques we need In Proceedings of IJ-CAI’07, Morgan Kaufmann Publishers Inc., San Francisco, CA, 2007, 617–622. needs. That is, implementing sociality- to implement them. Even though, sev- 20. Wooldridge, M. Reasoning about Rational Agents. MIT based agents will require other tech- eral approaches to model social aspects Press, 2000. 21. Wooldridge, M. An Introduction to Multiagent niques than those currently used in in agent behavior are available, there is Systems. Wiley, New York, 2009. either MAS or ABM,8 including the use not sufficient consensus on which char- Virginia Dignum ([email protected]) is an associate of simpler, context-specific decision acteristics are needed for what, nor on professor with the Faculty of Technology, Policy and rules, mimicking how people them- how to specify and integrate them. We Management, Delft University of Technology, Delft, selves are able to deal with complex have identified an initial set of charac- The Netherlands. decision making, for example, using teristics for sociability, proposed a re- Copyright held by author.

34 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 Inviting Young Scientists

Meet Great Minds in Computer Science and Mathematics

As one of the founding organizations of the Heidelberg Laureate Forum http://www.heidelberg-laureate-forum.org/, ACM invites young computer science and mathematics researchers to meet some of the preeminent scientists in their field. These may be the very pioneering researchers who sparked your passion for research in computer science and/or mathematics. These laureates include recipients of the ACM A.M. Turing Award, the Abel Prize, the Fields Medal, and the . The Heidelberg Laureate Forum is September 23–28, 2018 in Heidelberg, Germany. This week-long event features presentations, workshops, panel discussions, and social events focusing on scientific inspiration and exchange among laureates and young scientists.

Who can participate? New and recent Ph.Ds, doctoral candidates, other graduate students pursuing research, and undergraduate students with solid research experience and a commitment to computing research How to apply: Online: https://application.heidelberg-laureate-forum.org/ Materials to complete applications are listed on the site. What is the schedule? The application process is open between November 6, 2017 and February 9, 2018. We reserve the right to close the application website early depending on the volume Successful applicants will be notified bymid April 2018. More information available on Heidelberg social media

PHOTOS: ©HLFF / B. Kreutzer (top); ©HLFF / C. Flemming (bottom)

ACM_HLF_ad-2017_cacm.indd 1 10/2/17 5:46 PM practice

DOI:10.1145/3106627 TERRY COATTA: I’m curious about the Article development led by queue.acm.org original set of problems Hootsuite was looking to address in the switch to microservices. Can you provide A discussion with Edward Steel, Yanik Berube, some detail? Jonas Bonér, Ken Britton, and Terry Coatta EDWARD STEEL: Mostly, it had to do with our ability to send out notifica- tions to user mobile devices when- ever something relevant happened on Twitter. By the time we started having Hootsuite: some concerns about how we were handling that, we were already servic- ing several hundreds of thousands of users, each with individual subscrip- tions tailored to their own specific In Pursuit interests. What was needed was some- thing that could stay connected to Twitter’s streaming endpoints. COATTA: I gather that at about the of Reactive same time you were making this move, you also took steps to move from PHP to Scala. What drove that? STEEL: Initially, it had a lot to do with Systems learning about all the success some other organizations had experienced with Scala. This was after Twitter had decided to go with Scala, for example, and that obviously lent a lot of legiti- macy to it. Also, the first team here to BASED IN VANCOUVER, Canada, Hootsuite is the most work with Scala came from quite a var- ied background. We had some people widely used SaaS (software as a service) platform for who were lobbying for a more strongly managing social media. Since its humble beginnings in typed functional language—some- thing on the order of Haskell—and 2008, Hootsuite has grown into a billion-dollar company then there were some others with Clo- with more than 15 million users around the globe. jure and Java experience. In taking all As Hootsuite evolved over the years, so did the that into account, I guess Scala just seemed to check most of the boxes. technology stack. A key change was moving from LAMP JONAS BONÉR: What would you say (Linux, Apache, MySQL, PHP) to microservices. A shift was the principal benefit you saw with to microservices didn’t come without its challenges, Scala? Was it the functional nature of the language itself? Or did it have more however. In this roundtable chat, we discuss how Scala to do with the libraries available within and Lightbend (which offers a reactive application that ecosystem? STEEL: The language itself was the development platform) were an essential part of a biggest part of it. The main advocate successful transition. The exchange includes Jonas here for Scala was working on a Black- Bonér, CTO of Lightbend; Terry Coatta, CTO of Marine berry client at the time, so he had a lot of JVM (Java virtual machine) knowl- Learning Systems; Edward Steel, senior Scala developer edge and yet also had become frus- at Hootsuite; Yanik Berube, lead software developer at trated with Java itself. I guess he was just looking for a better way. Another Hootsuite; and Ken Britton, senior director of software aspect of our thinking had to do with development at Hootsuite. building a distributed system that

36 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 could take advantage of Akka as an thought being that perhaps we could pect to fire a lot of messages around, available library. That was a big part of then move toward a service-oriented you can expect the actor model to re- the decision as well. architecture. The idea of microservices ally shine. In fact, I think we were able to use wasn’t something that came up until a YANIK BERUBE: Just in terms of where some actors right from the start. That little later, and that was probably influ- this fits into our current infrastructure, was with a very early version of Akka, enced by some of the buzz around the we should note that all our microser- but it still offered a lot of compelling industry at the time. vices are powered by Akka. Internally, features we found useful. COATTA: You have already mentioned we have a server-type library that han- BONÉR: Were you already thinking in Akka a couple of times, so can you dles requests and responses via Akka, terms of microservices back then, even speak to how that fits in here? and we also have at least one, if not before that took off as a buzzword? Or STEEL: At that time, at least, the main more, back-end services that use sets were you more drawn by reactive prin- appeal of the Akka system for the JVM of actors to accomplish work at certain ciples having to do with things like a was that it provided people beyond the intervals of time. share-nothing architecture, strong iso- Erlang community with access to the COATTA: One of the things that lation, and loose coupling? actor model. The thing about actors is comes to mind when I think of the STEEL: Microservices were always in that they are both message based and Erlang actor system, beyond the in- the back of our minds. We already had highly resilient—which is to say that dependence you have already men- some batch processes written in PHP even when they crash, they can typi- tioned, is that it’s quite fine-grained. that were starting to run jobs at that cally be brought back in a useful way. So I wonder, given your focus on no- point. That sort of worked, but it was This probably explains why Erlang has tifications to user mobile devices, far from an ideal way of doing things. so often been used to develop resilient whether you might actually require an So I think we had already started to de- telecommunications systems. When- actor per user just to deal with that? velop an appetite for a system on which ever you are talking about distributed STEEL: In our case, no. But symp-

IMAGES COURTESY OF HOOTSUITE MEDIA INC. OF HOOTSUITE COURTESY IMAGES we could build a few services, with the systems or some system where you ex- toms of that definitely showed up as we

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 37 practice

JONAS BONÉR Something I find very interesting is that Akka and Erlang appear to be the only platforms or libraries that put an emphasis on embracing and managing failure— were first building our system. When sentially glue business logic onto the which is to say we were starting out, we learned more data part of the logic. But in terms of they are basically about how we should be building the the size of these things, I’d say we’re system, as well as about how actors re- still trying to figure that out, and we designed for ally ought to work. At first, we definitely haven’t come up with any hard and fast resilience. fell into the trap of putting too much rules so far. logic into single actors—for example, BONÉR: I’m guessing each of your mi- by putting recovery logic into each ac- croservices owns its own data store. If tor instead of relying on the supervi- so—with these things being stateful— sion hierarchy, which would have al- how are you then able to ensure resil- lowed us to code less defensively. It ience across outages? turns out it’s best just to embrace the STEEL: Each of these services abso- “let it crash” philosophy, since that ac- lutely owns its own data. When it comes tually offers a lot of robustness. to replacing parts of the monolithic sys- We also learned the model really tem, it generally comes down to dealing shines whenever you separate con- with a table or a couple of related tables cerns into single-purpose actors. Be- from the LAMP MySQL database. Gen- sides helping to clarify the design, erally speaking, the space is pretty min- it opens up a lot of opportunities in imal in terms of the services that need terms of scalability and configuration to be accounted for. It’s basically just a flexibility at the point of deployment. matter of retrieving and creating data. BONÉR: This applies to microservices BERUBE: I would say we make fairly as well. That is, there are plenty of heterogeneous use of various technol- opinions about what even defines a mi- ogies for data storage. And, yes, we do croservice. What does that term mean come from a LAMP stack, so there is to you? And how does that map concep- still a heavy reliance on MySQL, but we tually to how you view actors? also make use of MongoDB and other BERUBE: Internally, we are still try- data-storage technologies. ing to define what a microservice is The services typically each encap- and what the scale of that ought to be. sulate some area of the data. In theory, Today, most of our services focus on at least, they are each supposed to own accomplishing just one set of highly re- their data and rely on data storage dedi- lated tasks. Our goal is to have each of cated only to them. So, we have recently these services own some part of our do- started looking into storing a bit more main model. Data services, for exam- data within the services themselves for ple, would each control their own data, reasons of efficiency and performance. and nothing else would have access to COATTA: To be clear, then, there’s that. To get to that data, you would have some separate data-access layer to go through the data service itself. from which the services are able to Then we would also have functional pull in whatever information they services, which are the services that es- need to manipulate?

38 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice

TERRY COATTA One of the things that comes to mind when I think of the Erlang actor system, beyond its independence, is that it’s quite fine-grained.

STEEL: Yes, but you won’t see more ploy code on the Scala side with things than one service accessing the same switched off. data store. BONÉR: Looking back to when you BONÉR: How do you manage this in were doing this along with every- terms of rolling out updates? Do you thing else required to maintain a have some mechanism for deploying monolith, what do you see as the updates, as well as for taking them biggest benefits of having made the down and rolling back? Also, are move to microservices? these services isolated? If so, how did STEEL: In terms of what it takes to you manage to accomplish that? And, scale a team, I think it has proved to if not, what are you doing to mini- be much easier to have well-defined mize downtime? boundaries within the system, since BERUBE: Right now, every service that means you can work with people uses a broker/worker infrastructure. who have only a general idea of how No service can access another ser- the overall product works but augment vice without going through a pool of that with a strong, in-depth knowledge brokers that then will redistribute re- of the specific services they’re person- quests to multiple workers. This gives ally responsible for. us the ability to scale by putting more I also think the microservices ap- workers behind the brokers. Then, proach gives us a little more control when it comes to deployment, we can operationally. It becomes much easier do rolling deploys across the target to scale and replace specific parts of servers for those workers. In this way the system as those needs arise. we are able to deploy the newest ver- BERUBE: One clear example of this sion of a service gradually without is the data service my team has been affecting the user experience or the working on. It’s a very high-volume experience of any other services that service in terms of the amount of data need to make use of that service as it’s we store, and we knew it would be chal- being redeployed. lenging to scale that, given the storage STEEL: Another thing we have recent- technology we’re using. The ability ly pulled over from the LAMP side that to isolate all that data behind a data has proved to be useful for frequent service makes it a lot easier for us to rollouts is feature flagging. That’s ob- implement the necessary changes. viously something that was a lot easier Basically, this just gives us a lot more back when we just were working with control over what it takes to change a bunch of Web servers, since we had the persistent technology we’re using a central place for handling it. But re- in the background. So I certainly see cently we migrated those same capa- this as a big win. bilities to HashiCorp’s Consul to give ourselves a distributed, strongly con- Some of the benefits of moving over sistent store, and that now lets us de- to the reactive microservice model

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 39 practice

EDWARD STEEL Because of our ability to change the characteristics of actors by how we configure them, we have been able to adapt this core framework to all types of payloads supported by the Lightbend stack sur- COATTA: So far, we have talked only and traffic profiles faced almost immediately as the Hoot- about general issues. Now I would like for the various suite engineering team started discov- to hear about some of your more spe- ering opportunities for scaling down cific engineering challenges. services. on the underlying physical and virtual BONÉR: One thing I would like to infrastructure they had run previously know is whether you’re mostly doing on their LAMP stack (where there had reactive scaling, predictive scaling, or been a process for each request). Ac- some combination of the two to opti- cordingly, it soon became apparent mize for your hardware. that operations under the reactive mi- BERUBE: For now at least, our loads croservices model were going to put don’t really change a lot. Or perhaps much less strain on their resources. what I should say is that they change In fact, if anything, the engineers throughout the day, but predictably so at Hootsuite quickly learned that by from one day to the next. And the way continuing to employ some of the prac- we have designed our services to run tices that had made sense with a LAMP behind brokers means we are able just stack, they would actually be denying to spin off more workers as necessary. themselves many of the benefits avail- In combination with some great tool- able by relying to a greater degree on ing from AWS (Amazon Web Services), the Lightbend stack. For example, they we are able to adapt quickly to chang- found there was a real advantage to ing workloads. making greater use of the model class- STEEL: One thing we did decide to es supported by the Lightbend stack, do was to build a framework using Ze- since those classes come equipped roMQ to enable process communica- with data-layer knowledge that can tion between our various PHP systems. prove to be quite useful in a dynamic But then we saw later that we could have web-oriented system. just as easily pulled all that into Akka. Similarly, they learned that by us- COATTA: And I’m assuming, with ing individual actors to run substan- Akka, it would have been easier for you tial portions of their system instead of to achieve your goal of adhering to the decomposing those components into actor paradigm, while also taking ad- groups of actors, they had been unwit- vantage of better recovery mechanisms tingly depriving themselves of some and finer-grained control. of the features Akka offers for tuning STEEL: Yes, but I think the key is that parts of the system separately, parallel- because of our ability to change the izing them, and then distributing work characteristics of actors by how we efficiently among a number of differ- configure them, we have been able to ent actors capable of sharing the load. adapt this core framework to all types And then there were also a few other of payloads and traffic profiles for the things they learned ... various services. We can say, “This ser- vice is using a blocking database,” at

40 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice

YANIK BERUBE One of the fantastic lessons to come out of all this is it has allowed me to start thinking about how the system actually works in terms of handling failures and dealing with which point a large thread pool will be be fairly explicit about how you handle supplied. If the service happens to be failures as a supervisor. That is, as an ac- the external agents handling two different kinds of jobs, tor that spins off other actors, you must we communicate we can separate them into different ex- have rules in place as to what ought to ecution tracks. happen should one of your child actors with via messages. More recently, we have also had a end up failing. But, yes, people had to fair amount of success using circuit be educated about that. And even then, breakers in situations where we have it still took a bit of getting used to. had a number of progressive timeouts Now, as new developers come in, we as a consequence of some third party see them resort to the more traditional getting involved. But now we can just patterns of handling exceptions. But cut the connection and carry on. Much once you get some exposure to how of this comes for free just because of much saner it is just to leave that to all the tools Akka provides. We have the supervisor hierarchy, there’s gen- learned that we can take much better erally no turning back. Your code just advantage of those tools by keeping becomes a lot simpler that way, mean- our designs as simple as possible. ing you can turn your focus instead to BONÉR: Something I also find very figuring out what each actor ought to interesting is that Akka and Erlang ap- be responsible for. pear to be the only platforms or librar- COATTA: Talking about actor frame- ies that put an emphasis on embracing works in the abstract is one thing, but and managing failure—which is to say what does this look like once the rub- they are basically designed for resil- ber actually hits the road? For example, ience. The best way to get the fullest how do you deal with failure cases? benefit of that is if they are part of your BERUBE: You can just let the actor application from Day One. That way, fail, which means it will essentially you can fully embrace failure right at die, with a notification of that then be- the core of your architecture. ing sent off to the supervisor. Then the But, with that being said, how did supervisor can decide, based on the se- this newfound embrace of failure work verity or the nature of the failure, how out in practice for those of your devel- to deal with the situation—whether opers who had come from other envi- that means spinning off a new actor or ronments with very different mind- simply ignoring the failure. If it seems sets? Was this something they were the problem is something the system able to accept and start feeling natural actually ought to be able to handle, it about in fairly short order? will just use a new actor essentially to BERUBE: For some, it actually required send the same message again. a pretty substantial mindset shift. But the point is that the actor model But I think Akka—or at least the actor allows you to focus all the logic related model—makes it easier to understand to the handling of specific failure cases the benefits of that, since you have to in one place. Because actor systems are

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 41 practice

KEN BRITTON It has become apparent how critical frameworks and standards are for development teams when using microservices.

hierarchies, one possibility is that you some failures. They are just going to will end up deciding the problem isn’t happen. This means we should not be really the original actor’s responsibil- banking on some external service re- ity but instead should be handled by sponding in a short amount of time. that actor’s supervisor. That’s because We want to explicitly set timeouts. the logic behind the creation of these Then if we see that the service is start- hierarchies determines not only where ing to fail very quickly or with some the processing is to be done, but also high frequency, we will know it’s where the failures are to be handled— time to trip a circuit breaker to ease which is not only a natural way to orga- the pressure on that service and not nize code, but also an approach that have those failures echo across all very clearly separates concerns. services. I have to say that came as a BONÉR: I think that really hits the bonus benefit I certainly was not ex- nail on the head. It comes down to pecting when we first started working distinguishing between what we call with Akka. errors—which are things that the user is responsible for dealing with—and Providing for greater efficiency in the failures. Validation errors then natu- utilization of system resources by re- rally go back to the user while less sorting to a distributed microservices severe errors go to the component architecture is one thing. But to what that created the service. This creates degree is that liable to end up shifting a model that is easier to reason about, additional burdens to your program- rather than littering your code with mers? After all, coding for asynchro- try/catch statements wherever fail- nous distributed systems has long ures might happen—since failures been considered ground that only the can, and will, happen anywhere in a most highly trained Jedi should dare distributed system. to tread. BERUBE: One of the fantastic les- What can be done to ease the tran- sons that has come out of all this is sition to a reactive microservices en- it has allowed me to start thinking vironment for programmers more about how the system actually works accustomed to working within the in terms of handling failures and confines of synchronous environ- dealing with the external agents we ments? Won’t all the assumptions communicate with via messages. Ba- they typically make regarding the sically, I started to think about how state of resources be regularly vio- we should handle the communica- lated? And how to get a large team of tion between services around the coders up the concurrency learning way we handle failures. So now that’s curve in reasonably short order? something we always think about. Here’s what the Hootsuite team The reality is that any time we talk learned ... to external services, we should expect

42 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice

COATTA: Let’s talk a little about the BONÉR: That matches my experi- to operate in that environment, as impact the move to microservices ence as well. Actors are very object well as to train people to work in it. has had on your developers. In par- oriented, and they encapsulate state Whereas before maybe it was suffi- ticular, I would think this means you and behavior—all of which I think of cient just to find some new hires that are throwing a lot more asynchrony as mapping well to a traditional ap- were proficient in PHP, now you have at them than most developers are ac- proach. Futures, on the other hand, ZeroMQ and actors and Futures and customed to. I imagine they probably lend themselves to functional think- any number of other things for them also have a lot more data consistency ing—with all these small, stateless, to wrap their heads around. Without issues to worry about now. one-off things you can compose eas- question, your environment has be- BERUBE: Although the asynchrony ily. But have you found you can actu- come more complex. But is it now in problem hasn’t been fully addressed, ally make these things derived from some respects also actually an easier Scala Futures (data structures used to two very different universes work well place in which to operate? retrieve the results of concurrent oper- together? Do you blend them or keep BERUBE: I think the act of dividing ations) actually make it really easy to them separate? the logic into a lot of different self-con- work with asynchronous computation. BERUBE: We have used them togeth- tained services has made it easier at I mean, it still takes some time to ad- er in parallel, and I think they work some level to reason about how the sys- just to the fact that anything and every- well that way. Ken mentioned this tem works. But we are not finished yet. thing can and will fail. But, with Scala idea of generating Futures and then There is still plenty of work to do and Futures, it’s actually quite easy for rela- piping them either to yourself as an lots of challenging areas to continue tively uninitiated programmers to actor or to some other actors. I think reasoning about. learn how to express themselves in an that pattern works quite well. It’s And, yes, of course, the environ- asynchronous world. both simple and elegant. ment has become a bit more complex. STEEL: If you are coming at this from STEEL: I have to admit I stumbled I have to agree with that. But the ben- the perspective of thread-based con- over that mental shift a bit early on. efits outweigh the drawbacks of roll- currency, it’s going to seem much scar- But, yes, I’d say we have been able to ing in all this technology, since we now ier for a lot of use cases than if you’re blend actors and Futures successfully have more layers of abstraction to take coming at it from a Futures point of for the most part. advantage of. We have teams that are view. Also, when you’re working direct- BONÉR: Do you feel that certain types generally aware of the big picture but ly inside actors, even though messages of problems lend themselves better to are mostly focused on just a few mi- are flying around asynchronously and one or the other? croservices they understand really well. the system is doing a thousand things STEEL: In our simpler services, the That’s an approach that will have huge at the same time, you are insulated routing of a request to the code is all ac- benefits for our operations as we scale from what it takes to synchronize any tor based, and then the actual business them moving forward. state modifications, since an actor will logic is generally written as calls to other BRITTON: It has become apparent process only one message at a time. things that produce Futures. I suppose how critical frameworks and stan- KEN BRITTON: I have noticed when that when you’re thinking about infra- dards are for development teams developers first start writing Scala, structure and piping things around, it’s when using microservices. People of- they end up with these deeply nested, very natural to think of that in terms ten mistake the flexibility microser- control-flow-style programs. You see a lot of actors. Business logic, on the other vices provide with a requirement to of that in imperative languages, but there hand, perhaps maps a little more readily use different technologies for each is no penalty for it. In a strongly typed to the functional point of view. service. Like all development teams, functional language, however, it is much BRITTON: We are also finding that a we still need to keep the number of more difficult to line up your types rich object-oriented model is helpful in technologies we use to a minimum so through a complex hierarchy. Developers our messaging. For example, we have we can easily train new people, main- learn quickly that they are better served started defining richer success and tain our code, support moves between by writing small function blocks and failure messages containing enough teams, and the like. then composing programs out of those. detail to let an actor know exactly how We have also seen a trend toward Akka takes this one step further by to respond. So, now our message hi- smaller services. Our initial microser- encouraging you to break up your logic erarchy has expanded to encapsulate vices were actually more like loosely with messages. I’ve observed a com- a lot of information, which we think coupled macroservices. Over time, mon evolution pattern where develop- nicely aligns object-oriented concepts though, we have pushed more of the ers will start off with these very bulky, with the actor model. deployment, runtime, and so forth into complex actors, only to discover later COATTA: One thing that occurred to shared libraries, tooling, and the like. that they could have instead piped a me as you talked about your environ- This ensures the services are focused on Future to their own actor or any other ment is that it seems you have moved logic rather than plumbing, while also actor. In fact, I’ve witnessed a number not only from a monolithic architec- sticking to team standards. of aha moments where developers hit ture, but also, in some sense, from upon the realization that these tools a monolithic technology to a much actually encourage them to build wider array of technologies. So I won- Copyright held by authors/owners. smaller composable units of software. der if you now find it more difficult Publication rights licensed to ACM. $15.00.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 43 practice

DOI:10.1145/3132267 people to know, and he responded Article development led by queue.acm.org that it didn’t matter—a good software engineer can work on anything. This has been the thinking at many We all wear many hats, but make sure large software companies in the past, you have one that fits well. and there are definitely merits to it—especially when you are hiring in- BY KATE MATSUDAIRA experienced candidates straight out of school. As I have worked longer in the industry, however, I have started changing my thinking. I would argue that it is important to go deep in at least one area, and it is almost always better to hire people who have a solid depth of experience in the tools and technology they are Breadth using. Why do you need to have deep knowledge? Really good software engineers for a particular language or technology will exhibit qualities such as these: and ˲˲ They are productive. They pro- duce an amount of work that is above average, and they are able to get things done. This means they know how to use the tools of their trade well and aren’t slowed down by not under- Depth standing something. They use their brainpower for harder problems, not learning how to do the basics. ˲˲ They make smart trade-offs. They are able to understand the risks of their decisions. They have failed I am often asked for career advice. One of the things before and so can avoid mistakes. software engineers always want to know is if they If there is a library or prebuilt code should learn some new tool or language. In fact, I somewhere, they are probably aware of it—they may even have contributed cannot think of a performance review I have read for a to it or used it in past projects. software developer that didn’t include something about ˲˲ They help others. Teammates go growing their skills around a particular technology. to them with questions because they have done this before and know how That is the nature of our work: it is constantly to do it right. This expertise makes changing. You have to keep learning, or you will them natural leaders or mentors. ˲˲ They hit their deadlines. Their esti- become obsolete. mates are almost always accurate. They But for your career, is it better to go wide and learn know how long something should take a lot of different things, or is it better to go deep and them because they have experience working on similar projects. learn a few things really well? ˲˲ They are still growing person- Making the case for going deep. Recently, I was ally and professionally. Since they are comfortable with the technology talking to another engineering leader about hiring and stack, they can use their time on learn- staffing. I asked which technologies he wanted ing skills in other areas such as lead-

44 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 ership, communication, or even new technologies. ˲˲ They have a deep understanding. When engineers work with a particu- lar technology for a long time, they learn the nooks and crannies of how it works—the good parts and the ugly parts. This can force them to think about the very constructs of how the technology was built. For example, discovering a bug in a foundational library teaches them to write better software. This depth also allows them to pick up other technologies faster. When you hire engineers who have experience in a particular technology and can operate at this level, they are able to get up to speed quickly. They will have to learn the way your partic- ular systems and software work, but they will not have to learn the tools and technology they were written in. While there is definitely value in gaining that deep experience—and it In these cases, it isn’t enough to you can be the best in the world at. Spe- should be part of your career plan—it know just one technology well; in fact, cialize in it. Pursue it above everything is also important to have breadth. if that is the case, you won’t be success- else. Success will come along with it. Making the case for going wide. ful. You have to have breadth and be In many ways this is true. When you When I look at my own career path, I able to think through other aspects of are truly exceptional at something, can say it is my breadth that has helped the problems you are solving, and you you build career capital, and you me the most. As a developer, the fact must have a broader understanding to can trade that capital for bigger pay- that I understood operations, lower- work with people in different roles. checks, more flexibility, or even fancy level operating systems, and compil- job titles.1 ers helped me write better code. And So Should You Go Wide or Deep? When people ask me the ques- as a manager, having experience with I once had the privilege of having a tion of where they should focus their other disciplines helps me work with mentoring session with a VP at one time—should I keep learning one those people better. of the great software companies. I technology or spend time learning a For example, we have all written asked him about his background and new one?—I ask them that very ques- software tests, but when you work what he felt got him to his position. tion: What is the one thing you could with really great testers and learn the He told me that he started his career be the best in the world at? way they think and approach prob- as a software engineer and soon real- The answer might be going deep or lems, this helps you not only work ized that being the best software en- going wide—the important thing is to better alongside them, but also write gineer would be hard; software test- spend your time on building the skills more robust code. ing, on the other hand, was much less that will move you to where to you As a technology executive, you can- competitive, and a lot of people didn’t want to go. not just understand the technology— have the passion for it that he did. you also have to understand the busi- He changed his career path to pur- Reference 1. Shin, L. 7 steps to developing career capital—and ness. You have to focus on customers, sue software testing and went on to achieving success. Forbes; http://bit.ly/2wgoieh. think about product, and be able to write several books. He became one understand the financial implications of the best in the world at testing and Kate Matsudaira (katemats.com) is the founder of her own company, Popforms. Previously she worked at of your decisions. Often, you are man- was asked to speak at conferences all Microsoft and Amazon as well as startups like Decide, aging people doing a job you have nev- over the world. The VP title and re- Moz, and Delve Networks. er done yourself—but you still need to sponsibility came later but was a natu-

be able to measure their impact and ral progression. Copyright held by author.

IMAGE BY LIGHTSPRING BY IMAGE mentor them to success. His advice to me: pick something Publication rights licensed to ACM. $15.00.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 45 practice

DOI:10.1145/3106637 such as Scrum and SAFe (Scaled Agile Article development led by queue.acm.org Framework) may be part of the future, but you will also need many new meth- ods and practices—some of which are Essence can keep software development not even known today. Extending a for the IoT from becoming unwieldy. single method to incorporate all that is required would result in something BY IVAR JACOBSON, IAN SPENCE, AND PAN-WEI NG that is way too big and unwieldy. In- stead, the new Object Management Group (OMG) standard Essence can be used to describe modular prac- tices that can be composed together Is There to form a multitude of methods, not only to provide for all of today’s needs, but also to be prepared for whatever the future may bring. a Single The software world is continu- ously innovating and opening up new areas of opportunity and challenge. A decade ago developers were busy with trends such as service-oriented Method for architecture and product-line archi- tecture—still very much around, but now a commoditized part of a larger system-of-systems landscape, and also the Internet extended to cloud computing with big data and mobile applications. New software development approaches have accompanied these new trends, of Things? most of them being agile in different flavors and size: Scrum, Kanban, DAD (Disciplined Agile Delivery), SAFe, LeSS (Large-scale Scrum), and SPS (Scaled Professional Scrum) being among these approaches. These trends have impacted the The Industrial Internet Consortium predicts the Internet software industry in many different ways—producing more pervasive and of Things (IoT) will become the third technological powerful technology-based products, revolution after the Industrial Revolution and the for example. None of them, however, Internet Revolution. Its impact across all industries has had a truly transformational or radically disruptive impact. and businesses can hardly be imagined. Existing The Industrial Revolution in the software (business, telecom, aerospace, defense, 19th century moved us from essentially building things as a craft to manufac- among others) will likely be modified or redesigned, and turing. The Internet Revolution at the a huge amount of new software, solving new problems, end of the 20th century was another such will be developed. As a consequence, the software transformation of the world or, as Bill Gates said in 1999, “A fundamental new industry should welcome new and better methods. rule for business is that the Internet This article makes the case that to be a major player changes everything.” The Internet has driven the need for faster turnaround in this space you will need a multitude of methods, not time with less precise requirements— just a single one. Existing popular approaches hence, sparking the trend toward light-

46 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 weight, empirical, and iterative meth- doesn’t know what happens to it once pact on the way we live and work, re- ods. It has also driven the rise of social it has left the factory gate. By connect- ducing waste and inefficiencies and networking, which places the Internet ing the product via sensors to the IoT, delivering major social and environ- at the heart of everyone’s life. the manufacturer can fundamentally mental benefits in security, health Now the Industrial Internet Consor- change its value proposition. Instead care, transportation and logistics, tium3 claims that the IoT, building on of selling only assets, the manufac- education and energy, amongst many the cloud, mobile Internet, big data, turer can sell services, including the other sectors of the economy.”13 The and so on, is a third such fundamental assets, which enables it to build long- IoT will eventually reach all areas transformation. The Industrial Inter- term relationships with its custom- where humans are providing prod- net Consortium was founded in March ers. For example, suppliers of aircraft ucts or services, both today and in the 2014 by ATT, Cisco, General Electric, engines can offer their products as future. Moreover, it will use all of the IBM, and Intel to remove roadblocks a service (the fee based on the num- kinds of systems in use today: com- to widespread adoption of the IoT. Its ber of flying hours). The effect is that munication, mobile, distributed, big mission is “to accelerate growth of the the supplier is now highly motivated data, cloud computing, among oth- Industrial Internet by coordinating to keep the machines running, since ers, and it will drive new technologies ecosystem initiatives to connect and otherwise it will lose revenue, and the not yet seen. integrate objects with people, process- airline can increase its revenue since To be successful, companies will es and data using common architec- it will have reduced downtime. There need to be able to respond quickly to tures, interoperability and open stan- are many similar examples from both the changing demands of the network dards that lead to transformational government and industry. Basically, while maintaining appropriate levels of business outcomes.” every industry will be affected, includ- engineering discipline, particularly for The IoT touches everything. What ing banking, insurance, telecoms, air- the cloud-based services upon which is it then about the IoT that will dra- lines, and defense. the distributed devices will depend. matically change the business model As voiced by Alex Sinclair, CTO of Moreover, the space to be addressed for all industries? Here is an exam- the GSMA: “We believe that with the covers all levels of complexity—from ple: Traditionally, a company sells a right standards and regulation in very simple software running on basic

IMAGE BY CONCEPT CAFE/SHUTTERSTOCK CONCEPT BY IMAGE product and, as long as all goes well, place it will have a fundamental im- sensors and other simple devices to

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 47 practice

the high-performance, highly reliable, The Ignite IoT Methodology. Ig- (AIA). The intention is that the PD highly governed, secure, resilient, scal- nite is an enterprise methodology for practice should be used to conduct able systems needed to process, ana- a major player in the IoT. It is a “big project self-assessments, compare lyze, and respond to the vast amounts method” covering all aspects of de- different IoT options, and select the of data they produce, and everything veloping for the IoT. It has two major solution architecture and technolo- in between. Not only that, the rate of practice areas. (In this article, practice gies to be used in a project. The AIA change and the need for innovation is defined as a repeatable approach practice is then used to identify the will never have been higher. to doing something with a specific devices, gateways, and services, and purpose in mind.9 Practices are the their responsibilities for an enter- The IoT Needs Everything things that practitioners actually do.) prise solution. Ignite provides a set The IoT does not lack methods. Re- These areas are strategy execution and of technology patterns (such as ma- searching the space shows clearly, solution delivery. Strategy execution is chine-to-machine connectivity, and and not surprisingly, that there is about agreeing what to build (that is, sensor networks, among other). not a one-size-fits-all approach. In- the solution) and involves the prac- The benefit of Ignite is that it is stead, methods for waterfall and Ag- tices of opportunity identification, based on real-world experience, captur- ile, methods for small applications opportunity management, and initia- ing this experience and best practice in (apps) and for complex systems of tion. Solution delivery is about deliv- a well-thought-out and comprehen- systems, and methods for systems ering the solution to users, and it has sive methodology. Naturally, the first engineering (that is, for systems with a life cycle consisting of planning, thought of the authors of the method- hardware and software integrated) building, and running (that is, oper- ology was not so much about the mod- are all still needed. What is really ating the solution). Planning involves ularity of the practices described but new is that a larger vendor needs all project initiation, whereas building about the completeness and relevance this at the same time and with com- and running are carried out through of the method as a whole. pressed time scales, which increases parallel project workstreams. The IoT Methodology. In compari- complexity significantly. Thus, for Project initiation is a set of prac- son, the IoT Methodology2 is a light- larger vendors a multitude of meth- tices that results in a number of dif- weight method highly inspired by lean ods are needed. A smaller vendor ferent artifacts, including solution startup12 and design thinking.1 It in- needs a more specific and focused sketches, a milestone plan, user inter- volves the following iterative steps: approach, but one that can grow as face mockups, and software architec- 1. Co-create. Communicate with end new products evolve and new prob- ture. Project workstreams consist of a users and stakeholders to identify pain lems emerge. Thus, methods such as complementary set of practices (called problem areas in a nontechnical way. Rational Unified Process (RUP) and workstreams): project management, 2. Ideate. Simplify discussions to com- SAFe, and practices such as Scrum, cross-cutting, solution infrastructure municate requirements to designers, user stories, and use cases are all be- and operations, back-end services, implementers, and project managers. ing applied. As always with any new communication services, on-asset 3. Question and answer. Trans- trend, new branded methods are components, and asset preparation. late soft concepts into hard require- born. Literature regarding methods At a high level, these might seem to ments, analyze solutions, and brain- for the IoT is extremely sparse at the all be very general practices, but em- storm options. time of this writing. We have found bedded within are two domain-specif- 4. Map IoT OSI (Open Systems Inter- two methods within the domain: Ig- ic practices: project dimensions (PD) connection). Map requirements to nite13 and the IoT Methodology.2 and asset-integration architecture a valid architecture, infrastructure, and business frameworks, similar to Figure 1. An abundance of practices. the layered approach used in the ISO/ OSI model. 5. Prototype. Use standardized tool- Unified Model- Comp’s Comp’s kits to build prototypes and iterate to- Scrum-of Use-Cases Process for Service Driven for ••• for ward minimal viable products. Lifecycle -Scrums Arch Re-Use .Net Def’n 6. Deploy continuously to close the feedback loop and improve the products. Iterative Scrum Use-Case User Architecture Component Test-Driven Like Ignite, this seems to be a very Essentials Essentials Essentials Stories Essentials Essentials ••• Develop’t generic method at the high level. What’s so special about IoT Method- QA Process Agile Team Measurem’t ology is its use of an IoT Canvas and PSP Essentials Essentials Modeling Essentials ••• Essentials an IoT OSI reference architecture. The IoT Canvas is an adaptation of Org the business model/lean canvas used Practice Essential Distributed Virtual Process ••• in brainstorming sessions to validate Harvesting Imp UML Team Team minimal viable product requirements for IoT projects. The IoT OSI refer- ence model is an adaptation of the

48 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice seven-layer ISO/OSI reference model of a practice, which consists of the for use with IoT solutions. This IoT things to work with, things to do, and OSI reference model consists of five competencies and patterns to provide layers, with endpoints at the bottom, minimal explicit guidance to apply connectivity, middleware, IoT ser- the practice effectively. This does not vices, and, finally, applications at the Essence provides just make the practices more acces- top. Stakeholders and developers use a common sible, but it also makes them easier to the IoT Canvas and IoT OSI reference learn, change, and use for teams that model to co-create and co-evolve a so- framework for adopt them. Later, we look at how one lution definition before prototyping. describing all of them—the Ignite Methodology— The IoT Methodology has taken ag- could be essentialized. ile thinking as a starting point but is practices and then also a monolithic method. The IoT Needs Essence composing them As discussed previously, the IoT re- New Practices Are Needed into many methods. quires many methods and practices, It is clear from these methods, and some of them specific to the domain our own experience handling emerg- and others that are generally accepted ing technologies, that new domain- good software development practices. specific practices will be needed to For example, they need to deal with handle the very nature of the IoT— specific problems of distribution and particularly practices to handle mobility, yet at the same time they these concerns: must be grounded in sound architec- ˲˲ Distributed. These systems are ture practices. typically far more distributed than Essence and practices. The soft- most other software systems. Experi- ware development world has already ence from the development of telecom- identified and described hundreds munication systems will come into of different practices, some of which play: new failure modes (due to com- are shown in Figure 1. Those shaded munications), reliability engineering, in green are selected for an IoT team. redundant systems development, and In an ideal world teams would be able so on. to select the set of practices they need ˲˲ Mobile. Again telecommunica- to address their current situation and tion vendors have practices to develop easily assemble them into a method. mobile systems, which are applicable. For example, a team building software For example, these systems have to for the IoT with a high level of engi- degrade gracefully, security is critical, neering complexity and a high rate and they must be robust. of change may choose to base their ˲˲ Human out-of-the-loop. The whole method on the practices highlighted idea of the IoT is to sense/analyze/acti- in green using Use Case and Archi- vate without a human in the loop—for tectural Essentials to provide the re- example, self-driving cars, automat- quired engineering rigor, and Scrum ed trading systems, and population and Agile Modeling to cope with the health integration systems. There may high rates of change. be practices to be designed here, The problem is these practices around reliability, failure manage- come from different sources and do ment/failover, and exception condi- not share the common ground needed tion management. to allow them to be readily composed What isn’t needed are new manage- into an effective method. This isn’t a ment practices. problem unique to the IoT; it is a prob- Both Ignite and IoT Methodology lem that has been plaguing the soft- are monolithic methods that reuse ware industry since its inception and many existing generic practices, com- one that gets worse with every advance bining these with new innovative prac- in technology. tices specifically for the IoT—sadly, in How can teams be empowered to a way that makes the new practices own and control their methods while difficult to reuse and share. This issue providing them with the guidance can be easily fixed, however, by taking they need to be successful, and reflect- them to the next level by essentializ- ing the owning organization’s need ing them and freeing their practices. for governance and compliance? How This means capturing the essence can teams benefit from the growing

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 49 practice

number of proven practices while con- standard Essence,9 which provides a Composing practices into methods. tinuing to innovate and rise to the new foundation that allows teams to share In the past, different methods have pri- challenges that they face every day? and free the practices from the shack- marily been described as isolated, con- These are issues that particularly affect les of monolithic methods. ceptual islands. Every method is basi- companies moving into the IoT, as they Essence provides the following: cally a unique phenomenon, described will need a variety of methods. ˲˲ A kernel of elements that estab- in its own language and vocabulary and What is needed is some concrete lishes a common ground for carrying not standing on any widely accepted common ground that the practices out software engineering endeavors common ground. Any method, howev- can share, providing both a shared vo- and assembling methods er, may be considered to be composed cabulary for practice definition and a ˲˲ A simple, easy-to-understand, vi- from a number of practices. framework for the assembly and anal- sual, intuitive language for describ- For example, the agile method ysis of methods. ing practices that can be used both to of extreme programming (XP) is This will allow organizations to represent the kernel and to describe described as having 12 practices, prepare a library of practices suitable practices and methods in terms of including pair programming, test- for their industry/domain—practices the kernel driven development, and continu- that teams can easily share, adapt, By combining these capabilities, Es- ous integration. Scrum, on the other and plug and play to create the inno- sence provides a common framework hand, introduces practices such as vative ways of working that they need for describing all practices and then maintaining a backlog, daily scrums, to excel and improve. composing them into many methods. and sprints. Scrum is not really a com- This common ground has already The power of Essence in address- plete method, though; it is a compos- been prepared in the form of the Es- ing the method complexity inherent in ite practice built from a number of sence kernel, part of the new OMG developing software for the IoT comes other practices designed to work to- from its ability to enable the composi- gether. Scrum can itself be composed Figure 2. The essence practice architecture. tion of practices into methods; help with other practices from, say, XP, clearly define life cycles and check- to form the method used by an agile Domain-Specific Practices points, enabling practice-independent team. This composition is typically governance; and support the creation done tacitly, as Scrum and XP are not Generic Practices of practice libraries from which prac- provided in a format that allows them tices can be selected to be composed to be explicitly composed. Essence Kernel: Method Agnostic into methods. As discussed previously, Essence Let’s now look at each of these in provides a framework and language for more detail. describing and composing practices. This framework provides a practice ar- Figure 3. Three teams sharing a simple practice library. chitecture where, as shown in Figure 2, both generic and domain-specific Shared Practice Library practices are described and assembled on top of the Essence kernel. Use Case Component Scrum Kanban Now individual practices can be de- scribed using Essence. A practice can be expressed by extending the kernel with practice-specific elements, by describing the activities used to prog- TDD ress the work and the work products produced, and by describing the spe- Kernel Architecture Iterative Test Driven User Story cific competencies needed to carry out Development these activities. Liberating practices in this way is Team A Team B Team C very powerful. Once practices are codi- Use Use Case Kanban fied in Essence, teams can take owner- Case Architecture Kernel Kernel Component ship of their way of working and start to assemble their own methods. This can start with even a simple library of prac- tices, as shown in Figure 3. TDD This capturing and sharing of prac- tices, both generic and domain-specif- Test Driven Kernel ic, in a way that lets them be applied Component Iterative Development User Scrum alongside popular management prac- Story tices (agile or otherwise), provides the raw materials that teams need to com- pose their own ways of working.

50 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice

Bringing a set of practices into this common system also allows gaps and overlaps to be more easily identified. The gaps can then be filled with ad- Introducing Essence ditional practices and the overlaps re- The new Object Management Group (OMG) standard Essence9 is designed to support solved by connecting the overlapping organizations and communities in becoming learning organizations with empowered teams that own their own ways of working and share their practices. practices together appropriately. In addition to liberating the practices by enabling them to play well together, Governance and compliance. The Essence does the following: Essence kernel allows you to define ˲˲ Makes methods significantly lighter by focusing on the essentials. life cycles easily in a practice-inde- ˲˲ Helps teams measure progress and health in a method-independent way. ˲˲ Allows organizations to build a library of practices from which teams can select the pendent way. Having a selection of ones needed for a particular solution (some teams need a “big” method, while others different life cycles is incredibly use- need only a small one). ful when tackling a domain as com- ˲˲ Helps organizations build “forever” learning organizations. plex as the IoT—particularly when Essence provides a foundation for software engineering methods. This foundation helps in two ways: enables teams to understand and visualize the progress and health the life cycles can be combined with of their endeavors, regardless of their ways of working; and, allows teams to easily whichever set of practices the team share, adapt, and plug and play their practices to create the innovative ways of working wants to use, ensuring that appro- that they need to excel and continuously improve.6,7 It guides developers in achieving measurable results and reusing their knowledge in priate governance is applied without systematic ways. compromising the other aspects of It helps executives lead programs and projects in balanced ways, without more the team’s way of working. governance than necessary, and develop learning organizations. Using the Essence kernel makes it very easy to assemble a number of life cycles, each built using the same Note that Essence is generic enough Parallel to these efforts, existing building blocks but addressing a dif- to support a waterfall life cycle, as well methods such as dynamic systems de- ferent context and containing its own as agile approaches. velopment method (DSDM) and the contextualized checkpoints. For ex- Building a practice library. It is easy Unified Process are being essential- ample, Munich Re11 defined a family to see how the use of Essence would ized.8,10 An essentialized method is of life cycles, each addressing a differ- readily allow the assembly of a com- first structured in terms of its inher- ent context: prehensive practice library contain- ited practices, and then each practice ˲˲ Exploratory: A lightweight agile de- ing all the practices needed for a par- is essentialized without changing its velopment life cycle for experiments, ticular domain in a way that empowers original idea. proof of concept, and small creative teams to select just the practices they All of these practices are built on top endeavors need to build their methods. Over the of the kernel and can be assembled to ˲˲ Feature growth: A rigorous engi- past few years, working in many ar- prime the pump for the methods that neering life cycle to support rapid fea- eas of software development, includ- your teams will use. For example, or- ture growth with a strong architectural ing embedded systems, financial sys- ganizations have used these practices foundation tems, telecommunications, modems, to create lightweight agile methods, ˲˲ Maintenance and small enhance- and many other areas affected by the robust software engineering methods, ments: A lightweight life cycle to enable IoT, Ivar Jacobson International (IJI; pull-based flow methods, and flexible the continuous flow of small en- https://practicelibrary.ivarjacobson.com/ method families. They have been used hancements and bug fixes for a start has built an Essence-based prac- to create both agile and waterfall meth- fixed, funded period of time (typi- tice library). Its library caters to both ods that share many of the same prac- cally a year) the craft and engineering ends of the tices but apply them with a very differ- ˲˲ Support: A support-focused life development spectrum. ent emphasis. cycle to aid in the transition be- The practice library is constantly What is powerful here is that these tween the development and support evolving as more and more practices methods all share the same founda- organizations are captured in the Essence language. tion and can adapt to changing cir- The ability to capture checkpoints At press time, IJI has essentialized cumstances by dropping and adding and life cycles in a practice-indepen- close to 30 practices, including: practices. The methods can also share dent way is incredibly powerful. It lib- ˲˲ Agile essentials, such as daily practices, helping the teams—and the erates the practices, allowing them stand-ups, product ownership, and ag- software they produce—to align and to be used where appropriate and not ile retrospectives, collaborate with one another. constraining them to any predefined ˲˲ Common agile practices such as To make the practices accessible type or style of development. It also Scrum, user stories, and continuous flow, and easy to learn, they are all available makes it possible to address the entire ˲˲ Proven architectural practices in card and electronic formats. Easy-to- IoT methods space with a minimal, ex- such as Use-Case 2.0, architectural es- use tools are available for practice and tensible, evolving set of practices, and sentials, and component-based devel- card creation, for method composi- allows teams to get the help they need opment, and tion and publication, and for practice without compromising their agility or ˲˲ Life cycles such as the ones defined exchange and community building. engineering rigor. by Munich Re, discussed earlier. These tools make it easy to extend ex-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 51 practice

isting practices to meet your needs and plications, with their high distribution local standards, add your own prac- and ubiquity, require serious attention tices, define practice-independent life to architecture. cycles, and build your own frameworks At the kernel layer, Essence provides and methods. guidelines for working with the soft- This allows you to leverage not just The IoT will ware system. IJI’s generic library has a the industry best practice captured in eventually reach practice for working with architecture, the IJI practices, but also your own best including guidelines for creating a practices, be they technical, financial, all areas where sound architecture description (a work motivational, or managerial. humans are product) in an agile and lightweight manner. The Ignite method recom- Building a Practice providing products mends using AIA as a way to describe Library for the IoT architecture, and IoT Methodology Examining the practices found in Ignite or services, recommends using its IoT OSI model. helps illustrate how to add domain-spe- both today and An application that uses EPC and REST cific practices to a practice library. would have technology specifics about Ignite describes a number of IoT- in the future. how to name products and connec- specific practices, including the AIA tions and so on. practice discussed earlier. Today, the Let’s dive into the practices identi- generic practices in Ignite are not de- fied in Figure 4. The Essence language scribed in any detail, a gap that can specifies a number of constructs. For easily be addressed by reusing the brevity, this article illustrates only al- generic practices available in the IJI phas and work product. An alpha is “an practice library. essential element that is relevant to an Essentializing Ignite in this way assessment of the progress and health helps distinguish the IoT-specific prac- of a software engineering endeavor.”9 tices in a way that allows them to be ad- The alphas provide descriptions of the opted separately and applied alongside kinds of things that a team will man- whatever generic practices the team or age, produce, and use in the process of commissioning organization deems to developing, maintaining, and support- be the most appropriate. ing software and, as such, are relevant New domain-specific practices. By to assessing the progress and health of their very nature, the practices in the a software endeavor. “A work product is IJI practice library are very generic an artifact of value and relevance for a and applicable to many software-de- software engineering endeavor. A work velopment domains. These generic product may be a document or a piece practices are useful for many kinds of of software.”3 Practices are a kind of software, including for the Internet package consisting of these elements. of Things. The Essence kernel, which stands The specific practices from Ignite at the bottom of Figure 4, is made up and IoT Methodology are useful do- of a number of elements. The figure main-specific practices that help ad- specifically shows the Software Sys- dress specific challenges for IoT ap- tem alpha. The Essence kernel does plications. In addition, practitioners not have an explicit notion of archi- would have to work with specific tech- tecture because in simple develop- nologies such as EPC (Electronic Prod- ment, this is left for teams to define. uct Code) to identify smart objects over For more sophisticated development, an RFID network communicating with the architecture practice fills the gap REST (representational state transfer) by providing explicit guidance on cre- interfaces.5 Thus, there would be other ating an intentional architecture. The domain-specific practices to use EPC architecture practice introduces an and REST correctly. Architecture alpha that is described Let’s take a peek at how domain-spe- by an architecture description work cific practices are added to the practice product. The Architecture alpha pro- architecture. A method has many as- vides guidance on how to determine pects, such as team collaboration, how architecture goals and how to identify to manage requirements, architecture, and validate architecture scenarios. and so on. In the discussion to follow, The two domain-specific practic- as shown in Figure 4, the focus is on es—namely, AIA practice and IoT OSI architecture aspects because IoT ap- practice—provide specializations on

52 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 practice

Figure 4. Architecture practices in Ignite and IoT Methodology. to adopt. Instead, the focus should be on essentialized practices that provide Work an incremental and safe path for teams Products Layer Practices Alphas and organizations to evolve and grow

(IoT) Domain EPC/ their ways of working. Specific Practices EPC/REST REST By using Essence as the foundation based IoT Pattern for a new practice library, we can lib- (extends) (extends) (extends) erate the practices and provide devel- opment teams with the guidance they Asset-Integration- IoT OSI AIA IoT OSI need to innovate, improvise, and excel. Architecture (AIA) Model Pattern Pattern We can avoid the traps of the past and enable software-engineering methods (extends) (extends) (extends) to evolve at Internet speeds while build- Generic ing on established, proven practices. Practices Architecture 1 Practice References (Architecture Views) 1. Brown, T. Design thinking. Harvard Business Review Architecture 86, 6 (2008), 84. 1 Description 2. Collins, T. A methodology for building the Internet of Architecture Things; http://www.iotmethodology.com/ 3. Evans, P.C., Annunziata, M. Industrial Internet: Essence Pushing the boundaries of minds and machines. Kernel GE, 2012; www.ge.com/docs/chapters/Industrial_ Kernel Software System Internet.pdf. 4. Fontoura, M., Pree, W., Rumpe, B. The UML Profile for Framework Architectures. Addison-Wesley Longman Publishing, 2000. 5. Guinard, D., Mueller, M., Pasquier-Rocha, J. Giving RFID a REST: Building a Web-enabled EPCIS. how an IoT application architecture is learn a domain quickly. This practice Internet of Things. IEEE, 2010, 1–8. 6. Jacobson, I., Ng, P.-W., McMahon, P. E., Spence, I., described. The way teams work on an separation is in contrast to monolithic Lidman, S. The Essence of software engineering: The IoT architecture is similar to the way methods where salient aspects of such SEMAT kernel. Commun. ACM 55, 12 (Dec. 2012); and acmqueue 10, 10; http://queue.acm.org/detail. they work on other kinds of architec- methods often drown in the sea of ge- cfm?id=2389616. tures. Thus, they do not introduce a neric information. It also helps prac- 7. Jacobson, I., Ng, P.-W., McMahon, P. E., Spence, I., Lidman, S. The Essence of Software Engineering: new Architecture alpha but reuse the titioners differentiate methods—for Applying the SEMAT Kernel. Addison-Wesley, 2013. Architecture alpha and description example, Ignite and IoT Methodol- 8. Jacobson, I., Ng, P.-W., Spence, I. The Essential Unified Process.Dr. Dobb’s Journal (Aug. 2006); http:// from the generic architecture prac- ogy—from the way they work with ar- www.drdobbs.com/architecture-and-design/the- tice. There are specific considerations essential-unified-process/191601687. chitecture and to understand if they 9. Object Management Group. Essence—Kernel and peculiar to IoT applications, however. are truly different. Practice separation language for software engineering methods, 2014; http://www.omg.org/spec/Essence/. Hence, each of these domain-specific also helps practitioners pick the best 10. Page, V., Stimson, R. Essentializing the DSDM Agile practices introduces a pattern for de- parts from different methods, pro- Project Framework. Agile Methods Conference, London, 2016. Ivar Jacobson International; https:// scribing an IoT application. A pattern vided they have been decomposed, www.ivarjacobson.com/sites/default/files/field_iji_file/ provides domain/technology-specific as shown in Figure 4. This mix-and- article/essentializingdsdm_1.pdf. 11. Perkens-Golomb, B., Folkjaer, P., Rauch, F., Spence, stereotypes to model the IoT applica- match approach helps teams become I. Ending method wars: The successful utilization of tion. In Unified Modeling Language innovative with methods, as well as Essence at Munich Re. Ivar Jacobson International, 2015; https://www.ivarjacobson.com/sites/default/ (UML) speak, this corresponds to a the solutions they produce. files/field_iji_file/article/essence_munichre_0.pdf. UML profile.4 UML profiles are a com- Thus, architecture is one area that 12. Ries, E. The Lean Startup: How Today’s Entrepreneurs Use Continuous Innovation to Create Radically mon approach to describe domain- needs special attention when building Successful Businesses. Random House, 2011. specific architectures, and IoT is one IoT applications. Security and privacy 13. Slama, D., Puhlmann, F., Morrish, J., Bhatnagar, R. Enterprise IoT: Strategies and Best Practices for such domain. The AIA practice intro- also need special consideration. The Connected Products and Services. O’Reilly, 2015. duces an AIA pattern for the architec- IoT opens the world to new ideas and ture description, whereas the IoT OSI use cases, and, as such, product idea Ivar Jacobson, chair of Ivar Jacobson International, is practice introduces an IoT OSI pattern. a father of components and component architecture, use generation and formulation also need cases, the Unified Modeling Language, and the Rational At the very top is a technology-specific special considerations. Each of these Unified Process. He has contributed to modern business architecture practice for EPC/REST- areas require generic practices and modeling and aspect-oriented software development. Ian Spence is CTO at Ivar Jacobson International and based IoT applications. This contains a domain-specific practices. the team leader for the development of the SEMAT kernel. specific pattern for EPC/REST.5 An experienced coach, he has introduced hundreds of The layering of practices helps Welcome to the Future projects to iterative and agile practices. Pan-Wei Ng coaches large-scale systems development practitioners understand what is The IoT promises a new dawn for all involving many millions of lines of code and hundreds truly different when working with sorts of industries, fundamentally of people per release, helping them transition to a lean and agile way of working, not forgetting to improve their IoT-based applications, as opposed changing the basics of everyday life. code and architecture and to test through use cases to a more general application. Under- Let’s make sure our software-engi- and aspects. standing this difference helps practi- neering practices do not get left be-

tioners quickly pinpoint the specifics hind. Let’s stop producing inflexible, Copyright held by ownes/authors. they need to be aware of and, hence, monolithic methods that are not easy Publication rights licensed to ACM. $15.00.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 53 contributed articles

DOI:10.1145/3139453 in 2015, can be controlled via Wi-Fi and Multiple computational cameras Bluetooth.8 And at the high-perfor- mance end of the camera market, RED can be assembled from a common set offers a modular camera with inter- of imaging components. changeable parts, including lenses, bat- tery packs, and broadcast modules.10 BY MAKOTO ODAMAKI AND SHREE K. NAYAR Although they provide some level of flex- ibility, such cameras are limited in the types and quality of images they are ac- tually able to produce. In the realm of research, Adams et Cambits: al.1 proposed a computational pho- tography platform called Franken- camera, including API, sensor inter- A Reconfigurable face, and image-processing unit. That system can be used to implement var- ious computational-imaging meth- ods. However, its hardware is relative- Camera System ly rigid, limiting the extent to which it can be reconfigured. Manakov et al.5 proposed a camera system that can accommodate different optical add- ons, including kaleidoscope-like im- aging to make optical copies of the captured image. Different filters are THE CAMERAS IN our phones and tablets have turned then used to produce high dynamic range (HDR), multispectral, polariza- us all into avid photographers, regularly using them to tion, and light-field images. The sys- capture special moments and document our lives. One tem provides some reconfigurability notable feature of camera phones is they are compact but is bulky and difficult to scale in terms of functionality. Finally, recon- and fully automatic, enabling us to point and shoot figurability is a well-explored topic in without having to adjust any settings. However, when the field of science education;16 for in- stance, Schweikardt and Gross14 de- we need to capture photos of high aesthetic quality, we veloped a related robot kit, including resort to more sophisticated DSLR cameras in which a blocks with multiple functionalities variety of lenses and flashes can be used they call Cublets. And littleBits Elec- tronics Inc. developed a modular interchangeably. This flexibility is important for electronic system for experiential spanning the entire range of real-world imaging scenarios, while enabling us to be more creative. key insights Many developers have sought to make these cameras ˽˽ Cambits includes a set of physical blocks for building computational cameras with even more flexible through both hardware and multiple functionalities, including high dynamic range, wide angle, panoramic, software. For example, Ricoh’s GXR camera has collage, kaleidoscopic, post-focus, light field, stereo imaging, and even a interchangeable lens units, each with a different type microscope. 12 of sensor. Some manufacturers make their cameras ˽˽ Blocks include sensors, actuators, more flexible through application program interfaces lenses, optical attachments, and light sources attached through magnets (APIs) developers then use to control various camera without screws or cables. parameters and create new image-processing tools. For ˽˽ The configuration of the blocks can be changed without rebooting any of the

example, Olympus’s Open Platform Camera, released related hardware or software. BERG ALEXANDER BY IMAGES

54 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 CREDIT TK Communications haspublishedfourdifferentcovers, eachonefeaturingadifferentCambitconfiguration. Cambit piecescanbeassembledtocreateadozendifferentimagingsystems.To thisassortment, celebrate NOVEMBER 2017| VOL. 60| NO. 11 | COMMUNICATIONS OFTHEACM 55 contributed articles

learning in which the modules can be System Architecture snapped together through a magnetic A number of Cambits attributes follow interface to create circuits with vari- the design of the Cambits hardware and ous functionalities.4 software architecture: Here, we present Cambits, a set of Mechanical and electrical connections. physical blocks that can be used to build Cambits is Each Cambits block is 40mm along at a variety of cameras with different func- a scalable system, least two of its three dimensions. We 3D tionalities. Blocks include sensors, ac- printed the chassis of each block, in- tuators, lenses, optical attachments, allowing users cluding sockets close to its corners de- and light sources, assembled with mag- to add new blocks signed to hold magnets; the magnets nets without screws or cables. When are used to attach blocks to each other, two blocks are attached, they are con- and computational as well as to mechanically align them, as nected electrically through spring-load- in Figure 2a; the alignment is aided by ed pins that carry power, data, and con- photography convex and concave bumps on the sur- trol signals. The host computer always algorithms to face of the chassis. The polarities of the knows the current configuration and magnets in each block are also chosen automatically provides a menu of imag- the existing set. such that it is not possible for the user to ing functionalities from which the user attach two blocks that are otherwise in- can choose. Cambits is a scalable sys- compatible. For instance, a lens block tem, allowing users to add new blocks cannot be directly attached to an actua- and computational photography algo- tor block. When two blocks are at- rithms to the current set. tached, a set of either four or six spring- loaded pins on one block (see Figure 2b) Concept is aligned and electrically connected to Figure 1a shows the set of blocks that contact pads on the other block. The make up Cambits. They come in a vari- system uses USB 2.0 for the data signal ety of colors, each indicating a specific and I2C for the control signal. function: white for base, red for image Tree structure with bucket brigade. sensor, blue for flash, green for actua- Each Cambits block has three types of tors and spacers, yellow for lenses, and pins for conveying power, data signals, orange and purple for optical attach- and control signals. The data signal ments. Figure 1b shows the host com- conveys image data from the sensor puter, which always knows the current block. The control signal conveys the Cambits configuration, using a suite of configuration data upstream and vari- computational photography algorithms ous commands (such as the actuator to produce a variety of images. The sys- block’s rotation parameters and the tem reflects a number of attributes: flash’s strobing parameters) down- Ease of assembly. The blocks are at- stream. tached using magnets, and the configu- These signals communicate through ration of blocks can be changed without a tree structure. The host device—the requiring a reboot of the hardware or root of the Cambits structure—provides software; electrical power to all blocks in the con- Self-identification. The host comput- figuration, detects the current configu- er can detect the system’s current con- ration of the entire tree structure, and figuration, information that is conveyed controls all blocks within the tree. This to the user through 3D visualization and design ensures each block is able to a menu of functionalities it can per- connect with multiple other blocks. Up- form; stream is defined as the direction to- Diverse functionality. Since there are ward the host device and downstream many types of blocks, many controlla- as the direction in which components ble by the user, a diverse set of camera proceed forward from the host device systems can be configured in which (see Figure 3). each is able to produce a different type The data signal flows upstream di- of image; and rectly from each sensor block to the Scalability. The design of the sys- host device. However, the control sig- tem’s hardware and software architec- nals are passed from component to ture makes it inherently scalable so new component in bucket-brigade fashion. blocks and computational photography Each block is able to communicate algorithms are added readily to the ex- through control signals with only the

isting set. blocks that are connected to it. When a BERG ALEXANDER BY IMAGE

56 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 contributed articles

Figure 1. Cambits overview: (a) Cambits components; (b) a Cambits configuration, with host computer displaying a 3D visualization of the current configuration and a menu of functionalities it can perform; and (c) Cambits blocks and their specifications.

(Q)

(F) (G) (M) (P)

(B) (C) (E) (H) (K) (O)

(A) (C) (D) (E) (I) (J) (L) (N) (a)

Block Type Specification Host Device Functionality Menu Base (A) Single (B) Dual Sensor (C) 1.3MP, 1288 x 964, 1/3” CCD, 15 fps Flash (D) 4 LEDs with controller, max. current 25mA/LED Actuator (E) Single-axis rotary actuator, 180° range Spacer (F) Right angle Lens (G) 12mm, F2.0, horizontal FOV 22.2° (H) 8mm, F2.0, horizontal FOV 33.8° (I) 4.3mm , F2.0, horizontal FOV 87.7° (J)1.3mm , fisheye, F2.8, horizontal FOV 180° (K) 16mm, F1.4, with piezoelectric linear actuator OpticalAttachment (L) Teleidoscope (M) Lens array, 7 lenslets (N) Warm Current Configuration Shoot Button (O) Soft focus (P) Polarization Microscope (Q) Objective lens, x1.45, with LED illumination (b) (c) block is attached to the system, it scans blocks have the same address, as is pos- unit (MCU) (Texas Instruments the components downstream. If it de- sible when using the I2C interface. MSP430F5510) with two serial ports for tects any blocks, it reads the configura- When the host device seeks to con- the bucket brigade. The controller tion data of the blocks that are down- trol a specific block in the tree structure, board has an upstream interface and a stream, adds its own identity and it sends its command and the address downstream interface (see Figure 4). address to the data, and then sends the of the block to the base block. The base When a block with the controller is information upstream. As a result, the block and subsequent blocks pass the attached to the system, it turns on auto- host device is able to detect the com- command downstream in bucket-bri- matically, thus triggering the firmware plete order of the configuration. If we gade fashion. The addressed block ulti- on its MCU to start scanning down- had instead used a conventional electri- mately receives the command and exe- stream for approximately 100msec to cal bus for the control signals, the sys- cutes it. communicate with its adjacent blocks. tem would not have been able to detect Controller board. A key aspect of the When the block is removed from the the order of the blocks. Moreover, a con- design is the controller board inside the system, it loses power, and the firmware ventional bus would not be able to de- base, actuator, spacer, and sensor stops. tect configurations in which multiple blocks. It includes a microcontroller Each block has a power circuit to pre-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 57 contributed articles

Figure 2. Cambits detachable connector: (a) mechanical assembly and alignment of blocks vent inrush current and voltage drop using magnets; and (b) electrical connection between blocks using spring-loaded pins that when attached, thus maintaining a carry power, data, and control signals. steady input voltage. Due to the circuit, the system can be reconfigured without requiring a reboot of the hardware (blocks) or the software running on the host computer. VCC Power The controller board can also control D+ Data signal other devices (such as the servo motor D - in the actuator block and the LED con- SCL Control signal SDA troller in the flash block) through the GND Power I2C bus, pulse-width modulation (PWM), and general-purpose input out- puts (GPIO) based on commands it re- ceives from the host device. For exam- (a) (b) ple, when an actuator block receives the command for rotation, the MCU gener- ates the pulse signal needed to drive its Figure 3. Tree architecture used to implement Cambits; power flows downstream, data servo motor. flows upstream, and control signals are communicated in bucket-brigade fashion. The I2C interface is also useful in terms of scalability because it is widely Base Spacer Actuator Sensor Lens Optical Attachment used in the field of embedded systems, allowing us to add various extra devices Hub Sensor Host (such as a light sensor, acoustic sensor, MCU MCU MCU MCU ID IR sensor, GPS, IMU, and multispectral Servo light source) to the Cambits set. Power MCU Lens blocks and optical attachments. Upstream Data signal LEDs Downstream The lens block includes an identifica- Control signal Flash tion board with I2C expander device that can detect the identification num- ber of the lens type itself and an addi- Figure 4. The base, actuator, spacer, and sensor blocks include a controller board that allows a block to communicate with its adjacent blocks. tional optical attachment connected to the lens (such as soft focus filter, lens array, and “teleidoscope,” or lens for Sensor Block Controller Board creating kaleidoscope-like images). The Microcontroller unit Downstream interface (6 pins) optical attachment includes no electri- Power circuit cal parts but does have up to three PWM bumps that push against mechanical Upstream interface (6 pins) Downstream interface (4 pins) switches on the lens block to generate a three-bit code the lens block can use to identify the attachment. The lens block GPIOs then sends this information upstream. Debug I/F Sensor Board Sensor block. The sensor block in- cludes a Point Grey camera board (BFLY-U3-13S2C-CS) that can produce 1.3-megapixel video in various formats

Figure 5. Cambits software architecture. (such as YUV411 and RGB8) and send the video upstream as a USB 2.0 data signal. In designing Cambits, we aimed to minimize the length of the data sig-

Computational Photography Processing nal line and the number of connectors so as to enable high-frequency Cambits API (480Mbps) transmission needed to pre- 15 Cinder serve the integrity of the video. Users are able to control various imaging pa- Serial Port Driver Point Grey SDK Open CV rameters of the sensor board (such as exposure time and gain) from the host device. Windows 8 Mechanical design. As mentioned, the Cambits blocks attach to each oth-

58 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 contributed articles

er through magnets invisible to the user, as they are embedded within the plastic (polylactic acid, or PLA) enclos- ing the blocks. To ensure the magnetic forces exerted through the PLA block enclosures are strong enough to keep the blocks attached, we used Neodymi- um block magnets in dimensions 3/8” x 1/4” x 1/16”. To ensure precise optical alignment between blocks, we de- signed the faces of the block covers with small mechanical bumps and in- dentations. The dimensions of the PLA enclosures of the blocks must be pre- cise enough to ensure that when lens and optical attachment blocks are at- tached to a sensor block, an image is achieved with the desired depth of field. Finally, as in Figure 4, the con- troller circuit board within each block includes not only the various electron- ic components but also the spring- loaded connectors used to electrically connect the block to the one to which it is attached. In our prototype, the PLA enclosure, magnets, and controller cir- cuit board add approximately 5.2mm in linear dimension to the compo- nent—sensor, actuator, and flash—in a block. Detailed mechanical design files are available at http://www.cs.co- lumbia.edu/CAVE/projects/cambits. Software. The software system that runs on the host device captures images from the Cambits system, giving users a 3D visualization of the current configu- ration and the option to apply various computational photography methods to the captured images. The current im- plementation runs on Windows and is based on open source libraries (such as Open CV-v2.4.10 and Cinder v1.20) (see use a base, a sensor block, and 8mm uses the image with the most corre- Figure 5). It also uses the Point Grey Fly- lens block. In HDR mode, the camera sponding features, with all remaining capture2 software interface to control captures six images with different expo- images as the center image of the col- the image sensors. sure times—where the exposures are lage, transforming the remaining imag- The Cambits API is also able to re- the geometric sequence t, 2t, 4t, 8t, 16t, es to align with the center image and ceive images from the Point Grey SDK and 32t seconds—computes an HDR overlay to obtain the collage.7 and control camera parameters (such as image using a triangular weighting Cambits also includes a variety of exposure time and gain) and various de- function,2 and then “tone maps” it us- lenses. For instance, a fisheye lens with vices on the tree architecture, including ing Reinhard’s algorithm.11 a focal length of 1.3mm and f-number servo motors, linear actuators, and Users can also move this basic cam- of 2.8 can be used to capture a wide-an- LEDs, through serial ports. The API and era around to capture a set of images gle image of a scene with horizontal the open libraries allow developers to that can be fused to obtain a scene col- field of view of 180 degrees. Since the add new blocks and image-processing lage. The system is able to detect the fea- host computer knows the type of lens algorithms to the system. tures in each image using the scale-in- being used, the software automatically variant feature transform (SIFT) maps the captured image to a perspec- Functionality algorithm, reduce the outliers using the tive without barrel distortion.13 We have used Cambits to assemble a random sample consensus (RANSAC) As with the lenses, users can attach a range of computational cameras (see algorithm, and find corresponding fea- variety of optical filters to the lens of the

IMAGE BY ALEXANDER BERG ALEXANDER BY IMAGE Figure 6). To construct a basic one, we tures between pairs of images. Cambits imaging system, including simple opti-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 59 contributed articles

Figure 6. Example results; for high-resolution versions, see http://www.cs.columbia.edu/CAVE/projects/cambits/.

(a) HDR (i) Microscope

(b) Collage

(c) Wide Angle (e) Light Field

(d) Refocusing (f) Kaleidoscope

(g) Stereo Panorama Anaglyph

(h) Depth from Stereo

60 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 contributed articles

cal filters like diffusion and polariza- “bright field” illuminate the sample. References tion, as well as more complex ones The user controls the LED in terms of 1. Adams, A. et al. The Frankencamera: An experimental platform for computational photography. ACM (such as a lens array and a teleidoscope). brightness through the host computer. Transactions on Graphics 29, 4 (July 2010), article 29. The Cambits lens-array attachment in- Alternatively, ambient illumination in 2. Debevec, P.E. and Malik. J. Recovering high-dynamic- range radiance maps from photographs. In Proceedings cludes seven acrylic ball lenses to pro- the environment can be used to back- of the ACM SIGGRAPH Conference (Los Angeles, CA, duce a 4D light-field image of the scene.3 light the sample. Aug. 11–15). ACM Press, New York, 2008, 31. 3. Georgiev, T. et al. Spatio-angular resolution trade-offs The teleidoscope attachment produces in integral photography. In Proceedings of the 17th a kaleidoscope image. An acrylic ball Conclusion Eurographics conference on Rendering Techniques (Nicosia, Cyprus, June 26–28). Eurographics lens in front of the attachment captures Cambits is a versatile modular imaging Association, Aire-la-Ville, Switzerland, 2006, 263–272. the scene image, and a set of first-sur- system that lets users create a range of 4. littleBits Electronics Inc., New York; http://littlebits.cc/ 5. Manakov, A. et al. A reconfigurable camera add-on for face planar mirrors between the ball computational cameras. The current high dynamic range, multispectral, polarization, and light-field imaging.ACM Transactions on Graphics 32, lens and the lens block creates multiple prototype is a proof of concept we use to 4 (July 2013), article 47. rotated copies of the image. demonstrate key aspects of Cambits: 6. Ng, R. et al. Light-field photography with a hand-held plenoptic camera. Computer Science Technical Report The focal stack lens block includes a ease of assembly, self-identification, CSTR 2, 11 (Apr. 2005), 1–11. linear actuator that physically sweeps and diverse functionality. We have thus 7. Nomura, Y., Zhang, Li, and Nayar, S.K. Scene collages and flexible camera arrays. InProceedings of the 18th the lens to capture a set of images cor- shown Cambits can be a powerful plat- Eurographics Conference on Rendering Techniques responding to different focus settings. form for computational photography, (Grenoble, France, June 25–27). Eurographics Association, Aire-la-Ville, Switzerland, 2007, 127–138. The linear actuator moves the lens in enabling users to express their creativity 8. Olympus Corporation, Tokyo, Japan; https://opc. steps of 0.05mm, with a total travel dis- along several dimensions. An impor- olympus-imaging.com/en_sdkdocs/index.html 9. Peleg, S. and Ben-Ezra, M. Stereo panorama with a tance up to 2.0mm, using a piezoelec- tant aspect of Cambits is that it is de- single camera. In Proceedings of the Conference on tric linear actuator to move the lens pre- signed to be an open platform that is Computer Vision and Pattern Recognition (Fort Collins, CO, June 23–25). IEEE Computer Society, Los cisely. The captured stack of images scalable. That design allows users to Alamitos, CA, 1999. helps compute an index map that repre- add multiple hardware blocks, includ- 10. RED Digital Cinema Camera Company, Lake Forest, CA; http://www.red.com/products sents the image in which each pixel is ing structured light sources, multispec- 11. Reinhard, E. Parameter estimation for photographic focused. The focal stack lens block then tral sources, telescopic optical attach- tone reproduction. Journal of Graphics Tools 7, 1 (Nov. 2002), 45–51. generates an interactive image that lets ments, and even non-imaging sensors 12. Ricoh Company, Ltd., Tokyo, Japan; https://www.ricoh. com/r_dc/gxr/ users click on any part of the image to for measuring acceleration, orienta- 13. Schneider, D., Schwalbe, E., and Maas, H.G. Validation bring it into focus.6,17 tion, sound, temperature, and pressure. of geometric models for fisheye lenses.Journal of Photogrammetry and Remote Sensing 64, 3 (May We designed Cambits so it would be We anticipate developing algorithms 2009), 259–266. possible to insert a rotary actuator be- that use such a diverse set of sensors to 14. Schweikardt, E. and Gross, M.D. roBlocks: A robotic construction kit for mathematics and science tween the base and the sensor to scan a trigger/control various image-capture- education. In Proceedings of the Eighth International panorama of a scene. If the camera is and-processing strategies. To encour- Conference on Multimodal Interfaces (Banff, Alberta, Canada, Nov. 2–4). ACM Press, New York, 2006, 72–75. rotated off-axis—with an offset between age others to modify or build on the cur- 15. USB Implementers Forum, Inc. High Speed USB the rotation axis and the center of pro- rent system, we have made the details of Platform Design Guidelines Rev. 1.0; http://www.usb. org/developers/docs/hs_usb_pdg_r1_0.pdf jection of the camera—users would be its hardware and software design avail- 16. Yim, M. et al. Modular self-reconfigurable robot able to take left and right image strips able at http://www.cs.columbia.edu/ systems: Grand challenges of robotics. IEEE Robotics &Automation Magazine 14, 1 (Apr. 2007), 43–52. from the captured sequence of images CAVE/projects/cambits/databases/ 17. Zhou, C., Miau, D., and Nayar, S.K. Focal Sweep Camera to generate a stereo panorama for creat- cambits_supporting_database.zip. for Space-Time Refocusing. Technical Report. Department of Computer Science, Columbia 9 ing virtual reality. In the example in Fig- University, New York, 2012; https://academiccommons. ure 6g, 120 images were taken while the Acknowledgments columbia.edu/catalog/ac:154873 actuator rotated 120 degrees and the We did this research at the Computer Makoto Odamaki ([email protected]) is an offset between the rotation axis and the Vision Laboratory of Columbia Uni- engineer of digital camera systems at Ricoh Company, center of projection of the camera lens versity in New York while Makoto Oda- Ltd., Tokyo, Japan. was 70mm. maki was a Visiting Scientist from Shree K. Nayar ([email protected]) is the T.C. Chang Professor of Computer Science at Columbia A second rotary actuator can be add- Ricoh Company, Ltd.; for the design University in New York where he also heads the Columbia ed to the system to configure a pan/tilt data covered here, see http://www. Vision Laboratory. camera system. cs.columbia.edu/CAVE/projects/cam- Copyright held by the authors. Cambits is not limited to a single im- bits. We thank William Miller for de- Publication rights licensed to ACM. $15.00. age sensor. Its second base can be used signing and 3D printing the chassis of with two sensor blocks and lenses to the Cambits blocks, Wentao Jiang for create a stereo camera system with a his contribution to the user interface, baseline of 44mm. Cambits processes and Daniel Sims for editing the dem- the left and right video streams from onstration video and formatting the this system in real time to produce a project webpage. Divyansh Agarwal, gray-coded-depth video of the scene. Ethan Benjamin, Jihan Li, Shengyi Cambits can also be used to assem- Lin, and Avinash Nair implemented ble a microscope that includes an objec- several of the computational-photog- Watch the authors discuss tive lens, a mechanism to adjust the raphy algorithms. The authors also their work in this exclusive Communications video. height of the sample slide to bring the thank Anne Fleming for proofreading https://cacm.acm.org/videos/ sample into focus, and an LED light to an early draft of the article. cambits

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 61 contributed articles

DOI:10.1145/3141771 (such as by Fu et al.,5 Galvis Carreño and The varying review dynamics seen in Winbladh,6 and Google Analytics7) and commercial efforts (such as Applause different app stores can help guide Analytics3) have proposed solutions to future app development strategies. help developers cope with large num- bers of reviews. BY STUART MCILROY, WEIYI SHANG, A 2013 study of reviews of iOS apps NASIR ALI, AND AHMED E. HASSAN by Pagano and Maalej20 found that on average a free app receives 37 reviews per day, while paid apps receive approx- imately seven reviews per day,20 and an- other study of iOS apps found that 50% of studied free apps receive only 50 re- User Reviews 11 views in their first year. Yet no prior research examined the reviews in the Google Play store, considering, say, “Is the data normally distributed or highly of Top Mobile skewed, with only a small number of apps receiving a substantial number of reviews on a daily basis?” Here, we explore the question of how Apps in Apple pervasive are the frequently reviewed apps in the Google Play store. In partic- ular, we empirically cover app reviews from the perspective of the developers and Google of the top apps there. Through an analy- sis of reviews for the top 10,713 apps in the Google Play store over a period of two months—January 1 to March 2, App Stores 2014—we found: More than 500 reviews daily. Only 0.19% of the studied apps received more than 500 reviews per day; Majority of studied apps. Almost 88% of the studied apps received only a small number (20 or fewer) reviews per day; ONE OF THE unique aspects of app stores is the and 13 Correlates with reviews. The number convenience of providing user feedback. Users can of downloads and releases correlated effortlessly leave a review and a rating for an app, with the number of received reviews, providing quick feedback for developers. Developers while the app category did not play a major role. are then better able to update their apps. This feedback Some of our observations differ from mechanism contrasts with traditional feedback other studies of user reviews of iOS mechanisms like bug-reporting systems (such as key insights Bugzilla), which are negative in nature, as they ˽˽ The characteristics of user reviews include only bugs, unlike reviews, which can be differ depending on app store. positive. Moreover, reviews can even serve as a means ˽˽ Few mobile apps in the Google Play store attract large numbers of user reviews. 7 for deriving additional app requirements. ˽˽ More app downloads and releases Developers of top apps might be overwhelmed by correlate with more reviews in the Google Play store, whereas app the large number of received reviews. Several papers category plays only a minor role.

62 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 apps,11 highlighting the need for addi- companies also provide developers buyers, finding reviews are a key deter- tional in-depth investigation of the re- overviews of user feedback and crash minant in their decisions to purchase viewing dynamics in both stores. reports. Google promotes its own ex- an app. A survey by Lim et al.16 found tensive analytics tools for Android reviews are one of the top reasons for Mobile App Analytics developers as a key competitive differ- users to choose an app. Likewise, Mu- A Vision Mobile survey of 7,000 devel- entiator relative to other mobile app dambi et al.18 showed that user reviews opers, also in 2014, found 40% of them stores. The tools measure how users have a significant effect on sales of on- made use of user-analytics tools and use an app (such as by identifying user line products. 18% used crash-reporting and bug- locations and how users reached the The importance of user reviews has tracking tools. Other studies also found app). They also track sales data (such motivated many studies, as well as our that developers need tools for app ana- as how developers generate revenue own work analyzing and summariz- lytics. For example, a 2013 study by Pa- through in-app purchases and the ing user reviews for mobile apps (see gano and Bruegge19 of how feedback effect of promotions on app sales2). Table 1). Iacob and Harrison12 built a occurs following initial release of a However, other than crash-reporting rule-based automated tool to extract software product identified the need to tools, many analytics tools today are feature requests from user reviews of structure and analyze feedback, particu- mostly sales-oriented rather than soft- mobile apps, an approach that identi- larly when it involves a large amount of ware-quality-oriented involving bugs, fies whether or not a user review con- feedback. performance, and reliability. tains a feature request. Chandy and Gu3 A number of app-analytics compa- Other studies have highlighted the identified spam reviews in the Apple nies, including App Annie,1 specialize effect of reviews of mobile apps on an (iOS) App Store, using a technique that in tools designed to help developers app’s success.9,15,19 Harman et al.9 found achieved high accuracy with both la- understand how users interact with a strong correlation between app rat- beled and unlabeled datasets. Carreño their apps, how developers can help ings and an app’s total download num- and Winbladh6 used opinion-mining generate revenue (such as through bers. User reviews include information techniques and topic modeling to suc- in-app purchases, e-commerce, and that could help developers improve the cessfully extract requirements changes direct buy), and how to leverage user quality of their apps and increase their from user reviews. Fu et al.5 introduced 15 IMAGE BY ANDRIJ BORYS ASSOCIATES/SHUTTERSTOCK ANDRIJ BORYS BY IMAGE demographics of the apps. These revenue. Kim et al. interviewed app an approach for discovering inconsis-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 63 contributed articles

tencies in apps, analyzing the negative terms of findings, methodologies, and We collected review information reviews of apps through topic analy- context, or Android vs. iOS (see Table 2). from 12,000 free-to-download apps sis to identify reasons for users liking from the Google Play store. From among or disliking a given app. Khalid et al.14 Studied Apps 30 different categories, including pho- manually analyzed and categorized one- Martin et al.17 noted that not all stores tography, sports, and education, we se- and two-star reviews, identifying the is- provide access to all their reviews, lead- lected the top apps in each category in sues (such as the hidden cost of using ing to biased findings when studying re- the U.S. based on app-analytic company an app) about which users complained. views. To avoid such bias, we collected Distimo’s (acquired by App Annie) rank- Chen et al.4 proposed the most exten- all reviews on a daily basis, ensuring we ing of apps for a total of 12,000; Distimo sive summarization approach to date, would include all available reviewers. ranked the top 400 apps for each of the removing uninformative reviews and However, the Google Play Store provides 30 categories. We used Distimo’s Spring prioritizing the most informative re- access to only the 500 latest reviews for 2013 list of top apps. Of the 12,000 top views before presenting a visualization an app. If more than 500 reviews are re- apps, 1,287 were not accessible during of the content of reviews. Guzman and ceived in the 24-hour period between our two-month crawl because some of Maalej8 performed natural language daily runs of our crawler, then the them might have been removed from processing techniques to identify app crawler does not collect those reviews. the store. We thus collected data from features in the reviews and leveraged This limitation means we thus offer a 10,713 top apps, with a total of 11,047 sentiment analysis to identify whether conservative estimate of the number of different releases during the studied users like such features. Our own work reviews for apps that receive more than time period. differs from these studies, as it aims to 500 reviews per 24-hour time period. Our own selection of top apps might provide context about when the other We based our Google Play store crawler have biased our results, possibly gener- techniques would be needed. on an open source crawler called the alizing to only the top, stable, free apps Pagano and Maalej20 and Hoon et Akdeniz Google Play crawler (https:// in the Google Play store. Nevertheless, al.11 analyzed the content of reviews of github.com/Akdeniz/google-playcrawl- we studied successful apps we felt were both free and paid apps in the Apple er) to extract app information (such as more likely to have a large user base and App Store, answering a similar research app name, user ratings, and reviews). receive a large number of reviews, rath- question as ours about the number of Running it meant we were simulating er than blindly study all apps. We chose received reviews, but there are major a mobile device over approximately two apps that had been popular one year be- differences between them and us in months—January 1 to March 2, 2014. fore we began our study because we were

Table 1. Our observations on Google Play apps compared to the Pagano and Maalej20 and Hoon et al.11 observations on the Apple (iOS) App Store.

In the Apple App Store, from Pagano In the Apple App Store, In the Google Play Item and Maalej21 from Hoon et al.11 Store, from us Notes Reviews received Average of 22 reviews Median of 50 reviews in Average of seven reviews We found fewer average and median user reviews per day, with 36.87 for first year for free apps and per day, with median of no compared to Pagano and Maalej21 and more user reviews free apps and 7.18 for 30 reviews in first year for reviews per day for free than Hoon et al.11 Reviews were skewed, with median paid apps paid apps apps number of received reviews at 0 and 88% of the studied apps receiving 20 reviews or fewer per day. Number of Facebook received 4,275 (not studied) Only 0.19% of apps Pagano and Maalej21 were the first to observe that some reviews received reviews in one day received more than 500 apps (for them, the Facebook app) might receive a large reviews, and the top 100 number of reviews per day. We were first to explore this most-reviewed apps had observation—apps receiving a large number of reviews 6,000 to 43,000 reviews per day—in depth, finding that while some apps might in the two-month study receive a large number of reviews, only 0.19% of all period. studied apps received more than 500 reviews per day. Most top apps might not benefit much from automated approaches that leverage sophisticated techniques (such as topic modeling) given the small number of reviews they received and their limited length. Effect of app Number of daily reviews Certain categories receive No relation Compared to both iOS studies, we found no relation category differs by category greater numbers of between an app’s category and number of received reviewers than others reviews, once we controlled for number of downloads and number of releases. Spike in number Number of reviews (not studied) The standard deviation of Both stores showed evidence of spikes in number of of reviews decreases over time received reviews deviates reviews immediately following a new release. decreases following a release from the median directly following following release and release returns back to normal afterward. Average length of Average of 106 Average of 117 characters Average of 64 characters Reviews in the Google Play Store were shorter than in the a review characters and median and median of 69 and median of 36 Apple App Store. Median length of reviews demonstrated of 61 characters characters characters that the distribution of review length is highly skewed, with long reviews as outliers.

64 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 contributed articles interested in stable, mature apps that Maalej20 and greater than Hoon et al.;11 We found fewer average reviews per had not been released within the past Finding 2. The number of user re- day than Pagano and Maalej20 possibly few months to avoid the expected burst views were skewed; similar findings due to any of several factors. The first of reviews following an app’s initial re- were reported by Pagano and Maalej;20 is we collected reviews from stable top lease.20 We focused on free-to-download and apps that had been released for at least apps, since recent work showed that Implication. Most top apps might one year, whereas Pagano and Maalej20 free apps receive five times as many re- not benefit much from automated ap- may have collected new apps and not views as paid apps.20 Moreover, over 90% proaches to analyzing reviews that le- focused on top apps. The second was of downloaded apps were, at the time, verage sophisticated techniques (such that our estimates for the frequently re- of the free-to-download variety, accord- as topic modeling) given the small num- viewed apps were conservative; we did ing to Gartner. Such apps use other rev- ber of received user reviews and their not count more than 500 reviews in a enue models (such as freemium, in-app limited length. day. For instance, Pagano and Maalej re- purchases, and ads). The developers of We plotted the number of reviews per ported that Facebook received 4,275 re- such apps are thus concerned about the day, as well as total number of received views in a day, with such large numbers effect of reviews on their revenue.9 reviews, using a beanplot combining increasing the overall reported average a boxplot with a kernel-density-esti- number of received reviews on a daily Findings mation function. Figure 1a reports the basis. We separated the apps into two Here, we present our findings, as in Ta- median number of reviews per day was groups: 100 most-reviewed apps and ble 2, concerning the reviews from the 0. We found 20, or 0.19%, of the 10,713 all other apps. Figure 1b reports there Google Play store while comparing our studied apps received 500 or more re- was a large gap in the total number of results with prior studies. views; as mentioned earlier, 500 would reviews among the 100 most-reviewed Number of received reviews. On be a conservative estimate, whereas 88% apps. The total number of reviews of the the number of received reviews in the of the apps in our 10,713-app dataset re- 100 most-reviewed apps ranged from Google Play Store ceived fewer than 20 reviews per day. Ad- 43,000 to 6,000 in the two-month study Finding 1. Most apps (88% of those ditionally, the median total number of period. The reviews themselves were of the 10,713 we studied) received few reviews was 0 during the study period. short, much shorter (approximately reviews during our studied time period. We also calculated the number of words 40%) than the reviews in the Apple App The average and median number of in each of the received reviews, with me- Store. We also observed a notable skew reviews were fewer than Pagano’s and dian number of words per review at 46. in the length of reviews in both stores. Influence of app category and down- Table 2. Datasets of prior work mining reviews of mobile apps. loads on number of reviews. In the Google Play Store Finding 3. The number of downloads Paper App Store Apps Reviews and releases correlated with the num- Iacob and Harrison12 Google Play Store 161 3,279 ber of received reviews, whereas an Galvis and Carreno7 Google Play Store 2 710 app’s category did not play a major role Fu et al.6 Google Play Store 171,493 13,286,706 during the study period. On the other Chen et al.5 Google Play Store 4 169,097 hand, Pagano and Maalej20 and Hoon Pagano and Maalej21 Apple App Store 1,100 1,126,453 et al.11 both reported a relation between Hoon et al.11 Apple App Store 17,000 8,700,000 an app’s category and the number of re- ceived reviews; and Implication. The relationship be- Figure 1. Beanplots showing number of reviews per day and in total. tween number of received reviews and an app’s category should be explored further, especially in light of the dis- crepancy between the two app stores. ws

y Here, we investigate the effect of an 600 vi e app’s number of downloads, number

40,000 of releases, and app category on the ws Per D a 400 number of received reviews. We built a vi e

umber of R e regression model with an app’s number R e

20,00 0 of received reviews as the dependent 200

Total N variable. Due to the notable skew in the number of reviews, we log-transformed 0 0 the number of reviews before building All other 100 most reviewed the linear-regression model. Mobile Apps Mobile Apps Figure 2 plots the total number of re- views using the built-regression model. (a) (b) We included three plots, each keeping the median values of the other factors

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 65 contributed articles

Figure 2. Plots of the total number of reviews (logged) on the y-axis and three separate the same so we could see how each fac- graphs of app categories, number of downloads, and number of releases on the x-axis; tor affects the total number of reviews.10 the graphs reflect the relation between the three factors and the total number of reviews. The gray bands around the plotted lines are bootstrap confidence intervals for categories downloads versions our estimates. We generated a nomogram (see Fig- 10 ure 3) to visualize the results of our re- gression model,10 helping us examine the effect of each factor while control- 8

ws) ling for other factors. The nomogram

vi e consists of a series of scales. The Lin- 6 ear Predictor scale is the total number of reviews in log scale. To calculate the G G G G G G G G total number of reviews, we can draw a 4 G G G

Log(Total R e G G G G G G G G G G G G G G G G straight line from the value of the “to- G G G tal points” scale to the linear predictor 2 scale. The total points are calculated by summing the points of each of the 0e+00

4e+08 scales of the three factors: releases, 3e+08 5e+08 2e+08 1e+08 Category Variables 2 4 6 8 10 12 downloads, and categories. To calcu- late the points value of each factor, we can draw a line from the value in the factor scale to the points scale. The Figure 3. A nomograph of the effect of new releases, app category, and number of downloads value in the points scale becomes the on total number of reviews received. points for that factor. For example, re- leases = 2, downloads = 100,000, and

0 10 20 30 40 50 60 70 80 90 100 categories = tools. We found that 2-re- Points leases corresponded to approximately seven points, 100,000-downloads cor- Versions 0 2 4 6 8 10 12 14 16 18 20 responded to approximately 20 points, and the tools category corresponded to 10,000 100,000 500,000 1,000,000 5,000,000 50,000,000 100,000,000 Downloads approximately five points. The sum was 5,000 32 total points, which corresponded to approximately 2.5 log scale, or 316 total Categories user reviews. We found that as the number of Total Points downloads and releases increased, the 0 20 40 60 80 100 120 140 160 total number of reviews also increased. Linear Predictor We found no relation between indi- 0 1 2 3 4 5 6 7 8 9 10 11 12 vidual categories (such as communica- tions, social, tools, and review count) when we controlled for number of Figure 4. Standard deviation of new reviews every 24 hours before and after the first downloads and releases. In contrast, collected release for each studied app; each boxplot represents the standard deviation 20 11 from the median number of reviews for each app at that time. Pagano and Maalej and Hoon et al. observed a relation between categories and number of received reviews in the 2.0 Apple App Store; however, neither study controlled for the other metrics in its analysis. Those studies observed a rela- 1.5 tion between categories and number of v. reviews that may be due to the interac- tion between categories and number of Std. D e 1.0 downloads or between categories and number of releases. Spike in reviews following a release. 0.5 Finally, concerning the spike in reviews following a release of an app in Google 96 −72 −48 −24 0 24 48 72 Play Store Hours Finding 4. Both the Google Play store and the Apple App Store showed evi-

66 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 contributed articles dence of a spike in reviews following a N. Why people hate your app: Making sense of user feedback in a mobile app store. In Proceedings of release; and the 19th ACM SIGKDD International Conference on Implication. Greater effort examining Knowledge Discovery and Data Mining (Chicago, IL, Aug. 11–14). ACM Press, New York, 2013, 1276–1284. user reviews should follow a release in 6. Galvis Carreño, L.V. and Winbladh, K. Analysis of user order to improve app quality. comments: An approach for software requirements evolution. In Proceedings of the 2013 International 20 Pagano and Maalej reported that Frequent releases Conference on Software Engineering (San Francisco, the number of received reviews de- CA, May 18–26). IEEE Press, Piscataway, NJ, 2013, ensure an app’s 582–591. creased over time after a release, sug- 7. Google Analytics; http://www.google.ca/analytics/mobile/ user base is 8. Guzman, E. and Maalej, W. How do users like this gesting releases contribute to new re- feature? A fine-grained sentiment analysis of app views. We observed the same kind of reviews. In Proceedings of the 22nd International more engaged Requirements Engineering Conference (Karlskrona, correlation for the Google Play store. Sweden, Aug. 25–29). IEEE Press, Piscataway, NJ, Figure 4 outlines a boxplot of the medi- as it begins 2014, 153–162. an number of reviews for each studied 9. Harman, M., Jia, Y., and Zhang, Y. App store mining providing feedback. and analysis: MSR for app stores. In Proceedings of app across each of its releases, showing the Ninth Working Conference on Mining Software Repositories (Zurich, Switzerland, June 2–3). a spike in reviews directly on and after Piscataway, NJ, 2012. an app’s release day. 10. Harrell, F.E. Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and However, still not clear is if these Survival Analysis. Springer, New York, 2001. spikes were due to an app attracting 11. Hoon, L., Vasa, R., Schneider, J.-G., Grundy, J. et al. An Analysis of the Mobile App Review Landscape: Trends and new users following its release or to cur- Implications. Technical Report. Swinburne University of rent users becoming more inclined to Technology, Faculty of Information and Communication Technologies, Melbourne, Australia, 2013. review the app. Looking closer at our 12. Iacob, C. and Harrison, R. Retrieving and analyzing nomogram, we note that many releases mobile apps feature requests from online reviews. In Proceedings of the 10th International Workshop on (more than 20) for an app has as much Mining Software Repositories (San Francisco, CA, May 18–19). IEEE Press, Piscataway, NJ, 2013, 41–44. of an effect as an app with 10 million 13. Johns, T. Replying to User Reviews on Google Play. downloads. Frequent releases thus en- Android Developers Blog, June 21, 2012; http:// android-developers.blogspot.ca/2012/06/replying-to- sure an app’s user base is more engaged user-reviews-on-google-play.html as it begins providing feedback. 14. Khalid, H., Shihab, E., Nagappan, M., and Hassan, A. What do mobile app users complain about? IEEE Software 32, 3 (May-June 2015), 70–77. Conclusion 15. Kim, H.-W., Lee, H.L., and Son, J.E. An exploratory study on the determinants of smartphone app A very small percentage of the top apps purchase. In Proceedings of the 11th International we studied (0.19% of 10,713) have ever DSI Decision Sciences Institute and 16th APDSI Asia Pacific Region of Decision Sciences Institute Joint received more than 500 reviews per day, Meeting (Taipei, Taiwan, July 12–16, 2011). yet most studied apps received only a 16. Lim, S.L., Bentley, P.J., Kanakam, N., Ishikawa, F., and Honiden, S. Investigating country differences few reviews per day. The number of re- in mobile app user behavior and challenges for ceived reviews for the studied apps did software engineering. IEEE Transactions on Software Engineering 41, 1 (Jan. 2015), 40–64. not vary due to the category to which the 17. Martin, W., Harman, M., Jia, Y., Sarro, F., and Zhang, Y. app belonged, varying instead based The app-sampling problem for app store mining. In Proceedings of the 12th Working Conference on Mining on number of downloads and releases. Software Repositories (Florence, Italy, May 16–17). IEEE Press, Piscataway, NJ, 2015. Some of our results highlight differenc- 18. Mudambi, S.M. and Schu, D. What makes a helpful es between the Google Play store and online review? A study of customer reviews on Amazon.com. MIS Quarterly 34, 1 (2010), 185–200. the Apple App Store. 19. Pagano, D. and Bruegge, B. User involvement Additional studies are needed to in software evolution practice: A case study. In Proceedings of the 2013 International Conference better understand the review dynamics on Software Engineering (San Francisco, May 18–26). across both stores. Researchers should IEEE Press, Piscataway, NJ, 2013, 953–962. 20. Pagano, D. and Maalej, W. User feedback in the App thus examine whether other empiri- Store: An empirical study. In Proceedings of the cal findings hold across them. In par- 21st IEEE International Requirements Engineering Conference (Rio de Janeiro, Brazil, July 15–19). IEEE, ticular, techniques designed to assist Piscataway, NJ, 2013. mobile-app developers should be opti- mized for each store. Stuart Mcilroy ([email protected]) is a Ph.D. student at Dalhousie University, Halifax, Canada.

References Weiyi Shang ([email protected]) is an assistant 1. App Annie Analytics; http://www.appannie.com/app- professor and Concordia University Research Chair store-analytics/ in Ultra-Large-Scale Systems in the Department of 2. Applause; https://www.applause.com/testing/ Computer Science and Software Engineering at Concordia 3. Chandy, R. and Gu, H. Identifying spam in the iOS app University, Montreal, Canada. store. In Proceedings of the Second Joint WICOW/ AIRWeb Workshop on Web Quality (Lyon, France, Apr. Nasir Ali ([email protected]) is an assistant research 16). ACM Press, New York, 2012, 56–59. professor at the University of Memphis, Memphis, TN. 4. Chen, N., Lin, J., Hoi, S,C.H., Xiao, X., and Zhang, B. Ahmed E. Hassan ([email protected]) is Canada AR-Miner: Mining informative reviews for developers Research Chair in Software Analytics and NSERC/ from the mobile app marketplace. In Proceedings BlackBerry Software Engineering Chair in the School of th of the 36 International Conference on Software Computing at Queen’s University, Kingston, Canada. Engineering (Hyderabad, India, May 31–June 7). ACM Press, New York, 2014, 767–778. 5. Fu, B., Lin, J., Li, L., Faloutsos, C., Hong, J., and Sadeh, © 2017 ACM 0001-0782/17/11

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 67 review articles

DOI:10.1145/3127874 There have been many recent excit- Healthcare robotics can provide health and ing examples of robotics technology, such as autonomous vehicles, package wellness support to billions of people. delivery drones, and robots that work side-by-side with skilled human work- BY LAUREL D. RIEK ers in factories. One of the most excit- ing areas where robotics has a tremen- dous potential to make an impact in our daily lives is in healthcare. An estimated 20% of the world’s Healthcare population experience difficulties with physical, cognitive, or sensory functioning, mental health, or be- havioral health. These experiences Robotics may be temporary or permanent, acute or chronic, and may change throughout one’s lifespan. Of these individuals, 190 million ex- perience severe difficulties with ac- tivities of daily living tasks (ADL).a These include physical tasks (basic ADLs), such as grooming, feeding, and mobility, to cognitive func- tioning tasks (instrumental ADLs), THE USE OF robots in healthcare represents an exciting which include goal-directed tasks opportunity to help a large number of people. Robots such as , finance management, and housekeep- can be used to enable people with cognitive, sensory, ing.14 The world also has a rapidly and motor impairments, help people who are ill aging population, who will only or injured, support caregivers, and aid the clinical add to this large number of people who may need ADL help. Of all of workforce. This article highlights several recent these individuals, few want to live advancements on these fronts, and discusses their in a long-term care facility. Instead, impact on stakeholders. It also outlines several key a World Bank; http://documents.worldbank. technological, logistical, and design challenges faced org/curated/en/2011/01/14440066/world-re- in healthcare robot adoption, and suggests possible port-disability avenues for overcoming them. key insights Robots are “physically embodied systems capable ˽˽ Over 20% of the world’s population of enacting physical change in the world.” They experience physical, cognitive, or sensory impairments. Robots can fill care gaps enact this change with effectors, which can move the and support independence. ˽˽ Robots can help caregivers and the robot (locomotion), or objects in the environment clinical workforce, who are overloaded (manipulation). Robots typically use sensor data and experience high rates of injury themselves.

to make decisions. They can vary in their degree ˽˽ In health, most problems are open- of autonomy, from fully autonomous to fully ended, and there is no “one-size-fits-all” solution. Every person, task, and care teleoperated, though most modern system have mixed setting are different, and require robots to be able to robustly learn and adapt initiative, or shared autonomy. More broadly, robotics on the fly. technology includes affiliated systems, such as related ˽˽ Technologists, researchers, providers, and end users must closely collaborate to 28 sensors, algorithms for processing data, and so on. ensure successful adoption. LEZZA/GETTY LAURA IMAGES BY PHOTO

68 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 At a nursing residence in Florence, Italy, a robot performs caregiving and support duties for 20 elderly guests. The robot was developed through the Robot-Era project supported by the European Union.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 69 review articles

Figure 1. The main stakeholders for healthcare robotics, and exemplar contextualizations of their relationship to the technology.

Stakeholder Context for Robotics Examples of Robotics Use Primary Stakeholders: Direct Robot Users (DRU): A DRU may directly use robotics technology to ˲˲ A person with a lower limb amputation uses a robotic People who directly use robots to aid them help them accomplish daily living activities, with arm to grasp objects with daily living or wellness activities. physical, cognitive, or social tasks. ˲˲ A person with autism works with a robot to learn to This may include people who experience read facial expressions difficulties with physical, cognitive, or sensory ˲˲ A person who has low vision uses a smart cane to functions, mental health, or behavioral health. sense obstacles These experiences may be temporary or permanent, acute or chronic, and may change throughout one’s lifespan. Clinicians (CL): A CL may use robotics technology while ˲˲ A therapist employs a therapeutic robotic pet in a Persons who may provide healthcare or work providing care, in the course of their training, treatment regiment with DRU. These individuals may be: nurses, or to help them with day-to-day administrative ˲˲ A nurse uses a robot to help lift a DRU from their physicians, mental healthcare providers, tasks. wheelchair to a bed rehabilitation professionals, pharmacists, ˲˲ A surgeon uses a robot to aid with a minimally EMTs, among others. invasive procedure ˲˲ A medical student uses a robotic patient simulator to learn how to treat a stroke Care Givers (CG): A CG may use robotics technology to directly or ˲˲ An adult child uses a telepresence robot to Family members, neighbors, volunteers, or indirectly support a DRU communicate with an older parent other unpaid persons who may support DRU. ˲˲ A friend may use a robot to perform household tasks in the DRU’s home

Secondary Stakeholders: Robot Makers (RM): A RM may work with DRU, CL, CG, PM, and ESW ˲˲ A company builds a hospital discharge robot Individuals who design, build, program, to perform their work. ˲˲ A student writes sensing algorithms for a robot to lift instrument, or research robotics technology. people out of a wheelchair ˲˲ A Maker club adapts toys to be accessible by children with motor impairments Environmental Service Workers (ESW): An ESW may use robotics technology to ensure ˲˲ An ESW teleoperates a disinfecting robot which emits Persons who provide secondary care to DRUs care environments are safe and sanitary to help UV light to kill superbugs in a hospital room by helping prevent the spread of infection prevent the spread of infection. Their use of ˲˲ An ESW uses a waste removal robot to safely through cleaning services. These can include robotics directly affects DRU’s quality of care, transport medical waste environmental service workers in hospitals, and CL’s workplace safety. housekeeping staff in nursing homes, and so on. Health Administrators (HA): An HA may purchase robots to support staff, ˲˲ A chief medical officer reviews clinical effectiveness Individuals who provide leadership to a patients, or visitors, or set policy on their usage. data of a rehabilitation robot care setting by planning, coordinating, and ˲˲ A HA preforms a cost effectiveness study of acquiring directing care delivery. robots for their institution

Tertiary Stakeholders: Policy Makers (PM): A PM may work with DRU, CL, CG, ESW, RM, and ˲˲ A Federal Food and Drug Administration (FDA) worker People who work for or with federal, state, AG to understand how to best craft policy for the establishes new policy for Home Use Devices and local governments to design policy use of robots. ˲˲ A Federal Trade Commission (FTC) worker sets regarding: how robots will be used, which privacy policies for robot sensors robots will be used, and how their costs will be managed. Insurers (IC): ICs may work with PM, AG, HA, RM, and CL to ˲˲ An IC worker explores the robotic exoskeletons Public or private organizations who makes establish guidelines for reimbursable robot- evidencee base to establish reimbursement policy decisions about benefits to DRU and CG, related services. ˲˲ An IC worker consults with a company to understand including service payments to CL and RM. a robot’s control system Advocacy Groups (AG): AGs may work with with DRU, CL, CG, RM, PM, ˲˲ An muscular dystrophy AG supports new research on Organizations who work on behalf of DRU and others to ensure robots are employed in exoskeletons populations ways that are of the best interest of their DRU ˲˲ An MS advocacy group lobbies congress to fund new population. robotic therapies

many people would prefer to live and $30,000 and $85,000 per year in pro- people who need care than healthcare age gracefully in their homes for as vider wages alone.b workers available to provide it.33 While long as possible, independently and Second, there is a substantial health- family members and friends attempt with dignity.22 However, for people care labor shortage—there are far more to fill these care gaps, they too have full- requiring help with ADL tasks, this time jobs and other familial obligations, goal is challenging to meet for a few b U.S. Department of Health and Human Servic- and thus cannot meet the need. Health- reasons. First, this level of care is quite es; http://longtermcare.gov/costs-how-to-pay/ care workers are not only overburdened expensive; in the U.S. it costs between costs-of-care/ by this labor shortage, but face increas-

70 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 review articles ingly hazardous work environments, use robots to provide assistance with cians, and caregivers, all of whom are and are themselves at great risk of de- daily living or wellness activities likely to use robotics technology on a bilitating injury and disability. Accord- (direct robot users (DRU)), health regular basis; Secondary beneficiaries: ing to the National Institute for Occu- professionals who use robots to health administrators, robot makers, pational Health and Safety (NIOSH), provide care (clinicians (CL)), non- and environmental service workers, health care workers have the most haz- CL individuals who support DRUs all of whom are involved in the use of ardous industrial jobs in America, with (care givers (CG)), technologists and robotics technology in healthcare set- the greatest number of nonfatal occu- researchers (robot makers (RM)), tings but do not directly use the robots pational injuries and illness.c health administrators (HAs), policy- to use robots to support the health Thus, there is an incredible opportu- makers (PMs), advocacy groups (AGs), and wellness of DRUs; and tertiary nity for robotics technology to help fill and insurers (IC). Figure 1 introduces beneficiaries: policymakers and advo- care gaps and help aid healthcare work- these stakeholders. cacy groups, who have interest in the ers. In both the research and commer- These stakeholders can be grouped use of robots to provide care to their cial space, robotics technology has been into three beneficiary groups: Primary constituents, but are unlikely to use used for physical and cognitive rehabili- beneficiaries: direct robot users, clini- them directly. tation, surgery, telemedicine, drug de- livery, and patient management. Robots Selected care settings where robots may be used. have been used across a range of envi- ronments, including hospitals, clinics, Care Setting Definition homes, schools, and nursing homes; Longer-Term and in both urban and rural areas. Assistive Living Facility “Congregate residential facility with self-contained living units Before discussing these applica- providing assessment of each resident’s needs and on-site support 24 tions, it is important to first contextual- hours a day, 7 days a week, with the capacity to deliver or arrange for ize the use of robots within healthcare. services including some health care and other services.” This article begins by identifying who Group Home “A residence, with shared living areas, where clients receive supervision and other services such as social and/or behavioral will be providing, receiving, and sup- services, custodial service, and minimal services (e.g., medication porting care, where this care will take administration). place, and key tasks for robots within Custodial Care Facility “A facility which provides room, board and other personal assistance these settings. Examples of new tech- services, generally on a long- term basis, and which does not include a medical component” nologies aimed at supporting these Nursing Facility “A facility which primarily provides to residents skilled nursing care stakeholders will be introduced, and and related services for the rehabilitation of injured, disabled, or sick key challenges and opportunities to persons, or, on a regular basis, health-related care services above realizing the potential use of robots in the level of custodial care to other than [people with intellectual healthcare that research and industry disabilities]” are encouraged to consider, will be ad- Home Care “Location, other than a hospital or other facility, where [a person] receives care in a private residence.” dressed. These adoption issues include a robot’s capability and function (Does Shorter-Term a robot have the required capabilities Inpatient Hospital “A facility, other than psychiatric, which primarily provides diagnostic, to perform its function?), cost effective- therapeutic (both surgical and nonsurgical), and rehabilitation services ness (What is the robot’s value to stake- by, or under, the supervision of physicians to patients admitted for a holders relative to its cost?), clinical ef- variety of medical conditions.” fectiveness (Has the robot been shown On/Off Campus “A portion of a… hospital provider based department which provides Outpatient Hospital diagnostic, therapeutic (both surgical and nonsurgical), and to have a benefit to stakeholders?), us- rehabilitation services to sick or injured persons who do not require ability and acceptability (How easy is hospitalization or institutionalization.” the robot to use, modify, and maintain? Urgent Care Facility “Location, distinct from a hospital emergency room, an office, or Is the robot’s form and function accept- a clinic, whose purpose is to diagnose and treat illness or injury able?), and safety and reliability (How for unscheduled, ambulatory patients seeking immediate medical attention.” safe and reliable is the robot?) Inpatient Psychiatric “A facility that provides inpatient psychiatric services for the diagnosis Facility and treatment of mental [health disorders] on a 24-hour basis, by or Stakeholders, Care Settings, under the supervision of a physician.” and Robot Tasks Hospice “A facility, other than a patient’s home, in which palliative and Stakeholders. For this article, stake- supportive care for terminally ill patients and their families are holders are defined as people who have provided.” a vested interest in the use of robot- Substance Abuse “A location which provides treatment for substance (alcohol and drug) Treatment Facility abuse on an ambulatory basis. Services include individual and group ics technology in healthcare. Stake- therapy and counseling, family counseling, laboratory tests, drugs and holders can be: people who directly supplies, and psychological testing.” Residential facilities also provide room and board.

Source: http://www.cms.gov/Medicare/Coding/ place-of-service-codes/Place_of_Service_Code_Set.html c National Institute for Occupational Safety and Health, http://www.cdc.gov/niosh/topics/ healthcare/

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 71 review articles

This article will focus on primary these tradeoffs in how they design and example, in surgical procedures, ro- beneficiaries; however, it is impor- test their systems. bots may provide clinicians with the tant to note that all other stakeholder Care tasks. Robots may be helpful ability to perform less invasive proce- groups are critical to the successful for many health tasks. Robots can pro- dures to areas of the body inaccessible end-deployment of robotics in health- vide both physical and cognitive task with existing instrumentation due to care, and should be included when support for both DRUs and clinicians/ issue or distance constraints. These possible in decision-making. caregivers, and may be effective and can include types of neurological, gas- Care settings. Another critical di- helping reduce cognitive load. Task tric, and fetal surgical procedures.41 mension to contextualizing the use of assistance is particularly critical as Direct robot users. When designing robotics in healthcare is to consider the the demand for healthcare services robots for DRUs, there is great value location of use. This can significant- is far outpacing available resources, in designing straightforward solutions ly impact on how suitable different which places great strain on clinicians to problems. At a recent workshop dis- technologies are for a given setting,12 and caregivers.33 cussing healthcare robotics, people and can affect the design of a robot Physical tasks. Clinicians. Tasks in- with Amyotrophic Lateral Sclerosis and its required capabilities. For ex- volving the “3Ds” of robotics—dirty, (ALS) and other conditions reported ample, while a 400-lb, 5’4” dual-arm dangerous, and dull—can be of partic- that most of all they just wanted “a ro- mobile manipulator may work well in ular value for clinical staff. Clinicians bot to change the oil.”30 In other words: a lab, it is ill-suited to an 80-sq. ft. room spend an inordinate amount of time help is most needed with basic, physi- in an assisted living facility. While it on “non-value added” tasks, for exam- cal ADL tasks, such as dressing, eating, is understandable robot makers may ple, time away from treating patients. ambulating, toileting, and housework. immediately be more concerned with The overburden of these tasks creates a Robots that can help people avoid fall- achieving platform functionality than climate for error; so robots, which can ing could also be incredibly beneficial, the particulars of care settings, to suc- help clinicians effectively, surmount as falls cause thousands of fatal and cessfully deploy healthcare robots, set- these challenges would be a boon. debilitating injuries per year. ting must be considered. Some of these non-value added tasks Currently, standalone robots that The accompanying table defines include: Transportation, such as mov- can successfully perform the majority different kinds of care settings, and ing materials or people from one place of these key physical ADL tasks are a includes longer-term care facilities in to another, Inventory, such as patients long way from reaching the consum- the community, as well as shorter-term waiting to be discharged, Search Time, er market. There are several reasons care facilities, such as hospitals. For such as looking for equipment or pa- for this. First, the majority of these longer-term care in the U.S., the Fair perwork, Waiting, for patients, mate- tasks remain challenging for today’s Housing Act, and Americans with Dis- rials, staff, medications, and Overbur- robots, as they require a high degree abilities Act set some general guide- dening of Staff and Equipment, such as of manual dexterity, sensing capabil- lines for living space accessibility; how- during peak surge times in hospitals.42 ity, prior task knowledge, and learn- ever, the majority of space guidelines is Two of the best tasks for robots in ing capability. Furthermore, most state-dependent, and can have a large this task space are material transporta- autonomous, proximate robots move degree of variation. For example, an tion and scheduling, which robots can extremely slowly due to safety and assisted living facility in Florida must be exceptionally skilled at given the computational purposes, which will provide 35-sq. ft. per resident for liv- right parameters. For example, robots undoubtedly be frustrating for end ing and dining, whereas in Utah it is that can fetch supplies, remove waste, users. Finally, even if robots could 100-sq. ft. An in-patient psychiatric fa- and clean rooms. Another task robots perform some of these more complex cility in Kentucky must provide 30-sq. can do that will help greatly improve ADL tasks, their power budgets may ft. per patient in social common areas, the workplace for clinicians is mov- make them impractical for deploy- Oregon requires 120-sq. ft. in total and ing patients. This is a very hazardous ment in most care settings. 40-sq. ft. per patient. task—hospital workers, home health However, there have been substan- Robots in healthcare can also affect workers, and ambulance workers ex- tial gains in recent years for other the well-being, health, and safety of perience musculoskeletal injuries be- tasks. For example, robots that pro- both direct robot users and clinicians. tween three and five times the national vide DRUs with additional physical The field of evidence based health- average when moving patients accord- reach (for example, smart on-body care design40 has produced hundreds ing to NIOSH. prostheses, wheelchair mounted ro- of studies showing a relationship Robots can also help clinicians with bot arms) and robots which provide between the built environment and other dangerous tasks, such as helping multi-setting mobility capability (for health and wellness, in areas including treat patients with highly infectious example, exoskeletons, accessible patient safety, patient outcomes, and diseases. Robot mediated treatment personal transportation devices).26 staff outcomes. When new technology has become particularly pertinent after These are likely to continue to be the such as a robot becomes part of a care the recent Ebola outbreak, where clini- types of systems that reach end users setting, it is now a possible disruptor to cians and caregivers can perform treat- first for the foreseeable future. health. HAs must balance the risks and ment tasks via telepresence robots.17 Cognitive tasks. Clinicians. Any benefits for adopting new technology, Finally, robots may help extend the technology that can effectively reduce and robot makers should be aware of physical capabilities of clinicians. For clinical workload is likely to be warm-

72 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 review articles ly embraced in healthcare. Many of pendence by providing sensory aug- these systems exist in a non-embodied mentation or substitution. For exam- fashion, for example, decision sup- ple, DRUs who are blind or low vision port tools to aid in emergency medi- may benefit from a robotic way finding cine,12 patient logistical management, tool, or DRUs using robotic prostheses or charting. However, robotic systems There is incredible might receive sensory feedback from a may have a place within this domain, opportunity for robotic finger in their shoulder. particularly if a robot is well integrated Robots also may be able to help into existing workflow and able to ac- robotics technology DRUs with regaining (or supplement- cess EHR data. For example, perhaps a to help fill care ing) cognitive function in neuroreha- medication management robot could bilitative settings, such as in cases of anticipate a clinician’s “next move” in gaps and aid stroke, post-traumatic stress disorder, treatment by prefetching a likely medi- or traumatic brain injury. Robots also cine from the pharmacy. Or perhaps a healthcare workers. may provide socio-emotional support robot could deliver personalized mes- Robotics have been to DRUs: to provide companionship, sages to family members in waiting teach people with autism to learn to rooms to update them on the status of used for physical read emotions, or to help reduce symp- their relative while clinicians are occu- and cognitive toms of dementia. However, there is a pied with other tasks. paucity of clinical effectiveness trials Another area where robotics has rehabilitation, showing DRU benefit compared to stan- been extensively used to aid clinicians surgery, dard treatment, so it is unclear what the with cognitive tasks is in clinical simu- future for these robots may be.29 lation and training. Robotic patient telemedicine, drug simulators are life-sized, humanoid delivery, and patient Recent Advances in robots that can breathe, bleed, speak, Healthcare Robotics expel fluids, and respond to medica- management. The 2016 U.S. Robotics Roadmap was tions. They are the most commonly recently released,1 which frames the used humanoid robot worldwide, and state of the art in robotics and future provide learners with the ability to si- research directions in the field. Over multaneously practice both procedural 150 robotics researchers contributed, and communication skills.22,23 These including the author of this article. The robots are used by inter-professional roadmap includes a detailed summary clinicians across a wide range of spe- of advancements in robotics relating cialties, including acute care, perioper- to health and wellness. Some key focus ative care, trauma, and mental health- areas include: aging and quality of life care. The author and her students have improvement, surgical and interven- been designing the next generation of tional robotics, rehabilitative robotics, these simulators, which can convey re- and clinical workforce support. alistic facial patient pathologies, such In general, robots used in these ar- as pain, stroke, and cerebral palsy, and eas can be divided into three categories: are integrated with on-board physi- inside the body, on the body, and out- ological models.23,29 side the body. Those inside and on the Direct robot users and care givers. body are primarily intended for direct The ways in which robots may be able robot users, and those outside the to provide cognitive task support to body for direct robot users, care giv- CGs has yet to be fully realized. How- ers, and clinicians. These robots have ever, similar to clinicians, the ability to the potential to be used across a range reduce cognitive load would be greatly of care settings and clinical foci, and welcomed. CGs in particular are often can provide both physical and cogni- overburdened when providing care; tive support. they frequently have other family mem- Inside the body. Recent advances bers to care for, other jobs, and their for internal robots have occurred in the own lives (and health) to manage.3 Ro- fields of microrobotics, surgical robot- bots might be able to cognitively sup- ics, and interventional robotics. Micro- port CGs by learning and anticipating robotics are micro-scale, untethered their needs, prefetching items, attend- devices that can move through the ing to time-intensive tasks which de- body and can perform a range of func- tract from care, and so on. tions, such as targeted therapy (that For DRUs, robotics technology is, localized delivery of medicine or might be able to help facilitate inde- energy), material removal (for exam-

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 73 review articles

Figure 2. Key examples of recent advances in healthcare robotics. Those inside and on the body are primarily intended for direct robot users, and those outside the body for direct robot users, caregivers, and clinicians. These robots have the potential to be used across a range of care settings and clinical foci, and can provide both physical and cognitive support. Image credits (clockwise from upper left): B. Nelson, R. Alterovitz, Mobius, TED, Ekso Bionics, B. Smart, L. Riek, S. Sabanovic, C. Kemp.

[6, 78, 33, 34]. In the natural orifice embodiment, transnasal skull base [12] and transoral throat [80] applications have been proposed, and it is likely that surgeries througInh theother natura bodyl orifices will be pursued in the future. In the percutaneous, On the body needle-like embodiment, applications that have been suggested include fetal umbil- ical cord blood sampling [29], ultrasound guided liver targeting and vein cannula- tion [74], vascular graft placement for hemodialysis [7], thermal ablation of cancer [8, 13], prostate brachytherapy [79], retinal vein cannulation [87, 91, 88], epilepsy treatments [19], and general soft tissue targeting procedures [45, 70, 35]. Of all these applications, the two that have been studied most extensively are the cardiac applications of Dupont et al. and the endonasal applications of Webster et al. This includes the first ever use ofaconcentric tube robot inalive animal by Gosline et al. [26, 33]. It also includes the first insertion of a concentric tube robot into a human cadaver by Burgner et al. [16, 12]. Many researchers have also explored the

Microrobots are micro-scale, Concentric tube robots (active Robotic prostheses and exoskeletons. People with forearm-to-shoulder untethered robots that can cannulas) can be used as small, amputations can use wearable robot prostheses, which can provide fine-grained move through the body and can teleoperated manipulators or as dexterity, reach, and strength. People with lower-limb amputations or lower-body per form targeted therapy,Fig. 1 A concentric tube robosteerablet next toastandar dneedles,da Vinci laparoscopi andc tool. enable muscle weakness can use powered-knee and ankle prostheses to do everything Seeing is Comforting: Effects omaterialf Teleoperato removal,r structuralVisibility procedures in areas inaccessible from running marathons to dancing. Exoskeletons have helped people with in Robot-Mediated Healtcontrol,h andCar sensing.e with traditional instruments. muscle weakness, movement disorders, or paralysis locomote.

Kory Kraft, William D. Smart Outside the body

Teleoperated robots can be used to provide medical care to patients in infectious disease outbreaks, alleviating work- ers from being in dangerous infectious zones longer than abso- lutely needed. Nevertheless, patients’ reactions to this technology have not been tested. We test three hypotheses related to patients’ comfort and trust of the operator and robot in a simulated Ebola Treatment Unit. Our findings suggest patients trust the robot

Providing medical care to people with highly infectious diseases exposes health care work to the risk of becoming infected themselves. This is especially significant when the Fig. 1: The robot moves an IV fluid pole durin g a study diseases are either highly contagious or extremely lethal. Mobile manipulators. Patient simulators. Over Mental and Behavioral Physical task support. session.Clinicians can safely 180,000 clinicians annually Healthcare. Robots can Robots can support people tele-operate mobile robots to train on high fidelity robotic support people with cognitive with motor impairments, treat patients with highly patient simulators, which can impairments, facilitate movement disorders, and infectious diseases such as simulate physiological cues, neurorehabilitation, support brain injuries to provide external Ebola Virus Disease. and sense and respond to wellness, or provide manipulation capabilities. learners. companionship.

ple, biopsy, ablation), structural con- for intuitive physical and cognitive in- ple with lower-limb amputations or trol (for example, stent placement), teraction between the user and robot, lower-body muscle weakness can use and sensing (for example, determin- new methods for managing uncertain- powered-knee and ankle prostheses to ing oxygen concentrations, sensing the ty, and providing 3D registration in real engage in a range of activities, includ- presence of cancer).25 Recent advances time while traversing both deformable ing everyday locomotion to running in the field have enabled actuating, and non-deformable tissue.1 marathons and dancing. Exoskele- powering, and controlling these robots On the body. In terms of wearable tons have helped people with muscle (see Nelson et al.25 for a review.) robots for DRUs, there have been re- weakness, movement disorders, and In surgical and interventional ro- cent advances in the areas of actu- paralysis locomote. botics, a range of advances have been ated robot prostheses, orthoses, and Several advances have been made made that enable clinicians to have exoskeletons. A prothesis supplants recently in how people interface with improved dexterity and visualization a person’s missing limb, and acts in these robots. For example, some ro- inside the body and reduce the degree series with a residual limb. An ortho- bot prostheses offer neural integra- of movement during operations.1 Fur- sis is a device that helps someone who tion to provide tactile feedback and thermore, promising advances have has an intact limb but an impairment, increasingly more intuitive control of been made in concentric tube (active and an exoskeleton provides either a the limb.1 Other advances include an cannula) robots. These robots are com- person with intact limbs (DRU or oth- increase in the workspace and range of prised of precurved, concentrically erwise) assistance or enhancement of motion of wearable robots, as well as nested tubes that can bend and twist existing physical capability. Orthoses improvements in user comfort. throughout the body. The robots can be and exoskeletons act in parallel to an Outside the body. Robots outside the used as small, teleoperated manipula- existing limb.39 body are being used across many clini- tors or as steerable needles. The robots All of these robots can be used to cal application spaces. For clinicians, can enter the body directly, such as enable DRUs to perform tasks. For ex- mobile manipulators are being used through the skin or via a body opening, ample, people with forearm-to-shoul- to help treat patients with highly infec- or could be used via an endoscope.11 der amputations can use wearable tious diseases,17 aid in remote surgical Some future research directions for robot prostheses, which can provide procedures,26 and help provide physi- in-the-body robots include new means dexterity, reach, and strength. Peo- cal assistance to CLs when moving pa-

74 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 review articles tients.24 They are also used extensively as 75% of hand rehabilitation robots avoid using anything which further in clinical training, as discussed earlier. were never actually tested with end us- advertises their differences, even if it Robots are also being explored in men- ers, rendering them completely unus- provides a health benefit.27,32,33 tal and behavioral healthcare applica- able in practice and abandoned.2 Shinohara and Wobbrock argue that tions. Robots are being used to support One the major challenges is that in addition to designers considering people with autism spectrum disorder clinicians, even those who are well-ed- the functional accessibility of system, and cognitive impairments, to encourage ucated and accomplished in their dis- they also consider its social accessi- wellness, and to provide companionship. ciplines, often have low technology lit- bility, and employ a “Design for Social (See Riek29 for a detailed review of eracy levels.19 Thus, if they themselves Acceptance” (DSA) approach.35 This these applications). find a robot unusable, the likelihood of means going beyond purely functional For physical task support, robots them successfully training a direct ro- designs, which may be “awkward and can provide external manipulation and bot user or caregiver to use the robot is clunky.”34 Robot makers are usually sensing capabilities to DRUs. For exam- greatly diminished. primarily concerned about a robot’s ple, wheelchair mounted robot arms Another challenge is that DRUs are functional capabilities; for example, can provide reach, smart wheelchairs often excluded from the robot design can the robot perform its task safely can help facilitate safe navigation and process, which leads to unusable and and reliably given workspace, envi- control, and telepresence robot sur- unsuitable technology. Robots with ronmental, and platform constraints. rogates can enable people with severe multiple degrees of freedom, such as However, the aforementioned litera- motor impairments the ability to fly, wearable prostheses or wheelchair- ture suggests that there may be great give TED talks, and make coffee.6,7,38 mounted arms, require a high level of value in also considering a robot’s ap- There are other examples of external cognitive function to control.38 How- pearance and behavior to help enable robots that are outside the scope of this ever, many people needing such robots technology adoption. paper, but could prove highly pertinent often have co-morbidities (that is, oth- Safety and reliability. When robots in healthcare. For example, autonomous er conditions), which can make control and people are proximately located, vehicles may provide new opportunities a further exhausting process. safety and reliability are incredibly for DRUs to locomote, or may enable There are several ways to address important. This is even more critical EMTs to focus on treating patients rather this issue. One approach is for robot for DRUs who may rely extensively on than driving ambulances. Telepresence makers to reduce robot complexity. robots to help them accomplish physi- may also have unforeseen applications Balasubramanian et al.2 argue for func- cal or cognitive tasks, and who may not in healthcare, such as through aerial tional simplicity in therapeutic robot have the same ability to recover from manipulation, drone delivery of medi- design, which will lead to robots that robot failures as easily as non-DRUs. cal supplies, among others. are easier for all primary stakehold- There has been a fair bit of work ers to use, control, and maintain. This on safe physical human-robot inter- Healthcare Robotics Adoption: concept is echoed in much of the reli- action, particularly with regard to im- Challenges and Opportunities ability and fault tolerance literature; proving collision avoidance, passive While there are exciting advances in lower-complexity robots are more like- compliance control methods, and new healthcare robotics, it is important to ly to be longitudinally reliable and fault advances in soft robotics to facilitate carefully consider some of the chal- tolerant. gentle interaction.37 There also have lenges inherent in healthcare robotics, Forlizzi and Zimmerman propose been recent advances on algorithmic and discuss ways to overcome them. the idea of a service-centered design verifiability for robots operating in par- Robots have the ability to enact physi- process, wherein rather than only tially unknown workspaces,18 which cal change in the world, but in health- think about a single user and a sys- may prove fruitful in the future. care that world is inherently safety tem, designers consider including However, there has been little work critical, populated by people who may the broader ecosystem surrounding to date on safe cognitive human-robot be particularly vulnerable to harm due a technology.10 This is a particularly interaction. People with cognitive dis- to their disability, disorder, injury, or beneficial idea in healthcare robotics. abilities and children are particularly illness. Stakeholders face five major Rarely will there be one DRU and one prone to being deceived by robots.29 considerations when considering de- robot; rather, there is a complex so- This is an important and under-ex- ploying robots in healthcare: Usability cial structure surrounding caregiving plored question in the robotics com- and acceptability, safety and reliabil- that should be considered carefully in munity, though a few efforts have been ity, capability and function, clinical robot design. made recently with regard to encourag- effectiveness, and cost effectiveness. Another important barrier to health- ing robot makers to employ value-cen- Each is explored here. care robot adoption is its acceptability. tered design principles. For example, Usability and acceptability. Robots The morphology, behavior, and func- ensuring the appearance of the robot that are difficult for primary stakehold- tionality of a robot play a major role in is well-aligned with its function (for ex- ers to use have a high likelihood of be- its adoption and use. When a DRU uses ample, avoiding false-advertising), en- ing abandoned. This phenomenon has a robot in public, they are immediately abling transparency into how a robot been well documented in the Assistive calling attention to their disability, dis- makes decisions, and maintaining the Technology Community.5,9,20 For exam- order, or illness. DRUs already face sig- privacy and dignity of DRUs.15,31 ple, a 2010 study reported that as many nificant societal stigma, so frequently Another way to help bridge the

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 75 review articles

safety gap is for robot makers to em- approaches, which help enable more ploy in-depth testing and training robust algorithms. Furthermore, life- regimens that enable direct robot long learning and longitudinal experi- users, care givers, and clinicians mental approaches have also enabled to fully explore the capabilities of researchers to surmount some of these a platform. This can help prevent Robotic patient perceptual challenges. Modeling situ- people from either over-relying or simulators are ational context and object and environ- under-relying on the robot, and help mental affordances within them can facilitate trust. life-sized, humanoid also be a useful tool in surmounting Capability and function. The field robots that can these issues.1,28 of robotics has seen amazing capa- Learning, too, is a challenge. It is bility gains in recent years, some of breathe, bleed, critical that primary stakeholders, which have been instrumental in who have a wide range of physical healthcare. However, despite these speak, expel fluids, abilities, cognitive abilities, and tech- advances, robotics is still an excep- and respond to nology literacy levels, are able to eas- tionally difficult problem. For exam- ily repurpose or reprogram a robot ple, many demonstrations in robotics medications. without a RM present. This level of technology remain demos, and fail adaptability and accessibility presents outside of highly constrained situa- robot makers with a complex techni- tions.8 This is particularly problem- cal and socio-technical challenge. As atic when designing technology for mentioned previously, simple is un- healthcare: most problems are open- doubtedly better; it helps constrains ended, and there is no “one-size-fits- the problem space and lowers the all” solution.12,28 Every person, task, complexity of the system. Another ma- and care setting are different, and jor aid will be the research community require robots to be able to robustly continuing to develop new datasets, learn and adapt on the fly. evaluation metrics, and common plat- As discussed previously, care set- forms;8 these have shown to be useful tings differ substantially. Even the in other computing domains, so are same type of care setting, such as an likely to be helpful here. emergency department or assisted Cost effectiveness. When robots living facility, have substantial differ- are being acquired in healthcare, it is ences in their environment, practices, important that their cost effectiveness and culture. In our prior work design- is considered beyond the purchase, ing health information technology, we maintenance, and training costs for have demonstrated that these differ- the system. For example, when elec- ences can be surmounted by conduct- tronic health records (EHRs) were first ing multi-institutional trials, and by employed in hospitals, they were tout- building solutions that are adaptable ed as a means to save clinicians and to different care settings.13 The same patients’ time. However, because EHR approach can be taken in robotics. systems were so poorly designed, diffi- Real-world, real-time, robust per- cult to use, and poorly integrated into ception in human environments is a existing they ended up creating sub- another major challenge in robotics. stantially more non-value added work. While the field of computer vision has This resulted in “unintended conse- seen advances in solving still-image, quences,” including increasing costs fixed-camera recognition problems, and patient harm.16 It is critical these those same algorithms perform poorly same pitfalls are avoided for robots. when both the cameras and people The Agency for Healthcare Research are moving, data is lost, sensors are and Quality (AHRQ) created a guide occluded, and there is clutter in the for reducing these unintended conse- environment. However, these situa- quences for EHRs;16 the same method- tions are highly likely in human social ological approach can be employed for settings, and it is an open challenge to robots. For example, when assessing sense, respond to, and learn from end the acquisition and deployment of a users in these settings.28 There have robot in a first place: been some recent advances, however: ˲˲ Are you ready for a robot (and is a the fields of social signal processing robot ready for you)? HAs must care- and human-robot interaction have fully consider their institution’s moved toward multimodal sensing robot readiness. Robots may solve

76 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 review articles some problems, but may make others ˲˲ Detailed deployment plans are cal studies on effectiveness, or can worse. For example, suppose a supply- provided but are not overly complicated; synthesize the existing literature in a fetching robot that is purchased help ˲˲ There are multiple ways to col- systematic review. nurses save time. However, it has dif- lect, analyze, and act on feedback from All consumer in-the-body robots ficulties functioning at high volume users; and many on-the-body robots must times of day due to sensor occlusion, ˲˲ Success metrics should be deter- undergo regulatory approval before so supply deliveries end up being de- mined in advance and evaluated con- they can be marketed and sold. In the layed. This causes a cascade effect, tinually; and, United States, this approval is through increasing the workload of nurses. ˲˲ Quality improvement should be the FDA, which typically requires a Situations like these can be remedied supported on an ongoing basis. strong level of evidence showing the through a careful exploration of exist- Recently, the IEEE released a docu- effectiveness and safety of a medical ing workflow in a unit, and by fully un- ment on “Ethically Aligned Design” device. Outside-the-body robots typi- derstanding a robot’s existing capa- which contains detailed suggestions cally do not need to undergo a device bilities and limitations. See Gonzales for how to engage in this value-cen- review process provided they fall with- et al.12,13 for examples on engaging in tered practice in engineering, which in existing classifications; for exam- this process with clinicians in safety could be helpful for all stakeholders ple, Paro the robot seal (see Figure 2, critical settings. moving forward.e bottom right) is classified by the FDA ˲˲ Why do you want a robot? It is im- Clinical effectiveness. Clinical ef- a neurological therapeutic device, and portant stakeholders define exactly fectiveness answers the question: thus is exempt from premarket review. why a robot is necessary for a given task “Does it work?” In particular, does a Shimshaw et al.36 argue this lack of in the first place. What are the goals of given intervention provide benefit to regulation of healthcare robots may the stakeholders? What is the plan for a primary stakeholder? This question be harmful to stakeholders both phys- deploying the robot, and how will suc- is answered by conducting thorough, ically and informationally, and should cess be measured? These questions evidence-based science. For robots be subject to premarket review on di- can also be explored through design directly affecting DRUs, this evidence mensions including privacy, safety, activities while assessing workflow and comes from comparative effectiveness reliability, and usability. institutional readiness. research (CER), which is “generated In the meanwhile, while the policy ˲˲ How do you select a robot? As men- from research studies that compare community races to catch up with tioned previously, functionality is only drugs, medical devices, tests, surger- technology, the robotics community one aspect to a robot; there is also: us- ies, or ways to deliver healthcare.”f can and should engage in research ability, acceptability, safety, reliabil- CER can include both new clini- that tests the clinical effectiveness of ity, and clinical effectiveness. While robots across care settings. Begum et there are not yet definitive guidelines al.4 suggest robot makers follow exist- to aid HAs in this process, science pol- e http://standards.ieee.org/develop/indconn/ ing clinical effectiveness benchmarks ec/ead_v1.pdf icy is starting to be shaped within this f Agency for Health and Research Quality, ef- within their intended care space and space. The CCC recently held an event fectivehealthcare.ahrq.gov/index.cfm/what-is- adopt them for use with robots. Fur- entitled “Discovery and Innovation in comparative-effectiveness-research1/ thermore, Riek29 suggests that when Smart and Pervasive Health,”d which brought together over 60 researchers Figure 3. Factors that will affect the widespread adoption of robotics in healthcare. from across academia, industry, and government, many of whom are ro- boticists who work in health. These efforts will hopefully begin to provide guidelines in the future. ˲˲ What are the recommended prac- tices for avoiding unintended conse- quences of robot deployment? Suc- cessfully deploying robots is a difficult process that may result in a disruptive care setting, and upset key stakehold- ers. To avoid unintended consequenc- es, it is important that: ˲˲ The robot’s scope is well-defined with clear goals; ˲˲ Key stakeholders are included and engaged in the deployment from the onset; d http://cra.org/ccc/events/discovery-innovation- smart-health/

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 77 review articles

conducting CER with robots, particu- real world, and will remain so for the ACM/IEEE International Conference on Human-Robot Interaction, 32–41. larly in cognitive support settings, it foreseeable future; however, they still 23. Moosaei, M., Gonzales, M.J., and Riek L.D. Naturalistic is not sufficient to simply test robot have the potential to be a remarkable pain synthesis for virtual patients. International Conference on Intelligent Virtual Agents, 2014. vs.no-robot, as the morphology can game changer in health. 24. Mukai, T., Hirano, S., Nakashima, H., Kato,Y., Sakaida, affect outcomes, but to instead to test Y., et al. Development of a nursing-care assistant robot RIBA that can lift a human in its arms. IEEE actuated vs. non-actuated. Acknowledgments Intelligent Robots and Systems, 2010. Some research reported in this ar- 25. Nelson, B.J., Kaliakatsos, I.K., and Abbott, J.J. Microrobots for minimally invasive medicine. Annual Discussion ticle is based upon work supported by Rreview of Biomedical Engineering 12 (2010), 55–85. 26. Okamura, A.M., Mataric, M.J., and Christensen, H.I. Healthcare robotics is an exciting, the National Science Foundation un- Medical and health-care robotics. Robotics and emerging area that can benefit all der Grant Nos. IIS-1253935 and SES- Automation 17, 3 (2010), 26–27. 27. Parette, P. and Scherer, M. Assistive technology use stakeholders across a range of set- 1457307, and the Luce Foundation. and stigma. Education and Training in Developmental tings. There have been a number of Disabilities, 2004, 217–226. 28. Riek, L.D. The social co-robotics problem space: Six exciting advances in robotics in recent References key challenges. Robotics challenges and vision. In years, which point to a fruitful future. 1. A roadmap for US robotics: From Internet to robotics Proceedings of the Workshop at Robotics: Science and (Nov. 2016); http:// jacobsschool.ucsd.edu/ Systems, 2013. How these robots ultimately will be in- contextualrobotics/docs/rm3-final-rs.pdf, November 2016. 29. Riek, L.D. Robotics technology in mental health care. tegrated into the lives of primary ben- 2. Balasubramanian, S., Klein, J., and Burdet, E. Robot- Artificial Intelligence in Behavioral and Mental Health assisted rehabilitation of hand function. Curr Opin Care. D. Luxton, (ed). Academic Press, 2015. eficiaries remains unknown, but there Neurol, 2010. 30. Riek, L.D, Hartzog, W., Howard, D.A., Moon, A., and is no doubt that robots will be a major 3. Bastawrous, M. Caregiver burden-a critical discussion. Calo, R. The emerging policy and ethics of human Int’l J of Nursing Studies 50, 3 (2013), 431–441. robot interaction. HRI (Extended Abstracts), 2015. enabler (and disruptor) to health. 4. Begum, M., Serna, R.W., and Yanco, H.A. Are robots 31. Riek, L.D and Howard, D. A code of ethics for the ready to deliver autism interventions? A comprehensive It is critical that both the research human-robot interaction profession. In Proceedings of review. International J. Social Robotics 8, 2 (2016). We Robot, 2014. and industrial communities work to- 5. Brose, S.W., Weber, D.J., Salatin, B.A., Grindle, G.G., 32. Riek, L.D and Robinson, P. Using robots to help people Wang, H., et al. The role of assistive robotics in the gether to establish a strong evidence- habituate to visible disabilities. In IEEE International lives of persons with disability. Am J Phys Med, 2010. Conference on Rehabilitation Robotics, 2011. base for healthcare robotics. As we 6. Carlson, T. and Demiris, Y. Collaborative control for 33. Shi, L. and Singh, D. A. Delivering health care in a robotic wheelchair: evaluation of performance, America. Jones & Bartlett Learning, 2014. have learned from the large-scale de- attention, and workload. IEEE Trans. Systems, Man, and 34. Shinohara, K. A new approach for the design of ployment of EHRs, technology devel- Cybernetics, Part B (Cybernetics) 42, 3 (2012), 876–888. assistive technologies: Design for social acceptance. 7. Chen, T.L. et al. Robots for humanity: using assistive ACM SIGACCESS Accessibility and Computing, 2012. opment and deployment cannot hap- robotics to empower people with disabilities. IEEE 35. Shinohara, K. and Wobbrock, J.O. In the shadow pen in a vacuum, or it is likely to cause Robotics & Automation 20, 1 (2013), 30–39. of misperception: Assistive technology use and 8. Christensen, H.I., Okamura, A., Mataric, M., Kumar, social interactions. In Proceedings of the SIGCHI grave harm to DRUs, overwhelming V., Hager, G., and Choset, H. Next generation robotics Conference on Human Factors in Computing Systems, stress to clinicians, and astronomical (2016); arXiv preprint arXiv:1606.09205. 2011. 9. Dawe, M. Desperately seeking simplicity: how young 36. Simshaw, D., Terry, N., Hauser, K., and Cummings, unseen costs. It is wise for all stake- adults with cognitive disabilities and their families M. Regulating healthcare robots: Maximizing holders to proceed cautiously and de- adopt assistive technologies. In Proceedings of opportunities while minimizing risks. Richmond J. of the Conference on Human Factors in Computing Law & Tech, 2016. liberately, and consider the full con- Systems, 2006. 37. Trivedi, D., Rahn C.D., Kier, W.M., and Walker, I.D. Soft 10. Forlizzi, J. and Zimmerman, J. Promoting service robotics: Biological inspiration, state of the art, and text of care as much as possible. design as a core practice in interaction design. In th future research. Applied Bionics and Biomechanics, It is also critical that direct robot Proceedings of the 5 IASDR World Conference on 2008. Design Research, 2013. users remain directly involved in the 38. Tsui, K.M., Kim, D.J., Behal, A., Kontal, D., and Yanco, 11. Gilbert, H.B., Rucker, D.C., and Webster III, R.J. H. A. ‘I want that:’ Human-in-the-loop control of a research, development, and deploy- Concentric tube robots: The state of the art and wheelchair-mounted robotic arm. Applied Bionics and future directions. Robotics Research. Springer, 2016, ment of future robots in health and Biomechanics 8, 1 (2011), 127–147. 253–269. 39. Tucker, M.R. et al. Control strategies for active lower wellness across the entire lifecycle of 12. Gonzales, M.J., Cheung, V.C., and Riek, L.D. Designing extremity prosthetics and orthotics: a review. J. of collaborative healthcare technology for the acute care th Neuroengineering and Rehabilitation, 2015. a project, as ultimately they are the workflow. InProceedings of the 9 Int’l Conference on 40. Ulrich R.S. et al. A review of the research literature ones who will be using these robots. Pervasive Computing Technologies for Healthcare, 2015. on evidence-based healthcare design. Health 13. Gonzales, M.J., Henry, J.M., Calhoun, A.W., and Riek, Environments Research & Design J., 2008. As discussed earlier, ignoring DRU L.D. Visual task: A collaborative cognitive aid for acute th 41. Webster R.J., Okamura, A.M., and Cowan, N.J. Toward input leads to unusable, unsuitable, care resuscitation. In Proceedings of the 10 Int’l active cannulas: Miniature snake-like surgical robots. Conference on Pervasive Computing Technologies for IEEE/RSJ Intelligent Robots and Systems. IEEE, and abandoned robots, which ben- Healthcare, 2016. 2006. efits no one. Secondary and Tertiary 14. Graf, C. The Lawton instrumental activities of daily 42. Wellman, J., Jeffries, H., and Hagan, P. Leading the living scale. The American J. Nursing 108, 4 (2008). Lean Healthcare Journey: Driving Culture Change to stakeholders should look to the Pa- 15. Hartzog, W. Unfair and deceptive robots. Maryland Increase Value. CRC Press, 2016. Law Review 74, 785 (2015). tient Centered Outcomes Research In- 16. Jones, S.S et. al. Guide to reducing unintended g consequences of electronic health records. Agency for stitute (PCORI) as a highly successful Laurel D. Riek ([email protected]) is an associate professor Healthcare Research and Quality, 2011. of computer science and engineering at the University of model for how-to engage with primary 17. Kraft, K. and Smart, W.D. Seeing is comforting: effects California, San Diego. She directs the Healthcare Robotics of teleoperator visibility in robot-mediated health care. stakeholders in clinical research and lab and builds autonomous robots that can sense, The Proceedings of the 11th ACM/IEEE International understand, and learn from real people in the real world. development. Conference on Human Robot Interaction, 2016, 11–18. Finally, it is important that ro- 18. Lahijanian, M., Maly, M.R., Fried, D., Kavraki L.E., Kress- Gazit, H., and Vardi, M.Y. Iterative temporal planning bot makers work with DRUs to help in uncertain environments with partial satisfaction © 2017 ACM 0001-0782/17/11 $15.00 bridge technology literacy gaps and guarantees. IEEE Trans. Robotics, 2016. 19. Lluch, M. Healthcare professionals’ organizational appropriately set expectations. Most barriers to health information technologies—A literature people’s experience with robotics review. International J. Medical Informatics, 2011. 20. Lu, E.C. et al. Development of a robotic device for comes from movies or media, which upper limb stroke rehabilitation: A user-centered design approach. Paladyn 2, 4 (2011), 176–184. rarely reflects the true state of af- 21. Milligan, C. There’s no place like home: Place and care Watch the author discuss fairs. Robots are quite fallible in the in an ageing society. Ashgate Publishing Ltd., 2012. her work in this exclusive 22. Moosaei, M., Das, S.K., Popa, D.O., and Riek, L.D. Communications video. Using facially expressive robots to calibrate https://cacm.acm.org/videos/ g http://pcori.org clinical pain perception. In Proceedings of the 2017 healthcare-robotics

78 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 research highlights

P. 80 P. 81 Technical Perspective Heads-Up Limit Hold’em Solving Imperfect Poker Is Solved Information Games By Michael Bowling, Neil Burch, By David Silver Michael Johanson, and Oskari Tammelin

P. 89 P. 90 Technical Perspective The Heat Method for Exploring Distance Computation a Kingdom by By Keenan Crane, Clarisse Weischedel, and Max Wardetzky Geodesic Measures By Marc Alexa

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 79 research highlights

DOI:10.1145/3131286 To view the accompanying paper, Technical Perspective visit doi.acm.org/10.1145/3131284 rh Solving Imperfect Information Games By David Silver

THE STUDY OF GAMES is as old as com- ta) both defeating human profession- puter science itself. Babbage, Turing, als using variants of CFR—although a and Shannon devised algorithms and The methods used near-perfect solution remains out of hardware to play the game of chess. to solve poker reach. The final challenge would be Game theory began with questions the variant most widely played by hu- regarding optimal strategies in card are quite general, mans: multiplayer no-limit poker. games and chess, later developed into and therefore The methods used to solve poker a formal system by von Neumann. are quite general, and therefore have Chess subsequently became the dro- have potential potential applications far beyond this sophila—or common fruitfly, the most applications far one game. Many other imperfect in- studied organism in genetics—of ar- formation games played by humans, tificial intelligence research. Early beyond this including a wide variety of card games, successes in chess and other games one game. board games, and video games, are shaped the emerging field of AI: many tractable to these methods. Further- planning algorithms first used in more, there are many real-world ap- games became pillars of subsequent plications, such as auctions, nego- research; reinforcement learning was tiations, and security, in which agents first developed for a checkers play- factual regret minimization (CFR), for receive different information, and ing program; and the performance of approximating a Nash equilibrium. must make a sequence of decisions to game-playing programs has frequent- Several years ago, a program called Po- maximize a final pay-off—and there- ly been used to measure progress in laris—created by many of the authors fore belong to the same class of im- AI. of the following paper—defeated for perfect information games as HULHE. Most of this research focused on the first time a human professional Solving a problem attains perfec- perfect information games, in which poker player in HULHE. tion in one domain. The frontier of all events are observed by all players, However, Polaris was still far from solved domains is an incontrovertible culminating in programs that beat perfect; indeed, it turns out in retro- measure of current computer capa- human world champions in check- spect that it was exploitable, due to bilities. That frontier has now been ers, chess, Othello, backgammon, and the approximations it made, by a very extended by one significant step, to most recently, Go. However, many large margin. The obvious remaining include for the first time a challenging applications in the real world have question was whether a “near- per- imperfect information game. imperfect information: each agent fect” solution could be found—a strat- observes different events. This leads egy so close to a Nash equilibrium David Silver leads the reinforcement learning research group at Google DeepMind, London, and is lead researcher to the possibility of deception and a that it cannot be differentiated in a on AlphaGo. wealth of social strategies. Imperfect lifetime of play. information games provide a micro- The following paper takes the CFR cosm of these social interactions, methods used in previous work to the while abstracting away the messiness next level. Using a number of innova- of the real world. tions—and several hundred machine- Among imperfect information years of computation—they were games, Poker is the most widely stud- able to find a near-perfect solution to ied—the latest drosophila—due to HULHE. Their solution also provides its enormous popularity and strate- insights into the game itself, showing gic depth. The smallest competitively the dealer holds a significant advan- played variant by humans, and the tage, and that seemingly poor hands most widely played by computers, is should be played surprisingly often. the two-player game known as Heads- For the game of poker, the next Up Limit Hold’Em (HULHE), in which step beyond HULHE is no-limit poker, each player holds two private cards in which has a much larger action space. addition to five public cards. Two de- This too has recently been cracked, cades of research in this game has led with the programs Libratus (from to powerful methods, such as counter- CMU) and DeepStack (also from Alber- Copyright held by author.

80 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 DOI:10.1145/3131284 Heads-Up Limit Hold’em Poker Is Solved By Michael Bowling, Neil Burch, Michael Johanson, and Oskari Tammelin

Abstract In perfect information games, all players are informed of Poker is a family of games that exhibit imperfect informa- everything that has occurred in the game prior to making tion, where players do not have full knowledge of past events. a decision. Chess, checkers, and backgammon are exam- While many perfect information games have been solved ples of perfect information games. In imperfect informa- (e.g., Connect-Four and checkers), no nontrivial imperfect tion games, players do not always have full knowledge of information game played competitively by humans has past events (e.g., cards dealt to other players in bridge and previously been solved. In this paper, we announce that poker, or a seller’s knowledge of the value of an item in an the smallest variant of poker in-play, heads-up limit Texas auction). These games are more challenging, with theory, hold’em, is now essentially weakly solved. Furthermore, this computational algorithms, and instances of solved games computation formally proves the common wisdom that the lagging behind results in the perfect information setting.b dealer in the game holds a significant advantage. This result And, while perfect information may be a common prop- was enabled by a new algorithm, CFR+, which is capable of erty of parlor games, it is far less common in real-world solving extensive-form games three orders of magnitude decision making settings. In a conversation recounted by larger than previously possible. This paper is an extended Bronkowski, John von Neumann, the founder of modern version of the original 2015 Science article,9 with additional game theory, made the same observation, “Real life is not results showing Cepheus’ in-game performance against like that. Real life consists of bluffing, of little tactics of computer and human opponents. deception, of asking yourself what is the other man going to think I mean to do. And that is what games are about in my theory.”12 1. INTRODUCTION Von Neumann’s statement hints at the quintessential Games have been intertwined with the earliest developments in game of imperfect information: the game of poker. Poker computation, game theory, and Artificial Intelligence (AI). At the involves each player being dealt private cards, with play- very conception of computing, Babbage had detailed plans for ers taking structured turns making bets on having the an “automaton” capable of playing tic-tac-toe and dreamt of his strongest hand (possibly bluffing), calling opponent bets, Analytical Engine playing chess.4 Both Alan Turing46 and Claude or folding to give up the hand. Poker played an important Shannon,40 on paper and in hardware respectively, developed role in the early developments of the field of game theory. programs to play chess as validation of early ideas in computa- Borel7 and von Neumann’s47, 48 foundational works were tion and AI. For over a half-century, games have continued to motivated by developing a mathematical rationale for act as testbeds for new ideas and the resulting successes have bluffing in poker, and small synthetic poker gamesc were marked significant milestones in the progress of AI: For exam- commonplace in many early papers.7, 29, 32, 48 Poker is also ple, the checkers-playing computer program Chinook becoming arguably the most popular card game in the world with the first to win a world championship title against humans,38 over 150mn players worldwide.1 The most popular vari- Deep Blue defeating Kasparov in chess,14 and Watson defeat- ant of poker today is Texas hold’em. When it is played ing Jennings and Rutter on Jeopardy!17 However, defeating top with just two-players (heads-up) and with fixed bet-sizes human players is not the same as “solving” a game, that is, com- and number of raises (limit), it is called Heads-Up Limit puting a game-theoretically optimal solution that is incapable Hold’Em (HULHE).10 HULHE was popularized by a series of losing against any opponent in a fair game. Solving games has also served as notable milestones for the advancement of b AI, for example, Connect-Four2 and checkers.39 For example, Zermelo proved the solvability of finite, two-player, zero-sum, perfect information games in 1913,51 while von Neuman’s more general Every nontrivial game played competitively by humans minimax theorem appeared in 1928.47 Minimax and alpha-beta pruning, the that has been solved to-date is a perfect information game.a fundamental computational algorithm for perfect information games, was developed in the 1950s, while Koller and Megiddo’s first polynomial-time technique for imperfect information games was introduced in 1992.26 a We use the word trivial to describe a game that can be solved without the c We use the word synthetic to describe a game that was invented for the use of a machine. The one near-exception to this claim is oshi-zumo, but it ­purpose of being studied or solved rather than played by humans. A synthetic is not played competitively by humans and is a simultaneous-move game game may be trivial, such as Kuhn poker,29 or nontrivial such as Rhode Island that otherwise has perfect information.13 Furthermore, almost all nontrivial hold’em.41 games played by humans that have been solved to-date also have no chance elements. The one notable exception is hypergammon, a three-checker variant of backgammon invented by Hugh Sconyers in 1993 which he then The original version of this paper was published in strongly solved, that is, the game-theoretic value is known for all board Science 347, 6218 (Jan. 2015) 145–149. Adapted with ­positions. It has seen play in human competitions. See http://www.bkgm.com/ permission from AAAS. variants/HyperBackgammon.html (accessed July 4, 2014).

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 81 research highlights

of high-stakes games chronicled in the book The Professor, In this paper, we announce that heads-up limit Texas the Banker, and the Suicide King.16 It is also the smallest hold’em poker is essentially weakly solved. Furthermore, we variant of poker played competitively by humans. HULHE bound the game-theoretic value of the game, proving that has 3.16 × 1017 possible states the game can reach mak- the game is a winning game for the dealer. ing it larger than Connect Four and smaller than check- ers. However, as an imperfect information game, many of 2. SOLVING IMPERFECT INFORMATION GAMES these states cannot be distinguished by the acting player The classical representation for an imperfect information as they involve information about unseen past events (i.e., setting is the extensive-form game. Here the word “game” private cards dealt to the opponent). As a result, the game refers to a formal model of interaction between self- has 3.19 × 1014 decision points where a player is required interested agents and applies to both recreational games to make a decision. and serious endeavors such as auctions, negotiation, and While smaller than checkers, the imperfect infor- security. See Figure 1 for a graphical depiction of a por- mation nature of HULHE makes it a far more challeng- tion of a simple poker game in extensive-form. The core of ing game for computers to play or solve. It was 17 years an extensive-form game is a game tree specifying branches of after Chinook won its first game against world champion possible events, namely player actions or chance outcomes. Marion Tinsley in checkers that the computer program The branches of the tree split at game states and each Polaris won the first meaningful match against profes- is associated with one of the players (or chance) who is sional poker players.34 While Schaeffer et al. solved check- responsible for determining the result of that event. The ers in 2007,39 heads-up limit Texas hold’em poker, until leaves of the tree signify the end of the game, and have now, was unsolved. This slow progress is not for lack of an associated utility for each player. The states associated effort. Poker has been a challenge problem for artificial with a player are partitioned into information sets, which intelligence, operations research, and psychology with are sets of states which the acting player cannot distin- work going back over 40 years.6 17 years ago, Koller and guish between (e.g., corresponding to states where the Pfeffer28 declared, “we are nowhere close to being able opponent was dealt different private cards). The branches to solve huge games such as full-scale poker, and it is from states within an information set are the player’s unlikely that we will ever be able to do so.” The focus on available actions. A strategy for a player specifies for HULHE as one example of “full-scale poker” began in ear- each information set a probability distribution over the nest over ten years ago,5 and became the focus of dozens of research groups and hobbyists after 2006 when it became the inaugural event in the Annual Computer Figure 1. Portion of the extensive-form game representation of 29 Poker Competition,53 held in conjunction with the main three-card Kuhn poker where player 1 is dealt a queen (Q) and the opponent is given either the Jack (J) or King (K). Game states conference of the Association for the Advancement of are circles labeled by the player acting at each state (“c” refers Artificial Intelligence (AAAI). This paper is the culmina- to chance, which randomly chooses the initial deal). The arrows tion of this sustained research effort toward solving a show the events the acting player can choose from, labeled with “full-scale” poker game.10 their in-game meaning. The leaves are square vertices labeled with Allis3 gives three different definitions for solving a game. the associated utility for player 1 (player 2’s utility is the negation of player 1’s). The states connected by thick gray lines are part A game is said to be ultra-weakly solved if for the initial of the same information set, that is, player 1 cannot distinguish position(s), the game-theoretic value has been determined; between the states in each pair since they represent a different weakly solved if for the initial position(s), a strategy has been unobserved card being dealt to the opponent. Player 2’s states are determined to obtain at least the game-theoretic value, also in information sets, containing other states not pictured in this diagram. for both players, under reasonable resources; and strongly

solved if for all legal positions, a strategy has been deter- 1:K 2:J c 1:J 2:K 1:Q 2:K mined to obtain the game-theoretic value of the position, for 1:K 2:Q

both players, under reasonable resources. In an imperfect 1:Q 2:J

information game, where the game-theoretic value of a posi- 1:J 2:Q tion beyond the initial position is not unique, Allis’s notion 11 of “strongly solved” is not well-defined. Furthermore, imper- pass pass fect information games, due to stochasticity in the players’ bet bet strategies or the game itself, typically have game-theoretic values that are real-valued rather than discretely valued 2222

pass (such as “win,” “loss,” and “draw” in chess and checkers), fold pass fold and only achieved in expectation over many playing of the call bet call bet game. As a result, game-theoretic values are often approxi- mated, and so an additional consideration in solving a game 211 1–21 1 –1 is the degree of approximation in a solution. A natural level fold fold of approximation under which a game is essentially weakly call call solved is if a human lifetime of play is not sufficient to estab- 2–1–2–1 lish with statistical significance that the strategy is not an exact solution.

82 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

available actions. If the game has exactly two players and only 288 information sets, is intractable having over 1086 the utilities at every leaf sum to zero, the game is called possible deterministic strategies. zero-sum. Sequence-Form Linear Programming. Romanovskii35 The classical solution concept for games is a Nash equi- and later Koller et al.26, 27 established the modern era of solving librium, a strategy for each player such that no player can imperfect information games, introducing the sequence- increase their expected utility by unilaterally choosing a form representation of a strategy. With this simple change of different strategy. All finite extensive-form games have at variables, they showed that the extensive-form game could least one Nash equilibrium. In zero-sum games, all equi- be solved directly as an LP, without the need for an expo- libria have the same expected utilities for the players, and nential conversion to normal-form. Sequence-Form Linear this value is called the game-theoretic value of the game. An Program (SFLP) was the first algorithm to solve imperfect ε-Nash equilibrium is a strategy for each player where no information extensive-form games with computation time player can increase their utility by more than ε by choos- that grows as a polynomial of the size of the game repre- ing a different strategy. By Allis’s categories, a zero-sum sentation. In 2003, Billings et al.5 applied this technique to game is ultra-weakly solved if its game-theoretic value is poker, solving a set of simplifications of HULHE to build the computed, and weakly solved if a Nash equilibrium strat- first competitive poker-playing program. In 2005, Gilpin and egy is computed. We call a game essentially weakly solved Sandholm19 used the approach along with an automated if an ε-Nash equilibrium is computed for a sufficiently technique for finding game symmetries to solve Rhode small ε to be statistically indistinguishable from zero in a Island Hold’em,41 a synthetic poker game with 3.94 × 106 human lifetime of played games. For perfect information information sets after symmetries are removed. games, solving typically involves a (partial) traversal of the Counterfactual Regret Minimization. In 2006, the Annual game tree. However, the same techniques cannot apply Computer Poker Competition was started.53 The competition to imperfect information settings. We briefly review the drove significant advancements in solving larger and larger advances in solving imperfect information games, bench- games, with multiple techniques and refinements being marking the algorithms by their progress in solving increas- proposed in the years that followed.36, 37 One of the tech- ingly larger synthetic poker games as summarized shown niques to emerge, and currently the most widely adopted in in Figure 2. the competition, is Counterfactual Regret Minimization (CFR).d Normal-Form Linear Programming. The earliest method CFR is an iterative method for approximating a Nash equilib- for solving extensive-form games involved converting it rium of an extensive-form game through the process of repeated into a normal-form game, represented as a matrix of val- self-play between two regret-minimizing algorithms.10, 52 ues for every pair of possible deterministic strategies in Regret is the loss in utility an algorithm suffers for not having the original extensive-form game, and then solving it with selected the single best deterministic strategy, which can a Linear Program (LP). Unfortunately, the number of pos- only be known in hindsight. A regret-minimizing algorithm is sible deterministic strategies is exponential in the num- one that guarantees its regret grows sub-linearly over time, ber information sets of the game. So, while LPs can handle and so eventually achieves the same utility as the best deter- normal-form games with many thousands of strategies, ministic strategy. The key insight of CFR is that instead of even just a few dozen decision points makes this method storing and minimizing regret for the exponential number of impractical. Kuhn poker, a poker game with three cards, deterministic strategies, CFR stores and minimizes a modi- one betting round, and a one bet maximum having a total fied regret for each information set and subsequent action, of 12 information sets (see Figure 1), can be solved with this which can be used to form an upper bound on the regret for approach. But even Leduc hold’em,42 with six cards, two any deterministic strategy. An approximate Nash equilib- betting rounds, and a two bet maximum having a total of rium is retrieved by averaging each player’s strategies over all of the iterations, and the approximation improves as the number of iterations increases. The memory needed for the Figure 2. Increasing sizes of imperfect information games solved algorithm is linear in the number of information sets, rather over time. The shaded regions refer to the technique used to achieve than quadratic, which is the case for efficient LP methods.25 the result with references in the main text. CFR+ is the algorithm used in this work and the dashed line shows the result established in Since solving large games is usually memory-bound, CFR this paper. has resulted in as dramatic an increase in the size of solved games as Koller et al.’s advance. Since its introduction in 1015 1014 HULHE 1013 1012 d Another notable algorithm to emerge from the Annual Computer Poker 1011 Competition is an application of Nesterov’s excessive gap technique33 10 10 to solving extensive form games.18 The technique has some desirable 109 Rhode properties, including better asymptotic time complexity that what is known 108 Island for CFR. However, it has not seen widespread use among competition 107 Hold’em 106 participants due to its lack of flexibility in incorporating sampling schemes 5 Game size (Information sets) 10 and its inability to be used with powerful (but unsound) abstractions that SFLPCFR CFR+ 49 104 employ imperfect recall. Recently, Waugh and Bagnell have shown that 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 CFR and the excessive gap technique are more alike than different, suggesting Year that the individual advantages of each approach may be attainable in the other.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 83 research highlights

2007, CFR has been used to solve increasingly complex sim- empirically observed CFR+ to require considerably less com- plifications of HULHE, reaching as many as 3.8 × 1010 infor- putation than state-of-the-art sampling CFR,22 while also mation sets in 2012.20 being highly suitable for massive parallelization.

3. SOLVING HEADS-UP LIMIT HOLD’EM Theorem 1.e Given a set of actions A, and any sequence of T The full game of HULHE has 3.19 × 1014 information sets. value functions vt: A  ℜ with a bound L such that |vt(a) − vt(b)| Even after removing game symmetries it has 1.38 × 1013, that ≤ L for all t and a, b ∈ A, an agent acting according to the regret- is, three orders of magnitude larger than previously solved matching+ algorithm will have regret of at most . games. There are two challenges for established CFR vari- ants to handle games at this scale: memory and computa- Like CFR, CFR+ is an iterative algorithm that computes suc- tion. During computation CFR must store the resulting cessive approximations to a Nash equilibrium solution. The solution and the accumulated regret values for each infor- quality of the approximation can be measured by its exploit- mation set. Even with single-precision (four byte) floating ability: the amount less than the game value that the strategy point numbers, this requires 262TB of storage. Furthermore, achieves against the worst-case opponent strategy in expec- past experience has shown that a three order of magnitude tation.10 Computing the exploitability of a strategy involves increase in the number of information sets requires at least computing this worst-case value, traditionally requiring a three orders of magnitude more computation. In order to traversal of the entire game tree. This was long thought to tackle these two challenges we employ two ideas recently be intractable for games the size of HULHE. Recently it was proposed by Tammelin, a co-author of this paper.44 shown that this calculation could be dramatically acceler- To address the memory challenge we store the approxi- ated by exploiting the imperfect information structure of the mate solution strategy and accumulated regrets using com- game and regularities in the utilities.23 This is the technique pression. For the solution and regrets we use fixed-point we use to confirm the approximation quality of our resulting arithmetic by first multiplying all values by a scaling factor strategy. The technique and implementation has been veri- and truncating them to integers. The resulting integers are fied on small games and against independent calculations then ordered to maximize compression efficiency, with com- of the exploitability of simple strategies in HULHE. pression ratios around 13-to-1. Overall, we require under A strategy can be exploitable in expectation and yet, 11TB of storage during the computation, which is distrib- due to chance elements in the game and randomization uted across a cluster of computation nodes. This amount in the strategy, its worst-case opponent still is not guaran- is in-feasible to store in main memory, and so we store the teed to be winning after any finite number of hands. We compressed strategy and regret values on each node’s local define a game to be essentially solved if a lifetime of play disk. Each node is responsible for a set of subgames, that is unable to statistically differentiate it from being solved is, portions of the game tree partitioned based on publicly at 95% confidence. Imagine someone playing 200 hands observed actions and cards so that each information set is of poker an hour for 12hrs a day without missing a day for associated with one subgame. The regrets and strategy for a 70 years. Furthermore imagine them employing the worst- subgame are loaded from disk, updated, and saved back to case, maximally exploitive, opponent strategy, and never disk, using a streaming compression technique that decom- making a mistake. Their total winnings, as a sum of many presses and recompresses portions of the subgame as millions of independent outcomes, would be normally needed. By making the subgames large enough, the update- distributed. Hence, the observed winnings in this life- time dominates the total time to process a subgame. With time of poker would be 1.64 standard deviations or more disk pre-caching, the inefficiency incurred by disk storage is below its expected value (i.e., the strategy’s exploitability) approximately 5% of the total time. at least 1 time out of 20. Using the standard deviation of To address the computation challenge we use a variant of a single hand of HULHE, which has been reported to be CFR called CFR+.10, 44 CFR implementations typically sample around 5bb/g (big-blinds per game, where the big-blind is only portions of the game tree to update on each iteration. the unit of stakes in HULHE),11 we arrive at a threshold of They also employ regret-matching at each information set, . So, an approximate which maintains regrets for each action and chooses among solution with an exploitability under 1mbb/g (milli-big- actions with positive regret with probability proportional blinds per game) cannot be distinguished with high con- to that regret. Instead, CFR+ does exhaustive iterations over fidence from an exact solution, and indeed has a 1-in-20 the entire game tree, and uses regret-matching+, a vari- chance of winning against its worst-case adversary even ant of regret-matching where regrets are constrained to be after a human lifetime of games. Hence, 1mbb/g is the non-negative. Actions that have appeared poor (with less threshold for declaring HULHE essentially solved. than zero regret for not having been played) will be chosen again immediately after proving useful (rather than waiting 4. THE SOLUTION many iterations for the regret to become positive). Finally, Our CFR+ implementation was executed on a cluster of 200 in contrast with CFR, we have observed empirically that the computation nodes each with 24 2.1GHz AMD cores, 32GB exploitability of the players’ strategies during the computa- tion regularly converges to zero. Therefore, we skip the step e Theorem 1 and others providing the theoretical support for CFR+ did of computing and storing the average strategy, instead using not appear in the original version of this article, and were published in a the players’ current strategies as the CFR+ solution. We have subsequent­ paper.45

84 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

of Random Access Memory (RAM), and a 1TB local disk. strategy caps the betting less than 0.01%, and the hand most We divided the game into 110,565 subgames (partitioned likely to cap is a pair of twos, with probability 0.06%. Perhaps based on preflop betting, flop cards, and flop betting). The more importantly, the strategy chooses to play, that is, not subgames were split among 199 worker nodes, with one fold, a broader range of hands as the non-dealer than most parent node responsible for the initial portion of the game- human players (see the relatively small amount of red in tree. The worker nodes performed their updates in parallel, Figure 4b). It is also much more likely to re-raise when hold- passing values back to the parent node for it to perform its ing a low-rank pair (such as threes or fours).g update, taking 61 min on average to complete one iteration. While these observations are only for one example of The computation was then run for 1,579 iterations, taking game-theoretically optimal play (different Nash equilibria 68.5 days, and using a total of 900 core years of computationf may play differently), they both confirm as well as contradict and 10.9TB of disk space, including file system overhead current human beliefs about equilibria play, and illustrate from the large number of files. that humans can learn considerably from such large-scale Figure 3 shows the exploitability of the computed strat- game-theoretic reasoning. egy with increasing computation. The strategy reaches an exploitability of 0.986mbb/g, making HULHE essentially 5. IN-GAME RESULTS weakly solved. Using the separate exploitability values for In this extended version of the original paper,9 we present each position (as the dealer and non-dealer) we get exact additional results measuring Cepheus’ in-game perfor- bounds on the game-theoretic value of the game: between mance against computer agents and human opponents. 87.7mbb/g and 89.7mbb/g for the dealer, proving the com- HULHE has served as a common testbed for artificial intel- mon wisdom that the dealer holds a significant advantage ligence research for more than a decade, and researchers in HULHE. have produced a long series of computer agents for the The final strategy, as a close approximation to a Nash domain. This effort was largely coordinated by the Annual equilibrium, can also answer some fundamental and long- Computer Poker Competition (ACPC) which begain in 2006 debated questions about game-theoretically optimal play in with HULHE. While each year’s top agents outperformed the HULHE. Figure 4 gives a glimpse of the final strategy in two older agents in the competition, and so appeared to be con- early decisions of the game. Human players have disagreed verging to optimal play, their actual worst-case exploitability about whether it may be desirable to “limp,” that is, call as was unknown. In 2011, an efficient best response technique the very first action rather than raise, with certain hands. was developed that made it feasible to measure a computer Conventional wisdom is that limping forgoes the opportu- agent’s exploitability,23 and for the first time researchers nity to provoke an immediate fold by the opponent, and so were able to exactly measure their progress towards the goal raising is preferred. Our solution emphatically agrees (see of solving the game. A key result in that paper was that top the absence of blue in Figure 4a). The strategy limps just ACPC agents only defeated each other by tiny margins, and 0.06% of the time and with no hand more than 0.5%. In other yet had a wide range of exploitability. Using Cepheus, we can situations, the strategy gives insights beyond conventional now also evaluate these historical agents through matches wisdom, indicating areas where humans might improve. against an essentially optimal strategy. The strategy rarely “caps,” that is, makes the final allowed raise, in the first round as the dealer, whereas some strong g These insights were the result of discussions with Mr. Bryce Paradis, previously human players cap the betting with a wide range of hands. a professional poker player who specialized in HULHE. Even when holding the strongest hand, a pair of aces, the

Figure 4. Action probabilities in the solution strategy for two early f The total time and number of core years is larger than was strictly necessary decisions. Each cell represents one of the possible 169 hands (i.e., as it includes computation of an average strategy that was later measured to two private cards) with the upper diagonal consisting of cards be more exploitable than the current strategy and so discarded. The total with the same suit and the lower diagonal consisting of cards of space noted, on the other hand, is without storing the average strategy. different suits. The color of the cell represents the action taken: red for fold, blue for call, and green for raise, with mixtures of colors representing a stochastic decision. Figure 3. Exploitability of the approximate solution with increasing computation. Suited Suited AKQJT98765432 AKQJT98765432 104 A A K K Q Q 103 J J T T 2 9 9 10 8 8 7 7 Unsuited Unsuited 1 6 6 10 5 5

Exploitability (mbb/g ) 4 4 3 3 100 2 2 10–1 100 101 102 103 (a) First action as the (b) First action as the Computation time (CPU-years) dealer non-dealer after a dealer raise

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 85 research highlights

Figure 5 presents the exploitability of our historical strategy. However, even though many of these strategies were agents and their average loss in games played against highly exploitable, the rate at which they lose to Cepheus is Cepheus. To reduce the impact of luck, a duplicate poker quite low. This loss is difficult to measure with statistical format was used where each game is played twice, using confidence: a 100,000 game (non-duplicate) match would the same cards, but with the players in opposite positions. have a 95% confidence interval of 31mbb/g, larger than the PsOpti4 was the first game theoretic strategy produced for performance difference between Cepheus and every agent HULHE, and was also the University of Alberta entry to but PsOpti4. Further, Hyperborean 2009 did improve over the 2006 ACPC.5, h The University of Alberta entries to the its predecessors in terms of in-game performance against ACPC were named Hyperborean, and from 2007 onwards, Cepheus, and regressed in exploitability due to its use of all were created using variants of CFR.i The Polaris 2007 “Strategy Grafting,” an unsound solving technique that and 2008 agents were created by the University of Alberta solves an abstraction as a series of fragments.50 This tech- for its two Man-vs.-Machine Poker Championship matches, nique allows for a much larger and finer grained abstraction in which Polaris narrowly lost in 2007 and narrowly won in than would otherwise be feasible, resulting in improved in- 2008; an analysis of these matches is available in24 [Chapter game performance, but without theoretical guarantees on 8]. Finally, the CFR-BR agent was our closest equilibrium exploitability. Together, these results illustrate the difficulty approximation prior to this work.21 It used the same abstract in evaluating a strategy only through its competition perfor- game as Hyperborean 2011, but used an algorithm that mance, instead of calculating its exploitability. solved for the abstract strategy with the lowest real game We can also measure Cepheus’ performance against exploitability. human adversaries. After this article was first published in These results show that, with the exception of Hyper­ January 2015, our website allowed visitors to play against borean 2009, each new generation of strategies improved in Cepheus and inspect its strategy.8 Each visitor chose a user- both exploitability and in loss against an essentially optimal name and played any number of short 100-game matches against Cepheus. Over the last two years, 39,564 unique usernames have played 98,040 matches, with 3,564,094 total h PsOpti4 acts too slowly for an exploitability calculation to be practical, games played.j Over this set of games, Cepheus is winning or for a long match against Cepheus. i In the inaugural 2006 ACPC, PsOpti4 was the core component of at a rate of 169.9 ± 5.2mbb/g with 95% confidence. However, Hyperborean 2006. most of the players did not finish a single 100-game match (only 7,878 players did so, with 20,374 completed matches in total), and so this winrate is likely not reflective of Cepheus’ Figure 5. Exploitability and performance against Cepheus for earlier performance against strong opponents. computer strategies. Results are in mbb/g, and indicate the expected winnings by the strategy’s opponent (a best response or Cepheus, Determining which of these players are strong is non- respectively). The Cepheus matches involved 1mn games of trivial because of both variance in their matches, and the duplicate poker (2mn games total), except for PsOpti4 which played unequal amount of games played by each player. While both 20,000 duplicate games (40,000 games total). luck and skill contribute to a player’s performance, the high- 500 est-scoring players are more likely to be the luckiest rather than the strongest. Additionally, bias may be introduced if 400 players keep playing while ahead, but quit if they are losing. vs best response In order to limit the impact of bias and evaluate Cepheus’ 300 performance against different tiers of humans, we used 200 the following method. First, we eliminated usernames with insufficient data that had played fewer than 500 games, leav- 100 In-game loss (mbb/g) ing 821 usernames playing 33,752 matches with 1,765,656 vs cepheus 0 games. Next, we divided each username’s games into two sets, called Rank and Test.k Each username’s Rank games were evaluated, and the resulting winrates were used to sort

Hyper-07 Hyper-08 Hyper-09 Hyper-10 Hyper-11 Polaris-07 Polaris-08 the players by performance. This ordering reflected both PsOpti4-03 CFR-BR-12 Cepheus-14 Name Year Exploitability Cepheus their skill and luck. The players were then divided equally into five tiers: the bottom 20% of usernames, 21–40% etc. PsOpti4 2003 – 74.9 ± 23.7 Within each tier, the Test game results were averaged to pro- Hyperborean 2007 2007 298.106 27.4 ± 2.9 duce a winrate for the tier, independent from the luck that Polaris 2007 2007 275.880 26.2 ± 3.0 affected the Rank games. Hyperborean 2008 2008 266.797 22.5 ± 2.7 These results are shown in Figure 6. Cepheus’ esti- Polaris 2008 2008 235.294 22.2 ± 2.6 mated winrate varies from 225 to 87mbb/g as we advance Hyperborean 2009 2009 440.823 18.9 ± 2.6 through the tiers, decreasing as the quality of the human Hyperborean 2010 2010 135.427 10.8 ± 2.5 Hyperborean 2011 2011 106.035 8.0 ± 2.4 CFR-BR 2012 37.113 9.2 ± 2.6 Cepheus 2014 0.986 0 j Many players quit before finishing the 100-game match. k In each block of four sequential games, one pair (played in each position) was assigned to each set.

86 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

Figure 6. Games played by humans and Cepheus. Humans were Centre for Machine Learning (AICML), and was only possible divided equally into five skill groups, and the column and error bar due to computing resources provided by Compute Canada indicates the group’s average loss to Cepheus in mbb/g. and Calcul Québec. The authors would like to thank all of

300 the current and past members of the University of Alberta Computer Poker Research Group (CPRG), where the idea 250 to solve heads-up limit Texas hold’em was first discussed; Jonathan Schaeffer, Robert Holte, Duane Szafron, and 200 Alex Brown for comments on early drafts of this article; and 150 Bryce Paradis for insights into the conventional wisdom of top human poker players. 100

Cepheus winnings (mbb/g) 50 225195 119 106 87 References 0 1. Poker: A big deal. The Economist, 20. Jackson, E. Slumbot: An 0–19% 20–39% 40–59% 60–79% 80–100% Economist Newspaper Ltd., London, implementation of counterfactual December 22, 31–38, 2007. regret minimization on commodity Tier of humans 2. Allis, V.L. A Knowledge-Based hardware. In Proceedings of the 2012 Approach to Connect-Four. The Game Computer Poker Symposium. (2012). is Solved: White Wins. Master’s thesis, 21. Johanson, M., Bard, N. Burch, N., Vrije Universiteit, Amsterdam, The Bowling, M. Finding optimal abstract Netherlands, 1988. strategies in extensive form games. players improves. Even against the top 20% tier of play- 3. Allis, V.L. Searching for Solutions in AAAI, (2012), 1371–1379. ers in this experiment, Cepheus’ winrate of 87mbb/g Games and Artificial Intelligence. PhD 22. Johanson, M., Bard, N., Lanctot, M., thesis, Vrije Universiteit Amsterdam, Gibson, R., Bowling, M. Efficient Nash is higher than against any of our historical agents. It The Netherlands, 1994. equilibrium approximation through even exceeds 50mbb/g, a commonly cited benchmark 4. Babbage, C. Passages from the Life Monte Carlo counterfactual regret of a Philosopher. Longman, Green, minimization. AAMAS (2012). for what a professional poker player seeks to win from a Longman, Roberts, and Green, 23. Johanson, M., Waugh, K., Bowling, M., weaker opponent. London, 1864. Chapter 34. Zinkevich, M. Accelerating best 5. Billings, D., Burch, N., Davidson, A., response calculation in large extensive Holte, R.C., Schaeffer, J., Schauenberg, T., games. IJCAI (2011), 258–265. 6. CONCLUSION Szafron, D. Approximating game- 24. Johanson, M.B. Robust Strategies theoretic optimal strategies for full- and Counter-Strategies: From In this paper, we announced that heads-up limit Texas scale poker. IJCAI, (2003), 661–668. Superhuman to Optimal Play. 6. Billings, D., Davidson, A., Schaeffer, J., PhD thesis, University of Alberta, hold’em poker is essentially weakly solved. This is the first Szafron, D. The challenge of poker. Edmonton, Alberta, Canada, 2016. nontrivial imperfect information game played competi- Artificial Intelligence 134, 1–2 (2002), 25. Karmarkar, N. A new polynomial- 201–240. time algorithm for linear tively by humans to be solved. Even still, the reader may 7. Borel, E., Ville, J. Applications de la programming. In Proceedings of the ask what is the ultimate significance of solving poker? The théorie des probabilités aux jeux de Sixteenth Annual ACM Symposium hasard. Gauthier-Villars, 1938. on Theory of Computing (1984), breakthroughs behind this result are general algorithmic 8. Bowling, M., Burch, N., Johanson, M., ACM, New York, NY, 302–311. advances that make game-theoretic reasoning in large-scale Tammelin, O. The cepheus website, 26. Koller, D., Megiddo, N. The complexity 2015. http://poker.srv.ualberta.ca. of two-person zero-sum games in models of any sort more tractable. And, while seemingly 9. Bowling, M., Burch, N., Johanson, M., extensive form. Games Econ. Behav 4, playful, game theory has always been envisioned to have Tammelin, O. Heads-up limit hold’em 4 (1992), 528–552. poker is solved. Science 347, 6218 27. Koller, D., Megiddo, N., von Stengel, B. serious implications, for example, its early impact on cold (Jan. 2015), 145–149. Efficient computation of equilibria for war politics.31 More recently, there has been a surge in game- 10. Bowling, M., Burch, N., Johanson, M., extensive two-person games. Games Tammelin, O. Heads-up limit hold’em Econ. Behav 14, 2 (1996). theoretic applications involving security, including systems poker is solved: Supplementary online 28. Koller, D., Pfeffer, A. Representations being deployed for airport checkpoints, air marshall sched- material, January 2015. and solutions for game-theoretic 11. Bowling, M., Johanson, M., Burch, N., problems. Artificial Intelligence 94, 43 uling, and coast guard patrolling. CFR algorithms, based Szafron, D. Strategy evaluation in (1997), 167–215. extensive games with importance 29. Kuhn, H. Simplified two-person on those described in this paper, have been used for robust sampling. ICML, (2008), 72–79. poker. In Contributions to the Theory decision-making in settings where there is no apparent 12. Bronowski, J. Ascent of man. of Games, volume 1 of Annals of Documentary, 1973. Episode 13. mathematics studies. H. Kuhn and A. adversary, with potential application to medical decision 13. Buro, M. Solving the oshi-zumo game. Tucker, eds. Press, support.15 With real life decision-making settings almost Adv. in Comp. Games 135, (2004) Princeton, New Jersey, 1950, 97–103. 361–366. 30. Mirowski, P. What were von always involving uncertainty and missing information, algo- 14. Campbell, M., Hoane, A.J. Jr., Hsu, F. neumann and morgenstern trying rithmic advances, such as those needed to solve poker, are Deep blue. Artificial Intelligence 134, to accomplish? In Toward a (Jan. 2002), 57–83. History of Game Theory. Weintraub, needed to drive future applications. However, we also echo 15. Chen, k., Bowling, M. Tractable ed. Duke University Press, 1992, a response attributed to Alan Turing in defense of his own objectives for robust policy 113–147. Mirowski cites Turing as optimization. Adv. Neural Inf. Process. author of the paragraph containing work in games, “It would be disingenuous of us to disguise Syst. (NIPS) 25, (2012), 2078–2086. this remark. The paragraph the fact that the principal motive which prompted the work 16. Craig, M. The Professor, the Banker, appeared in [46], in a chapter 30 and the Suicide King: Inside the with Turing listed as one of three was the sheer fun of the thing.” Richest Poker Game of All Time. contributors. Which parts of the Grand Central Publishing, New York, chapter are the work of which NY, 2006. contributor, particularly the Acknowledgments 17. Ferrucci, D. Introduction to “this is introductory material containing The author order is alphabetical reflecting equal contribu- watson.” IBM J. Res. Dev 56, 3.4 (May this quote, is not made explicit. 2012) 1:1–1:15. 31. Morgenstern, O. The cold war is cold + tion by the authors. The idea of CFR and compressing the 18. Gilpin, A., Hoda, S., Peña, J., Sandholm, T. poker. N. Y. Times Mag. (Feb. 5 1961) 44 Gradient-based algorithms for finding pages 21–22. regrets and strategy originated with Oskari Tammelin. nash equilibria in extensive form 32. Nash, J.F., Shapley, L.S. A simple This research was supported by Natural Sciences and games. WINE, (2007), 57–69. 3-person poker game. In Contributions 19. Gilpin, A., Sandholm, T. Lossless to the Theory of Games I. Princeton Engineering Research Council (NSERC), Alberta Innovates abstraction of imperfect information University Press, Princeton, New Technology Futures (AITF) through the Alberta Innovates games. J. ACM 54, 5 (2007). Jersey, 1950, 105–116.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 87 research highlights

33. Nesterov, Y. Excessive gap technique Comp. Games, (2000), 333–345. on Computer Poker and Imperfect Congress Mathematics. Cambridge in nonsmooth convex minimization. 42. Southey, F., Bowling, M., Larson, B., Information, 2015. University Press, Cambridge, SIAM Journal on Optimization 16, 1 Piccione, C., Burch, N., Billings, D., 50. Waugh, K., Bard, N., Bowling, M. 1913, 501–504. (2005), 233–249. Rayner, D.C. Bayes’ bluff: Opponent Strategy grafting in extensive games. 52. Zinkevich, M., Johanson, M., Bowling, M., 34. Rehmeyer, J., Fox, N., Rico, R. Ante modelling in poker. UAI, (2005) In Advances in Neural Information Piccione, C. Regret minimization in up, human: The adventures of polaris 550–558. Processing Systems 22 (NIPS-09), games with incomplete information. the poker-playing robot. Wired 16, 12 43. Tambe, M. Security and Game Theory: 2009. http://webdocs.cs.ualberta. NIPS (2008), 905–912. (Dec. 2008), 186–191. Algorithms, Deployed Systems, ca/~games/poker/publications/ 53. Zinkevich, M., Littman, M. The AAAI 35. Romanovskii, I.V. Reduction of a game Lessons Learned. Cambridge University NIPS09-graft.pdf. computer poker competition. J. Inter. with complete memory to a matrix Press, Cambridge, England, 2011. 51. Zermelo, E. Über eine Anwendung Comp. Games Association 29, (2006), game. Soviet Mathematics 3, (1962), 44. Tammelin, O. Cfr+. CoRR, der Mengenlehre auf die Theorie News item. 678–681. abs/1407.5042, 2014. des Schachspiels. In Proceedings 36. Rubin, J., Watson, I. Computer 45. Tammelin, O., Burch, N., Johanson, M., of the Fifth International poker: a review. Artificial Intelligence Bowling, M. Solving heads-up limit 175,(2011), 958–987. texas hold’em. In Proceedings of the 37. Sandholm, T. The state of solving 24th International Joint Conference large incomplete-information games, on Artificial Intelligence, 2015, and application to poker. AI Mag. 31, 645–652. 4 (2010), 13–32. 46. Turing, A. Digital computers applied Michael Bowling ([email protected]), Oskari Tammelin ({[email protected]} {http:// 38. Schaeffer, J., Lake, R., Lu, P., Bryant, M. to games. In Faster Than Thought. Department of Computing, Science, jeskola.net}). Chinook the world man-machine B.V. Bowden, ed. Chapter 25. Pitman, University of Alberta, Edmonton, Alberta, checkers champion. AI Mag. 17, 1 1976. Canada. (1996), 21–29. 47. von Neumann, J. Zur theorie der 39. Schaeffer, J., Neil Burch, Y.B., gesellschaftsspiele. Mathematische Neil Burch and Michael Johanson Kishimoto, A., Müller, M., Lake, R., Annalen 100, 1 (1928), 295–320. ({nburch, johanson}@ualberta.ca), Lu, P. Sutphen, S. Checkers is solved. 48. von Neumann, J., Morgenstern, O. Department of Computing, Science, Science 317, 5844 (2007), 1518–1522. Theory of Games and Economic University of Alberta, Edmonton, Alberta, 40. Shannon, C.E. Programming Behavior. Princeton University Canada. a computer for playing chess. Press, Princeton, Second Edition, Philosophical Magazine, Series 7, 41, 1947. 314 (Mar. 1950), 256–275. 49. Waugh, K., Bagnell, J.A. A unified view 41. Shi, J., Littman, M.L. Abstraction of large-scale zero-sum equilibrium methods for game theoretic poker. In computation. In AAAI Workshop © 2017 ACM 0001-0782/17/11 $15.00

Plato and The Nerd The Creative Partnership of Humans and Technology Edward Ashford Lee How humans and technology evolve together in a creative partnership.

Foolproof, and Other Mathematical Meditations Brian Hayes A non-mathematician explores mathematical terrain, reporting accessibly and engagingly on topics from Sudoku to probability.

Robot Sex Social and Ethical Implications edited by John Danaher and Neil McArthur Perspectives from philosophy, psychology, religious studies, economics, and law on the possible future of robot-human sexual relationships.

mitpress.mit.edu/acm

88 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 DOI:10.1145/3131282 To view the accompanying paper, Technical Perspective visit doi.acm.org/10.1145/3131280 rh Exploring a Kingdom by Geodesic Measures By Marc Alexa

ONCE UPON A time, there was a particu- of finding a path, traveling only along fewer stones will result in the desired larly generous king. He decided to give the edges of the triangles will not geodesic path. Or, in more recent lan- all of his people a share of his land. yield the shortest possible path. A bet- guage, the gradient of the probability Any family owning a house would have ter option is tracking how a wavefront density function of a random walk is to mark the center of their land with a emanates from a source vertex, the parallel to geodesics. stone. Then the ruling said that every so-called fast marching method. Now, why is this observation useful point they could reach within 1,000 Unfortunately, this method may fail for computation? The density of ran- steps from the mark stone would be- to provide shortest paths for certain dom walks is related to heat diffusion, long to them. surfaces and shapes of triangles. It and heat diffusion is governed by a Initially, the king was praised for turns out the problem of shortest paths first-order partial differential equa- the simplicity and clarity of his rul- along a piecewise linear surface has a tion, which can be well approximated ing. Yet, as most new landowners de- superquadratic worst-case complex- by solving a sparse linear system. The cided to build a fence around their ity, and algorithms establishing this important insight from the authors property they soon realized that deter- complexity are surprisingly compli- is that a wide range of linear opera- mining the location of the fence was cated. The most recent algorithms of tors can be used to compute functions frustratingly difficult. Any point that this kind have much better asymptotic with the desired gradient directions was reached after walking 1,000 steps complexity in common practical cas- (yet different gradient lengths) was part of the property, sure; but es, yet they are still significantly more Given any such function, distances how would one have to walk to reach complex than a simple graph traversal can be computed by simply normaliz- a point that was as far away from the such as Dijkstra’s. ing the gradients and then integrat- stone as possible? The approach by Crane, Weisch- ing them. Integration amounts to Luckier landowners lived in flat re- edel, and Wardetzky in the follow- solving another sparse linear system. gions, where walking along a straight ing paper may be related to what the Together this means geodesic dis- line (for example, walking toward any landowners could have observed and tances can be computed by solving landmark further away than 1,000 exploited: assume one would start two (related) sparse linear systems, steps) would always reach a point from the mark stone, walk 1,000 ran- plus computing the gradients of the on the boundary of the land. Land- dom steps, and mark the endpoint of intermediate function and normaliz- owners in hilly regions had a prob- this walk with a pebble. Over many ing them. Note how the non-linearity lem. They needed to find a path that repetitions of this exercise a distribu- of the problem is pushed into the was the equivalent of a straight line. tion of pebbles on the land emerges: trivial step of normalizing a set of vec- Such paths are now known as geode- closer to the mark stone the pat- tors, while all global computations sics—the name indeed relating to the tern is denser, and fewer and fewer are linear. problem of measuring the earth. In pebbles are found further away. As it And lastly, turning the problem of today’s language, the land assigned turns out, starting at the mark stone computing geodesic paths to the solu- to each family is a geodesic disk. And and always walking in the direction of tion of a sparse linear system has an- it should come as no surprise that other very desirable feature: the most the distance between two points in time-consuming part lies in the step curved domains is a fundamental How would we of factorizing the system matrix. This problem with numerous applications factorization can be reused for arbi- in science and engineering. solve the problem trary distance computations on the So, how would we solve the prob- of computing domain. So all the king would have lem of computing geodesic paths, needed were the triangular factors of distances, and circles on a digital geodesic paths, a linear system, and then each geode- computer? Let’s assume the smooth distances, and sic circle could have been computed terrain is approximated as a triangu- almost instantaneously. lation. The shortest path from a ver- circles on a tex of this triangulation to all other Marc Alexa is a professor in the Faculty of digital computer? and Computer Science at vertices along the edges of the triangu- the Technical University of Berlin, and heads lation can be computed using Dijks- the Computer Graphics group. tra’s algorithm. While this provides an efficient solution to the problem Copyright held by author.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 89 research highlights

DOI:10.1145/3131280 The Heat Method for Distance Computation By Keenan Crane, Clarisse Weischedel, and Max Wardetzky

Abstract Figure 1. In contrast to algorithms that compute shortest paths along We introduce the heat method for solving the single- or a graph (left), the heat method computes the distance to points on a multiple-source on both flat and curved continuous, curved domain (right). A key advantage of this method domains. A key insight is that distance computation can be is that it is based on sparse linear equations that can be efficiently split into two stages: first find the direction along which dis- prefactored, leading to dramatically reduced amortized cost. tance is increasing, then compute the distance itself. The heat method is robust, efficient, and simple to implement since it is based on solving a pair of standard sparse linear systems. These systems can be factored once and subse- quently solved in near-linear time, substantially reducing amortized cost. Real-world performance is an order of magnitude faster than state-of-the-art methods, while maintaining a comparable level of accuracy. The method can be applied in any dimension, and on any domain that admits a gradient and inner product—including regular grids, triangle meshes, and point clouds. Numerical evi- dence indicates that the method converges to the exact dis- tance in the limit of refinement; we also explore smoothed approximations of distance suitable for applications where greater regularity is desired. measures the length along shortest and straightest curves through the domain rather than straight lines through space. Our key observation is that one can decompose dis- 1. INTRODUCTION tance computation into two stages: first determine the The multiple-source shortest path problem seeks the distance direction along which distance increases, then recover from each point of a domain to the closest point within a the distance itself. Moreover, since each stage amounts given subset; different versions of this problem are funda- to a standard problem in numerical linear algebra, one mental to a wide array of problems across computer science can leverage existing algorithms and software to improve and computational mathematics. Solutions date back at the efficiency and robustness of distance computation. least to Dantzig’s work on linear programs35; typically the Although this approach can in principle be used in the problem is formulated in terms of a weighted graph, as in context of graph distance, its real utility lies in approxi- Dijkstra’s algorithm. Often, however, one wishes to cap- mating the distance on continuous, curved domains. This ture the distance on a continuous domain; a key example approach has proven effective for a diverse range of appli- is illustrated in Figure 1 (left) where the graph distance will cations in computational neuroscience, geometric model- overestimate the straight-line Euclidean distance, no matter ing, medical imaging, computational design, and machine how fine the grid becomes. In 2D, an important development learning (Section 2), and has recently inspired more accu- was the formulation of “exact” algorithms, where paths rate variations of our original method.3 can cut through the faces of a triangulation8, 27; a great deal of subsequent work has focused on making these O(n2) 1.1. Formulation algorithms practical for large datasets.40,46 However, for Imagine touching a scorching hot needle to a single point problems in data analysis and scientific computing it is not on a surface. Over time heat spreads out over the rest of the

clear that the cost and complexity of exact algorithms are domain and can be described by a function kt, x(y) called the always well-justified, since the triangulation itself is only an heat kernel, which measures the heat transferred from a approximation of the true domain (see Figure 4). source x to a destination y after time t. A well-known rela- A very different approach is to formulate the problem tionship between heat and distance is Varadhan’s formula,42 in terms of partial differential equations (PDEs), where domain approximation error can be understood via, for example, traditional finite element analysis. However, the particular choice of continuous formulation has a sub- stantial impact on computation. The heat method was The original version of this paper was published in inspired by S.R.S. Varadhan’s classic result in differential ACM Transactions on Graphics 32, 5 (Sept. 2013). geometry42 relating heat diffusion and geodesic distance, which

90 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

Figure 2. Given an exact reconstruction of the heat kernel (top type of geometric discretization, including regular grids, left) Varadhan’s formula can be used to recover geodesic distance polygonal meshes, and point clouds. Second, it involves (bottom left) but fails in the presence of approximation or numerical only sparse linear systems, which can be prefactored once error (middle, right), as shown here for a point source in 1D. The and rapidly resolved many times—this feature substantially robustness of the heat method stems from the fact that it depends reduces the amortized cost for applications that require only on the direction of the gradient. repeated distance queries on a fixed geometric domain. Moreover, because the heat method is built on standard lin-

t ear PDEs that are widespread in scientific computing, it can k immediately take advantage of new developments in numer- ical linear algebra and parallelization. t 2. RELATED WORK The prevailing approach to distance computation is to solve –4 t log k the eikonal equation

(2)

which says that the distance φ between any pair of points x, y subject to boundary conditions φ|γ = 0 over some subset γ on a curved domain can be recovered via a simple pointwise of the domain (like a point or a curve). Intuitively, this equa- transformation of the heat kernel: tion says something very simple: as we move away from the source, the distance function φ must change at a rate of “one (1) meter per meter.” Computationally, however, this formula- tion is nonlinear and hyperbolic, making it difficult to solve The intuition behind this behavior stems from the fact that directly. Typically one applies an iterative relaxation scheme heat diffusion can be modeled as a large collection of hot such as Gauss-Seidel—special update orders are known as particles taking random walks starting at x: any particle that fast marching and fast sweeping, which are some of the most reaches a distant point y after a small time t has had little popular algorithms for distance computation on regular time to deviate from the shortest possible path. Previously, grids37 and triangulated surfaces.19 These algorithms can also however, this relationship had not been exploited by numer- be used on implicit surfaces,25 point clouds,26 and polygon ical algorithms that compute distance. soup,7 but only indirectly: distance is computed on a simplicial Why had Varadhan’s formula been overlooked in this con- mesh or regular grid that approximates the original domain. text? The main reason, perhaps, is that it requires a precise Implementation of fast marching on simplicial grids is chal- numerical reconstruction of the heat kernel, which is diffi- lenging due to the need for nonobtuse triangulations (which cult to obtain—applying the formula to a mere approxima- are notoriously difficult to obtain) or else an iterative unfold- tion of kt,x does not yield the correct result, as illustrated in ing procedure that preserves monotonicity of the solution; Figures 2 and 8. The heat method circumvents this issue by moreover these issues are not well-studied in dimensions working with a broader class of inputs, namely any function greater than two. Fast marching and fast sweeping have whose gradient is parallel to the gradient of the true distance asymptotic complexity of O(n log n) and O(n), respectively, function. We can then separate computation into two stages: first find the gradient, then recover the distance itself. Relative to existing algorithms, the heat method offers two major advantages. First, it can be applied to virtually any Figure 4. Convergence of distance approximations on the unit sphere with respect to mean edge length; as a baseline for comparison, we use the analytical solution φ (x, y) = cos−1(x • y). Notice that even with a nice tessellation, the exact distance along the polyhedron Figure 3. The heat method computes the shortest distance to a converges only quadratically to the true distance along the sphere subset γ of a given domain. Gray curves indicate isolines of the it approximates. (Linear and quadratic convergence are plotted as distance function. dashed lines for reference.)

∞ L error Heat method 0.050

Fast marching

0.010 Exact polyhedral 0.005 γ 0.001

–4 5x10

–4 1x10

0.030 0.050 0.070 0.100 0.150 0.200 0.300 Mean edge length (h)

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 91 research highlights

but sweeping is often slower due to the large number of Figure 5. The heat method has been applied to a diverse range of tasks 16 sweeps required to obtain accurate results. that demand repeated geodesic distance queries. Here, geodesic One drawback of these methods is that they do not reuse distance drives a differential growth model (left) that is used for information: the distance to different source sets γ must be computational design (right). Images courtesy Nervous System/Jesse computed entirely from scratch each time. Also note that Louis-Rosenberg. both sweeping and marching present challenges for paral- lelization: priority queues are inherently serial, and irregular meshes lack a natural sweeping order. In a different development, Mitchell et al.27 give an O(n2 log n) algorithm for computing the exact polyhedral distance from a single source to all other vertices of a tri- angulated surface. Surazhsky et al.40 demonstrate that this algorithm tends to run in sub-quadratic time in practice, and present an approximate O(n log n) version of the algo- rithm with guaranteed error bounds; Bommes and Kobbelt4 extend the algorithm to polygonal sources. Similar to fast marching, these algorithms propagate distance informa- tion in wavefront order using a , again making them difficult to parallelize. More importantly, the amor- iterating the heat method, or by applying more sophisti- tized cost of these algorithms (over many different source cated descent strategies. subsets γ) is substantially greater than for the heat method since they do not reuse information from one subset to the 3. THE HEAT METHOD next. Finally, although40 greatly simplifies the original for- A useful feature of the heat method is that the basic mulation, these algorithms remain challenging to imple- algorithm can be described in the purely continuous setting ment and do not immediately generalize to domains other (i.e., in terms of curved surfaces, or more generally, smooth than triangle meshes. manifolds) rather than in terms of discrete data structures and Closest to our approach is the recent method of Rangarajan algorithms. In other words, at this point one should not imag- and Gurumoorthy,32 who do not appear to be aware of ine that we have chosen a particular data structure (triangle Varadahn’s formula—they instead derive an analogous rela- meshes, grids, point clouds, etc.) or even dimension (2D, 3D, tionship between the distance function and solutions ψ to etc.). Instead, we focus on a general principle that can be the time-independent Schrödinger equation; this derivation applied on many different domains in different dimensions. applies only in flat Euclidean space rather than general curved We will later explore several particular choices of spatial and domains. Moreover, they compute solutions using the fast temporal discretization (Sections 3.1 and 3.2); further alter- Fourier transform, which limits computation to regular grids. natives have been explored in recent literature.13, 29, 45 A slight modification of the heat method allows us to com- In general, the heat method can be applied in any setting pute a smoothed distance function, useful in contexts where where one has a gradient operator Ñ, divergence operator Ñ, and sharp discontinuities can cause subsequent numerical dif- Laplace operator D = Ñ × Ñ—standard derivatives from vector cal- ficulties. Previous smooth distance approximations provide culus, possibly generalized to curved domains. Expressed in terms this regularity at the cost of poor approximation of the true of these operators, the heat method consists of three basic steps: geometric length10, 14, 21, 33; see Section 3.3 for a comparison. Recently, the heat method has facilitated a variety of tasks Algorithm 1 The Heat Method in computational science and data analysis. For example, I. Integrate the heat flow u˙ = Du for some fixed time t. Huth et al.15 use fast distance queries to optimize a proba- 41 II. Evaluate the vector field X = −Ñut/|Ñut|. bilistic model of cortical organization; van Pelt et al. use III. Solve the Poisson equation Dφ = Ñ × X. the heat method to assist cerebral aneurysm assessment; Zou et al.47 use the heat method for efficient tool path plan- ning; Solomon et al.38 leverage our approach to efficiently The function φ approximates the distance to a given source solve optimal transport problems on geometric domains; set, approaching the true distance as t goes to zero (Equation 1). Lin et al.20 apply this approach to vector-valued data in the For instance, to recover the distance to a single point x we

context of manifold learning. Figure 5 shows a real-world use initial conditions u0 = δ(x), that is, a Dirac delta encod- design application of the heat method based on differen- ing an “infinite spike” of heat. More generally we can obtain

tial growth. Various improvements have also been made the distance to any subset γ by letting u0 be a generalized to the original algorithm; for instance, de Goes et al.13 and Dirac distribution42—essentially an indicator function over Yang and Cohen45 describe two different ways to extend the γ; see Figures 3 and 7. Note that since the solution to (III) is method to accurate computation of anisotropic distance; it determined only up to an additive constant, final values are has also been adapted to voxelizations6, C1 finite elements,29 shifted such that the smallest distance is zero. and subdivision surfaces.12 Finally, Belyaev and Fayolle3 The heat method can be motivated as follows. Consider

provide a variational interpretation of our method, observ- an approximation ut of heat flow for a fixed time t. Unless

ing that more accurate results can be obtained by either ut exhibits precisely the right rate of decay, Varadhan’s

92 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

Figure 6. The three steps of the heat method. (I) Heat u is allowed to Figure 8. Left: Varadhan’s formula. Right: the heat method. Even for diffuse for a brief period of time (left). (II) The temperature gradient very small values of t, simply applying Varadhan’s formula does not Du (center left) is normalized and negated to get a unit vector field provide an accurate approximation of geodesic distance (top left); for X (center right) pointing along geodesics. (III) A function γ whose large values of t spacing becomes even more uneven (bottom left). gradient follows X recovers the final distance (right). Normalizing the gradient results in a more accurate solution, as indicated by evenly spaced isolines (top right), and is also valuable when constructing a smoothed distance function (bottom right).

IIu u IIX II φ

Figure 7. Distance to the boundary on a region in the plane (left) or a surface in space (right) is achieved by simply placing heat along the boundary curve.

(id – t ∆)υt = 0 on M\γ,

υt = 1 on γ⋅ (4)

which for a point source yields a solution υt equal to ut up to transformation will yield a poor approxima- a multiplicative constant. As established by Varadhan in his tion of the true geodesic distance φ because it is highly sen- proof of Equation (1), υt also has a close relationship with sitive to errors in magnitude (see Figures 2 and 8). The heat distance, namely method asks for something different: it requires only that the gradient ∇ut point in the right direction, that is, paral- (5) lel to ∇φ. Magnitude can safely be ignored since we know (from the eikonal equation) that the gradient of the true dis- This relationship ensures the validity of steps II and III since tance function has unit length. We therefore compute the the transformation applied to υt preserves the direction of normalized gradient field X = −∇ut/|∇ut| and find the closest the gradient. 2 scalar potential φ by minimizing ∫M|∇φ − X| , or equivalently, by solving the corresponding Euler-Lagrange equations 3.2. Spatial discretization ∆φ = ∇ ⋅ X.36 The overall procedure is depicted in Figure 6. Here we detail several possible implementations of the heat This procedure is used as the starting point for a family method on triangle meshes, polygon meshes, and point clouds. of discrete algorithms, as outlined in Sections 3.1–3.3. Note Note that the heat method can also be used on flat Euclidean that some details have been omitted from this manuscript, domains of any dimension by simply applying standard finite and can be found in Crane et al.11 differences on a regular grid; Belyaev and Fayolle3 outline implementation on tetrahedral (3D) meshes. 3.1. Time discretization Triangle meshes. Let u ∈ R|V| specify a piecewise linear To translate our continuous procedure (Algorithm 1) into a function on a triangulated surface with vertices V, edges E, discrete algorithm, we must replace derivatives in space and and faces F. A standard discretization of the Laplacian at a ver- time with suitable approximations. The heat equation from tex i is given by step I of Algorithm 1 can be discretized in time using a single backward Euler step for some fixed time t—in practice, this i means we simply solve the linear equation α β where Ai is one third the area of all ij ij

(id – t∆)ut = u0, (3) triangles incident on vertex i, the sum is taken over all neighbor- j over the entire domain M, where id denotes the identity ing vertices j, and αβij, βij are the operator. Note that at this point we still have not discretized angles opposing the corresponding edge.23 We can −1 space; spatial discretization is discussed in Section 3.2. We express this operation via a matrix L = M LC, where can get a better understanding of solutions to Equation (3) M ∈ R |V|×|V| is a diagonal matrix containing the vertex |V|×|V| by considering the elliptic boundary value problem areas and LC ∈ R is the cotan operator representing the

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 93 research highlights

Figure 9. Since the heat method is based on well-established remaining sum. Heat flow can then be computed by solving discrete operators like the Laplacian, it is easy to adapt to a variety the symmetric positive-definite system ui of geometric domains. Above: distance on a hippo composed of high- degree nonplanar (and sometimes nonconvex) polygonal faces. Ni (M –tLC)u = δγ, where δγ is a Kronecker delta (or indicator func- tion) over γ. The gradient in a given triangle can ei be expressed succinctly as

e2 θ1 X j θ2 i where Af is the area of the triangle, N is its e1

­outward unit normal, ei is the ith edge vector

(oriented counter-clockwise), and ui is the value of u at the opposing vertex. The integrated divergence associated with vertex i can be written as to note that the cost of the heat method depends primarily on the intrinsic dimension n of M, whereas methods based on fast marching require a grid of the same dimension m where the sum is taken over incident triangles j each with a as the ambient space25—this distinction is especially im-

vector Xj, e1, and e2 are the two edge vectors of triangle j con- portant in contexts like machine learning where m may be |V| taining i, and θ1, θ2 are the opposing angles. If we let b ∈ R significantly larger than n. be the vector of (integrated) divergences of the normalized Choice of time step. Accuracy of the heat method relies vector field X, then the final distance function is computed in part on the time step t. In the smooth setting, Equation (5) by solving the symmetric Poisson problem suggests that smaller values of t yield better approxima- tions of geodesic distance. In the discrete setting we in- stead observe the somewhat surprising behavior that the limit solution to Equation (3) depends only on the number As noted in Section 3.1, the solution to step I is a function of edges between a pair of vertices, independent of how we that decays exponentially with distance. Fortunately, nor- might try to incorporate edge lengths into our formulation— malization of small values is not a problem because floating see Crane et al.,11 Appendix A. Therefore, on a fixed mesh point division involves only arithmetic on integer exponents; decreasing the value of t does not necessarily improve accu- likewise, the large range of magnitudes does not adversely racy, even in exact arithmetic—to improve accuracy we must affect accuracy because gradient calculation is local. For simultaneously refine the mesh and decrease t accordingly. the calculation of phi itself we advocate the use of a direct Moreover, very large values of t produce an over-smoothed (Cholesky) solver in double precision; empirically we observe approximation of geodesic distance (Section 3.3). For a fixed roughly uniform pointwise relative error across the domain. mesh, we therefore seek an optimal time step t* that is neither Polygon meshes. Curved surfaces are often described too large nor too small. by polygons that are neither planar nor convex; although An optimal value of t* is difficult to obtain due to the com- such polygons can of course be triangulated, doing so plexity of analysis involving the cut locus.28 We instead use a can adversely affect an existing computational pipeline. simple estimate that works well in practice, namely t = mh2 We instead leverage the polygonal Laplacian of Alexa and where h is the mean spacing between adjacent nodes and Wardetzky1 to implement the heat method directly on po- m > 0 is a constant. This estimate is motivated by the fact that lygonal meshes—the only challenge in this setting is that for h2∆ is invariant with respect to scale and refinement; numeri- nonplanar polygons the gradient vector no longer has a clear cal experiments suggest that m = 1 yields near-optimal accuracy geometric meaning. This issue is resolved by noting that we for a wide variety of problems. In this paper the time step need only the magnitude |∇u| of the gradient; see Crane et al,11 Section 3.2.2 for further details. Figure 9 demonstrates distance computed on an irregular polygonal mesh. Point clouds. Raw geometric data is often represented is therefore used uniformly throughout all tests and exam- as a discrete point sample P ⊂ Rn of some smooth surface ples, except where we explicitly seek a smoothed approxima- M. Rather than convert this data into a polygon mesh, we tion of distance, as in Section 3.3. For highly nonuniform can directly implement the heat method using the point meshes one could set h to the maximum spacing, providing cloud Laplacian of Liu et al.,22 which extends previous a more conservative estimate. Numerical underflow could work by Belkin et al.2 Computation of the gradient and di- theoretically occur for extremely small t, though we do not vergence are described by Crane et al,11 Section 3.2.3. Other encounter this issue in practice. discretizations are certainly possible (see for instance Numerics. As demonstrated in Figures 10, 18, and 19, one the work of Luo et al.23); we picked one that was simple to does not need a particularly nice mesh or point cloud to get a implement in any dimension. It is particularly interesting reasonable distance function. However, as with any numerical

94 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

Figure 10. The heat method can be applied directly to scattered point Figure 11. A source on the front of the Stanford Bunny results in clouds. Left: face scan with holes and noise. Right: kitten surface nonsmooth cusps on the opposite side. By running heat flow for with connectivity removed. Yellow points are close to the source. progressively longer durations t, we obtain smoothed approximations of geodesic distance (right).

Figure 12. Top row: our smoothed approximation of geodesic distance (left) and biharmonic distance (center) both mitigate sharp “cusps” found in the exact distance (right), yet our approximation provides more even spacing of isocontours. Bottom row: biharmonic distance (center) tends to exhibit elliptical level lines near the source, while our smoothed distance (left) maintains isotropic circular profiles as seen in the exact distance (right). method, accuracy and other properties of the solution may be influenced by the quality of the mesh. For instance, in some applications one may wish to avoid “spurious minima,” that is, local maxima or minima that do not appear in the true (smooth) distance function. At present, there is no numerical scheme that guarantees the absence of spurious minima on arbitrary meshes, including exact polyhedral schemes.17 Empirically, however, we observe that the heat method produces fewer spurious minima than either fast marching or the bihar- monic distance (see Figure 20), in part due to regularization from the Hodge step (step III). In cases where one wishes to avoid spurious minima altogether, we advocate the use of Delaunay meshes.

3.3. Smoothed distance Geodesic distance fails to be smooth at points in the cut locus, that is, points at which there is no unique short- est path to the source—these points appear as sharp cusps in the level lines of the distance function. Non- any value of t due to normalization (step II); the solution smoothness can result in numerical difficulty for appli- is isometry invariant since it depends only on intrinsic cations which need to take derivatives of the distance operators. For a time step t = mh2, meaningful values of m function φ (e.g., level set methods), or may simply be are found in the range 1 − 106—past this point the term t∆ undesirable aesthetically. dominates, resulting in little visible change. Several distances have been designed with smoothness in mind, including diffusion distance,10 commute-time dis- 3.4. Boundary conditions tance,14 and biharmonic distance21 (see the last reference To solve the equations in steps I and II, we must define the for a more detailed discussion). These distances satisfy behavior of derivatives near the boundary. Intuitively, the a number of important properties (smoothness, isometry- behavior of our distance approximation should not be sig- invariance, etc.), but are poor approximations of true geo- nificantly influenced by the shape of the boundary (Figure desic distance, as indicated by uneven spacing of isolines 13)—for instance, cutting off a corner of a convex domain (see Figure 12, middle). They can also be expensive to eval- should not affect the distance at uate, requiring one to either solve a linear system for each the points that remain. For exact vertex, or compute a large number of eigenvectors of the distance computation, we

Laplace matrix (∼150 to 200 in practice). can apply standard zero-Neu- BUILD In contrast, one can rapidly construct smoothed approx- mann or zero-Dirichlet bound- imations of geodesic distance by simply applying the heat ary conditions, since this FACTOR method for large values of t (Figure 11). The computational choice does not affect the SOLVE cost remains the same, and isolines are evenly spaced for behavior of the smooth limit POTENTIAL

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 95 research highlights

solution (see Renesse44 Corollary 2 and Norri,30 Theorem 4. EVALUATION 1.1, respectively). Boundary conditions do however 4.1. Performance alter the behavior of our smoothed distance. Although there is A key advantage of the heat method is that the linear systems no well-defined “correct” behavior for this smoothed func- in steps I and III can be prefactored. Our implementation tion, we advocate the use of boundary conditions obtained uses sparse Cholesky factorization,9 which for Poisson-

by taking the mean of the Neumann solution uN and the type problems has guaranteed sub-quadratic complex- 5 Dirichlet solution uD, that is, . The intuition ity but in practice scales much better ; moreover there is behind this behavior again stems from a random walker strong evidence to suggest that sparse systems arising from interpretation: zero Dirichlet conditions absorb heat, elliptic PDEs can be solved in very close to linear time.34, 39 causing walkers to “fall off” the edge of the domain. Independent of these issues, the amortized cost for prob- Neumann conditions prevent heat from flowing out of the lems with a large number of right-hand sides is roughly domain, effectively “reflecting” random walkers. Averaged linear, since back substitution can be applied in essentially conditions mimic the behavior of a domain without linear time. See inset for a breakdown of relative costs in our boundary: the number of walkers leaving equals the num- ber of walkers returning. Figure 14 shows how boundary Figure 14. For path planning, the behavior of geodesics can be conditions affect the behavior of geodesics in a path-plan- controlled via boundary conditions and the time step t. Top-left: ning scenario. Neumann conditions encourage boundary adhesion. Top-right: Dirichlet conditions encourage avoidance. Bottom-left: small values of t yield standard straight-line geodesics. Bottom-right: large Figure 13. Effect of Neumann (top-left), Dirichlet (top-right) and values of t yield more natural trajectories. averaged (bottom-left) boundary conditions on smoothed distance. Averaged boundary conditions mimic the behavior of the same surface without boundary.

Figure 15. Meshes used in Table 1. Left to right: Bunny, Isis, Horse, Bimba, Aphrodite, Lion, Ramses.

96 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

Table 1. Comparison with fast marching and exact polyhedral distance

Heat method Fast marching Precompute Max Mean Max Mean Model Triangles (s) Solve error (%) error (%) Time (s) error (%) error (%) Exact time (s) Bunny 28k 0.21 0.01s (28x) 3.22 1.12 0.28 1.06 1.15 0.95 Isis 93k 0.73 0.05s (21x) 1.19 0.55 1.06 0.60 0.76 5.61 Horse 96k 0.74 0.05s (20x) 1.18 0.42 1.00 0.74 0.66 6.42 Kitten 106k 1.13 0.06s (22x) 0.78 0.43 1.29 0.47 0.55 11.18 Bimba 149k 1.79 0.09s (29x) 1.92 0.73 2.62 0.63 0.69 13.55 Aphrodite 205k 2.66 0.12s (47x) 1.20 0.46 5.58 0.58 0.59 25.74 Lion 353k 5.25 0.24s (24x) 1.92 0.84 10.92 0.68 0.67 22.33 Ramses 1.6M 63.4 1.45s (68x) 0.49 0.24 98.11 0.29 0.35 268.87 Best speed/accuracy in bold; speedup in orange.

implementation; Potential is the time taken to compute the Figure 16. Visual comparison of accuracy. Left: exact polyhedral right hand side in step III. distance. Using default parameters, the heat method (middle) and In practice, a number of factors affect the run time of fast marching (right) both produce results of comparable accuracy, the heat method including the choice of spatial discreti- here within less than 1% of the polyhedral distance—see Table 1 for a zation, discrete Laplacian, and geometric data structures. more detailed comparison. As a typical example, we compared the scheme from Triangle meshes section to the first-order fast march- ing method of Kimmel and Sethian19 and the exact algo- rithm of Mitchell et al.,27 using the state-of-the-art fast marching implementation of Peyré and Cohen31 and the exact implementation of Kirsanov.40 The heat method was implemented in ANSI C in double precision using a vertex-face adjacency list. Single-threaded performance was measured on a 2.4 GHz Intel Core 2 Duo (Table 1). Note that even for a single distance computation the heat method outperforms fast marching; more importantly, updating distance for new subsets γ is consistently an order of magnitude faster (or more) than both fast march- ing and the exact algorithm.

4.2. Accuracy We examined errors in the heat method, fast marching,19 and the polyhedral distance,27 relative to mean edge length h on triangulated surfaces. Both fast marching and the heat method appear to exhibit linear convergence; it is interesting to note that even the exact polyhedral distance Figure 17. Medial axis of the hiragana letter “a” extracted by provides only quadratic convergence. Keeping this fact in thresholding second derivatives of the distance to the boundary. Left: fast marching. Right: heat method. mind, Table 1 uses the polyhedral distance as a baseline for comparison on more complicated geometries—Max is the maximum error as a percentage of mesh diameter and Mean is the mean relative error at each vertex. Note that fast marching tends to achieve a smaller maximum error, whereas the heat method does better on average. Figure 16 gives a visual comparison of accuracy; the only notable dis- crepancy is a slight smoothing at sharp cusps, which may explain the larger maximum error. Figure 17 indicates that smoothing does not interfere with the extraction of the cut locus—here we visualize values of |∆φ| above a fixed threshold. Overall, the heat method exhibits errors of the same order and magnitude as fast marching (at lower com- More recent implementations of the heat method improve putational cost) and is therefore suitable in applications accuracy by using a different spatial discretization,29 or where fast marching is presently used; see Crane et al.11 for by iteratively updating the solution.3 The accuracy of fast more extensive comparisons. marching schemes is determined by the choice of update

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 97 research highlights

Figure 18. Smoothed geodesic distance on an extremely poor Figure 20. In any method based on a finite element approximation, triangulation with significant noise—note that small holes are mesh quality will affect the quality of the solution. However, because essentially ignored. Also note good approximation of distance even the heat method is based on solving low-order elliptic equations along thin slivers in the nose. (rather than high-order or hyperbolic equations), it often produces fewer numerical artifacts. Here, for instance, we highlight spurious extrema in the distance function (i.e., local maxima and minima) produced by the fast marching method (left), biharmonic distance (middle), and the heat method (right) on an acute Delaunay mesh (top) and a badly degenerate mesh (bottom). Inset figures show closeup view of isolines for the bottom figure.

Fast Marching Biharmonic Heat Method

Figure 19. Tests of robustness. Left: our smoothed distance (m = 104) appears similar on meshes of different resolution. Right: even for meshes with severe noise (top) we recover a good approximation of the distance function on the original surface (bottom, visualized on noise-free mesh).

since it tends to produce results more accurate than fast marching at a similar computational cost. However, accu- racy is measured relative to the polyhedral distance rather than the smooth geodesic distance of the approximated sur- face. Like fast marching, Surazhsky’s method does not take advantage of precomputation and therefore exhibits a sig- nificantly higher amortized cost than the heat method; it is also limited to triangle meshes.

4.3. Robustness Two factors contribute to the robustness of the heat method, namely (1) the use of an unconditionally stable time discret- ization and (2) an elliptic rather than hyperbolic formula- tion (i.e., relatively stable local averaging vs. more sensitive global wavefront propagation). Figure 19 verifies that the heat method continues to work well even on meshes that are poorly discretized or corrupted by a large amount of noise (here modeled as uniform Gaussian noise applied to the vertex coordinates). In this case we use a moderately large value of t to investigate the behavior of our smoothed distance; similar behavior is observed for small t values. rule—a number of highly accurate rules have been devel- Figure 18 illustrates the robustness of the method on a sur- oped for regular grids (e.g., HJ WENO18), but fewer options face with many small holes as well as long sliver triangles. are available on irregular domains such as triangle meshes, the predominant choice being the first-order update of 5. CONCLUSION Kimmel and Sethian.19 Finally, the approximate algorithm The heat method is a simple, general method that can be eas- of Surazhsky et al.40 provides an interesting comparison ily incorporated into a broad class of algorithms. However, a

98 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11

great deal remains to be explored, including further inves- Point-based manifold harmonics. 37. Sethian, J.A. Level Set Methods and IEEE Trans. Vis. Comput. Graph. 18, Fast Marching Methods: Evolving tigation of alternative spatial discretizations, and formal 10 (2012), 1693–1703. Interfaces in Computational analysis of convergence under refinement. Further explo- 23. Luo, C., Safa, I., Yusu, W. Geometry, Fluid Mechanics, Approximating gradients for meshes Computer Vision and Materials ration of the parameter t also provides an avenue for future and point clouds via diffusion metric. Science. Cambridge University Press, work (especially in the case of variable spacing), though one Comput. Graph. Forum 28, 5 (2009), Cambridge 1996. 1497–1508. 38. Solomon, J., de Goes, F., Peyré, G., should note that the existing estimate already outperforms 24. MacNeal, R. The solution of partial Cuturi, M., Butscher, A., Nguyen, A., fast marching in terms of mean error (Table 1). Another differential equations by means Du, T., Guibas, L. Convolutional of electrical networks. Ph.D. Wasserstein distances: Efficient natural question is whether a similar transformation can dissertation, Caltech (1949). optimal transportation on geometric be applied to a larger class of Hamilton-Jacobi equations; it 25. Memoli, F., Sapiro, G. Fast domains. ACM Trans. Graph. 34, 4 computation of weighted distance (2015), 66:1–66:11. DOI:http://dx.doi. is likewise enticing to apply a similar principle to distance functions and geodesics on implicit org/10.1145/2766963. computation on domains that do not immediately resemble hyper-surfaces. J. Comput. Phys. 173 39. Spielman,D.A., Teng, S.-H.. Nearly- (2001), 730–764. linear time algorithms for graph a continuous domain (such as a weighted graph). 26. Memoli, F., Sapiro, G. Distance functions partitioning, graph sparsification, and and geodesics on submanifolds of Rd solving linear systems. In Proc. ACM and point clouds. SIAM J. Appl. Math. Symp. Theory Comput. (STOC ‘04) Acknowledgments 65, 4 (2005), 1227–1260. (2004). ACM, 81–90. 27. Mitchell, J., Mount, D., Papadimitriou, C. 40. Surazhsky, V., Surazhsky, T., This work was funded by a Google PhD Fellowship and a The discrete geodesic problem. SIAM J. Kirsanov, D., Gortler, S.J., Hoppe, H. grant from the Fraunhofer Gesellschaft. Thanks to Michael Comput. 16, 4 (1987), 647–668. Fast exact and approximate geodesics 28. Neel, R., Stroock, D. Analysis of the on meshes. ACM Trans. Graph. 24, 3 Herrmann for inspiring discussions. Meshes are provided cut locus via the heat kernel. Surv. (2005), 553–560. courtesy of the Stanford Computer Graphics Laboratory, Diff. Geom. 9 (2004), 337–349. 41. van Pelt, R., Gasteiger, R., Lawonn, K., 29. Nguyen, T., Karciauskas, K., Peters, Meuschke, M., Preim, B. Comparative the AIM@Shape Repository, Luxology LLC, and Jotero GbR J. C1 finite elements on non-tensor- blood flow visualization for cerebral (http://www.evolution-of-genius.de/). product 2d and 3d manifolds. Appl. aneurysm treatment assessment. Math. Comput. (2016). Comput. Graph. Forum 33, 3 30. Norris, J. Heat kernel asymptotics (2014), 133–140. DOI:http://dx.doi. and the distance function in Lipschitz org/10.1111/cgf.12369 Riemannian manifolds. Acta Math. 42. Varadhan, S.R.S. On the behavior of References 10. Coifman, R.R., Lafon, S. Diffusion 179, 1 (1997), 79–103. the fundamental solution of the heat 1. Alexa, M., Wardetzky, M. Discrete maps. Appl. Comput. Harmon. Anal. 31. Peyré, G., Cohen, L.D. Chapter equation with variable coefficients. Laplacians on general polygonal 21 (2006), 5–30. geodesic computations for fast and Comm. Pure Appl. Math. 20, 2 (1967), meshes. ACM Trans. Graph 30, 4 11. Crane, K., Weischedel, C., Wardetzky, M. accurate surface remeshing and 431–455. (2011), 102:1–102:10. Geodesics in heat: A new approach parameterization. In Progress in 43. Villa, E. Methods of geometric 2. Belkin, M., Sun, J., Wang, Y. to computing distance based on Nonlinear Differential Equations and measure theory in stochastic Constructing Laplace operator from heat flow. ACM Trans. Graph. 32, 5 Their Applications. Volume 63 (2005). geometry. Ph.D. dissertation. point clouds in Rd. In Proceedings of (2013), 152:1–152:11. Springer, 157–171. Università degli Studi di Milano the Twentieth Annual ACM-SIAM 12. de Goes, F., Desbrun, M., Meyer, 32. Rangarajan, A., Gurumoorthy, K. A (2006). Symposium on Discrete Algorithms, M., DeRose, T. Subdivision exterior Fast Eikonal Equation Solver Using the 44. Von Renesse, M.-K. Heat kernel SODA‘09 (New York, 2009). calculus for geometry processing. Schrödinger Wave Equation. Technical comparison on Alexandrov spaces Society for Industrial and Applied ACM Trans. Graph. 35, 4 (2016). Report REP-2011–512. CISE, with curvature bounded below. Poten. Mathematics, Philadelphia, PA, USA, 13. de Goes, F., Liu, B., Budninskiy, M., University of Florida, 2011. Anal. 21, 2 (2004), 151–176. 1031–1040. Tong, Y., Debrun, M. Discrete 2-tensor 33. Rustamov, R., Lipman, Y., Funkhouser, T. 45. Yang, F., Cohen, L. Geodesic distance 3. Belyaev, A., Fayolle, P.-A. On fields on triangulations.Comput. Interior distance using Barycentric and curves through isotropic and variational and PDE-based distance Graph. Forum 33, 5 (2014), 13–24. coordinates. Comput. Graph. Forum anisotropic heat equations on images function approximations. Comput. 14. Fouss, F., Pirotte, A., Renders, J.-M., (Symposium on Geometry Processing) and surfaces. J. Math. Imaging Vis. Graph. Forum 34, 8 (2015), 104–118. Marco, S. Random-walk computation 28, 5 (July 2009), 1279–1288. 15, 2 (2015), 210–228. 4. Bommes, D., Kobbelt, L. Accurate of similarities between nodes of a 34. Schmitz,P.G., Ying, L. A fast direct 46. Ying, X., Xin, S.-Q., He, Y. Parallel computation of geodesic distance graph with application to collaborative solver for elliptic problems on general Chen-Han (PCH) algorithm for fields for polygonal curves on triangle recommendation. IEEE Trans. Knowl. meshes in 2D. J. Comput. Phys. 231, 4 discrete geodesics. ACM Trans. meshes. In Proceedings of Workshop Data Eng. 19, 3 (2007), 355–369. (2012), 1314–1338. Graph. 33, 1 (2014), 9:1–9:11. on Vision, Modeling, and Visualization 15. Huth, A., Griffiths, T., Theunissen, F., 35. Schrijver, A. On the history of the DOI: http://dx.doi. (VMV) (Saarbrücken, Germany, Gallant, J. PrAGMATiC: A probabilistic shortest path problem. Docum. Math. org/10.1145/2534161. November 7–9, 2007), 151–160. and generative model of areas tiling 1 (2012), 155–167. 47. Zou, Q., Zhang, J., Deng, B., Zhao, 5. Botsch, M., Bommes, D., Kobbelt, L. the cortex. arXiv:1504.03622 (2015). 36. Schwarz, G. Hodge Decomposition: A J. Iso-level tool path planning for Efficient linear system solvers for 16. Hysing, S., Turek, S. The eikonal Method for Solving Boundary Value freeform surfaces. Comput. Aid. Des. mesh processing. In IMA Conference equation: Numerical efficiency Problems. Springer, Berlin 1995. 55 (2014), 117–125. on the Mathematics of Surfaces, R.R. vs. algorithmic complexity. In Martin, H.E. Bez, and M.A. Sabin, eds. Proceedings of Algorithmy (2005). Volume 3604 of Lecture Notes in Solvak University of Bratslavia, Computer Science (2005). Springer, Bratslavia, 22–31. Keenan Crane ([email protected]), Clarisse Weischedel and Max 62–83. 17. Itoh, J-I, Sinclair, R. Thaw: A tool Carnegie Mellon University. Wardetzky (Clarisse.Weischedel@ 6. Caissard, T., Coeurjolly, D., Gueth, P. for approximating cut loci on a darmstadt.ihk.de, [email protected] DGtal DEC: Discrete Exterior Calculus triangulation of a surface. Exp. Math. goettingen.de), University of Göttingen. Package for the Digital Geometry 13, 3 (2004), 309–325. Tools and Algorithms Library (2015) 18. Jiang, G., Peng, D. Weighted ENO 7. Campen, M., Leif K. Walking on broken schemes for Hamilton–Jacobi mesh: Defect-tolerant geodesic equations. SIAM J. Sci. Comput 21 distances and parameterizations. (1997), 2126–2143. Comput. Graph. Forum 30, 2 (2011), 19. Kimmel, R., Sethian, J.A. Fast 623–632. marching methods on triangulated 8. Chen, J., Han, Y. Shortest paths on domains. Proc. Nat. Acad. Sci. 95 a polyhedron. Sympos. Comput. (1998), 8341–8435. Geom. In Proceedings of the Sixth 20. Lin, B., Ji, Y., He, X., Ye, J. Geodesic Annual Symposium on Computational distance function learning via heat Geometry SCG'90 (Berkley, California, flow on vector fields. InProceedings of USA, 1990). ACM, New York, NY, USA the 31st International Conference on 360–369. Machine Learning, ICML 2014 (Beijing, 9. Chen, Y., Davis, T.A., Hager, W.W., China, 21–26 June 2014) Inter. Conf. Rajamanickam, S. Algorithm 887: Mach. Learn. (2014), 145–153. CHOLMOD, supernodal sparse 21. Lipman, Y., Rustamov, R.M., cholesky factorization and update/ Funkhouser, T.A. Biharmonic distance. downdate. ACM Trans. Math. ACM Trans. Graph. 29, 3 (July 2010), Softw. 35, 3 (October 2008), 14 pp., 11 pp., Article 27. Article 22. 22. Liu, Y., Prabhakaran, B., Guo, X. Copyright held by authors. Publication rights licensed to ACM. $15.00.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 99 CAREERS

Auburn University Baylor University several years, and will seek to fill faculty positions Department of Computer Science and Software Chairperson, Department of Computer Science at all levels. We invite applications for a full-time, Engineering (CSSE) non-tenure-track faculty position, beginning in Multiple Faculty Positions in Cybersecurity The School of Engineering and Computer Sci- the 2018-2019 academic year. Candidates should ence invites nominations and applications for be committed to excellence in undergraduate CSSE invites applications for multiple tenure-track the position of Chair of the Department of Com- education, and should be able to both teach a faculty positions to begin in Fall 2018 or later. puter Science. The successful candidate must broad variety of undergraduate computer science Candidates at the Assistant Professor level will be have an earned Ph.D. in Computer Science or a courses, and to participate in the development of given preference, however outstanding candidates related field, leadership experience, a commit- new courses that reflect the evolving landscape of at senior levels will also be considered. A Ph.D. de- ment to undergraduate and graduate education, the discipline. gree in computer science, software engineering or a strong research record that includes significant Minimum requirements for the title of As- a closely related field must be completed by the external funding, and effective communication sistant Professor of the Practice include a Ph.D. start of appointment. Excellent communication and organization skills. in Computer Science or closely related disci- skills are required. We are interested in candidates The Department: Computer Science is one of pline. Candidates who have only attained a specializing in all areas related to security, such as three departments in the School of Engineering Master’s degree would be eligible for the title AI/machine learning applications to security, block- and Computer Science. It offers an ABET- of Lecturer. chain, cryptocurrency, cyberidentity, cyberinfrastruc- accredited B.S. in Computer Science degree, a B.A. Application review beings October 1, 2017. ture and critical infrastructure protection, digital degree with a major in Computer Science, a B.S. See www.cs.bc.edu for more information. forensics, reverse engineering, secure cloud, mobile in Informatics with a major in Bioinformatics, To apply go to http://apply.interfolio.com/44984. systems, networks and operating systems, secure soft- B.S. in Computing with a major in Computer ware engineering, and securing the Internet of Things. Science , and M.S. and Ph.D. degrees in CSSE is home to the Auburn Cyber Research Computer Science. The Department has 15 full- Boston College Center (http://cyber.auburn.edu), and is affiliated time faculty, over 370 undergraduate majors and Associate or Full Professor of Computer with the McCrary Institute for Critical Infrastructure 20 graduate students. The Department’s greatest Science Protection and Cyber Systems (http://mccrary.au- strength is its dedication to the success of the burn.edu). The department has 21 full-time tenure- students and each other. Interested candidates The Computer Science Department of Boston track faculty members and supports strong M.S. and may contact any faculty member to ask questions College aims to grow substantially over the next Ph.D. programs (with a new M.S. in Cybersecurity and/or visit the departmental web site at http:// several years, and will seek to fill faculty posi- Engineering projected to start in Fall 2018). Faculty www.ecs.baylor.edu/computerscience. tions at all levels. We invite applications for a research areas include artificial intelligence, archi- The University: Baylor University is a private senior-level (Associate or Full Professor) position, tecture, computational biology, computer science Christian university and a nationally-ranked re- starting in the 2018-2019 academic year. The suc- education, data science, energy-efficient systems, search institution, consistently listed with highest cessful candidate is expected to play a leadership human-computer interaction, Internet of Things, honors among The Chronicle of Higher Educa- role in the creation of a Data Science program in learning science, machine learning, modeling and tion’s “Great Colleges to Work For.” The university conjunction with the new interdisciplinary Insti- simulation, multi-agent systems, networks, security, is recruiting new faculty with a deep commitment tute for Integrated Science and Society, recently software engineering and wireless engineering. to excellence in teaching, research and scholar- announced at Boston College, and will also par- Auburn University is one of the nation’s pre- ship. Baylor seeks faculty who share in our aspira- ticipate in shaping the future of our growing de- mier public land-grant institutions. It is ranked tion to become a tier-one research institution while partment. 46th among public universities in the U.S. News and strengthening our distinctive Christian mission Applicants must have a Ph.D. in Computer World Report 2018 Rankings. It is nationally recog- as described in our strategic vision, Pro Futuris Science or closely related discipline, a record of nized for its commitment to academic excellence, (https://www.baylor.edu/profuturis). As the world’s strong research accomplishment and external a positive work environment, student engagement, largest Baptist University, Baylor offers over 40 doc- funding, and a commitment to quality in under- and its beautiful campus. Auburn residents enjoy a toral programs and has over 17,000 students from graduate and graduate education. Preference will thriving community, recognized as one of the “best all 50 states and more than 80 countries. be given to candidates whose research is in the ar- small towns in America.” The city is located on the Appointment Date: Fall 2018. eas of high-performance data mining / machine rapidly developing I-85 corridor between Atlanta, Application Procedure: Applications, includ- learning or data visualization, particularly those GA, and Montgomery, AL. The Auburn City school ing detailed curriculum vitae, a statement dem- with a history of interdisciplinary collaboration, system is ranked as one of the top school systems onstrating an active Christian faith, and contact but outstanding candidates in all research areas in the nation and the state. A nationally recognized information for three references should be either will be considered. hospital, East Alabama Medical Center, is located emailed to [email protected] or mailed to Application review begins October 1, 2017. close by in Opelika. The Auburn-Opelika metropol- Chair Search Committee, Department of Com- See www.cs.bc.edu for more information. itan area has a population of over 150,000. puter Science, Baylor University, One Bear Place To apply go to http://apply.interfolio.com/44982. Applicants should submit a cover letter, cur- #97141, Waco, TX 76798-7141. riculum vita, research vision, teaching philoso- For full consideration, applications should phy, and the names of references through the be received by January 1, 2018. However, applica- Boston College faculty hiring link on the department home page tions will be accepted until the position is filled. Tenure Track Assistant Professor in Computer http://www.eng.auburn.edu/csse. There is no ap- Science plication deadline. Application review will begin in November. Selected candidates must be able Boston College The Computer Science Department of Boston to meet eligibility requirements to work legally in Assistant Professor of the Practice or Lecturer College aims to grow substantially over the next the United States at the time of appointment for several years and will seek to fill faculty positions the proposed term of employment. Auburn Uni- The Computer Science Department of Boston at all levels. We invite applications for one or versity is an EEO/Vet/Disability Employer. College aims to grow substantially over the next more tenure-track faculty positions at the rank of

100 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 Assistant Professor, beginning in the 2018-2019 plied and computational mathematics, computer California State University – academic year. Successful candidates will be ex- science and statistics are invited to apply. Areas of Sacramento pected to develop strong research programs that interest include (but are not limited to) scientific Tenure-Track Assistant Professor can attract external research funding. The search computing, optimization, statistics, probability, will focus on candidates who can participate in networked systems, control and dynamical sys- California State University – Sacramento, Depart- cross-disciplinary research in conjunction with tems, robotics, theory of computation, security, ment of Computer Science. Three tenure-track as- the new Institute for Integrated Science and So- privacy, machine learning, and algorithmic eco- sistant professor positions to begin with the Fall ciety recently announced at Boston College, in nomics. In addition, we welcome applications 2018 semester. Applicants specializing in any area the areas of high-performance data mining / ma- from candidates who have demonstrated strong of computer science will be considered. Those chine learning, systems / networks, data visual- connections between computer science, engi- with expertise in areas related to software engi- ization, and human-computer interaction. How- neering and applied mathematics, and to other neering, embedded systems, or artificial intel- ever, outstanding candidates in all research areas fields such as the physical, biological, and social ligence are especially encouraged to apply. Ph.D. will be considered. sciences. in Computer Science, Computer Engineering, or Minimum requirements for all positions in- A commitment to world-class research as well closely related field required by the time of the ap- clude a Ph.D. in Computer Science or closely re- as high-quality teaching and mentoring is ex- pointment. For detailed position information, in- lated discipline, an energetic research program pected. The initial appointment at the Assistant- cluding application procedure, please see http:// that promises to attract external funding, and Professor level is for four years and is contingent www.csus.edu/about/employment/. Screening a commitment to quality in undergraduate and upon the completion of a Ph.D. degree in Applied will begin November 19, 2017, and remain open graduate education. Mathematics, Computer Science, Engineering, or until filled. AA/EEO employer. Clery Act statis- Application review begins October 1, 2017. a related field. tics available. Mandated reporter requirements. See www.cs.bc.edu for more information. Applications will be reviewed beginning Criminal background check will be required. To apply go to http://apply.interfolio.com/44980. November 15, 2017, and applicants are encouraged to have all their application materials on file by this date. For a list of documents Carnegie Mellon University California Institute of Technology required and full instructions on how to apply Faculty Hiring (Caltech) on-line, please visit http://www.cms.caltech.edu/ Tenure-Track Faculty Position search. Questions about the application process The School of Computer Science consists of may be directed to: [email protected]. seven departments, spanning a wide range of The Computing and Mathematical Sciences We are an equal opportunity employer and topics in computer science and the application (CMS) department at the California Institute of all qualified applicants will receive consideration of computers to real-world systems. Faculty Technology (Caltech) invites applications for a for employment without regard to race, color, positions are specific to each department, though tenure-track faculty position. CMS is a unique religion, sex, sexual orientation, gender identity, in certain cases, joint positions are also possible. environment where innovative, interdisciplinary, or national origin, disability status, protected vet- We are seeking tenure, research, and systems and foundational research is conducted in a col- eran status, or any other characteristic protected track faculty candidates with a strong interest in legial atmosphere. Candidates in all areas of ap- by law. research, an earned Ph.D., and outstanding aca-

Professor or Assistant Professor (Tenure Track) of Analog and Mixed Signal Interfaces

→ The Department of Information Technology and Electrical Engineering (www.ee.ethz.ch) at ETH Zurich invites applications for the above-mentioned position. → The successful candidate is expected to develop a strong and visible research program in the area of analog and mixed signal interfaces circuits and systems. He or she has a strong background in one or more of the following areas: (i) analog circuits and techniques for ultra-low power, ranging from basic building blocks (e.g. amplifiers, filters) to silicon sensors, interfaces, and novel clock generation architectures; (ii) data converters enabled by new techniques, architectures, or circuit topologies; (iii) wireless transceiver and RF circuits for low-power and energy-efficient links, cellular connectivity including machine-to-machine applications, emerging broadband networks (5G), millimeter-wave and THz systems (radar, sensing and imaging); (iv) wireline communications circuits and systems for chip-to-chip communication, including serial links, high-speed memory, accelerator, peripheral interfaces, backplanes, long-haul, and powerline communications. Candidates should hold a PhD. A track record of successfully manufactured chips and systems is highly desirable. In addition, commitment to teaching and the ability to lead a research group are expected. Ge- nerally, at ETH Zurich undergraduate level courses are taught in German or English and graduate level courses are taught in English. → Assistant professorships have been established to promote the careers of younger scientists. ETH Zurich implements a tenure track system equivalent to other top international universities. The level of the appointment will depend on the successful candidate’s qualifications. → Please apply online: www.facultyaffairs.ethz.ch → Applications should include a curriculum vitae, a list of publications, a statement of future research and teaching interests, and a description of the three most important achievements. The letter of application should be addressed to the President of ETH Zurich, Prof. Dr. Lino Guzzella. The closing date for applications is 15 December 2017. ETH Zurich is an equal opportunity and family friendly employer and is responsive to the needs of dual career couples. We specifically encourage women to apply.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 101 CAREERS

demic credentials. Candidates for tenure track Human-Computer Interaction Institute: electrical and computer engineering, as well appointments should also have a strong interest https://hcii.cmu.edu/careers/list as interdisciplinary areas such as robotics and in graduate and undergraduate education. Institute for Software Research: http://www. cyber-physical systems, bio-ECE, microsystems, We are also seeking teaching track faculty isri.cmu.edu/jobs/index.html machine learning, applied mathematics, and candidates. You should have a Ph.D. in Computer Language Technologies Institute: http://lti. energy. See our website, https://www.ece.cornell. Science or a related computing discipline, a back- cs.cmu.edu/news/lti-hiring edu/academics, for additional information on ground of demonstrated excellence and dedica- Machine Learning Department: http://www. our programs. tion to teaching, the ability to collaborate with oth- ml.cmu.edu/Faculty_Hiring.html Applicants must hold a doctorate in an er faculty in a fast-paced environment, and must Robotics Institute: http://ri.cmu.edu/about/ appropriate field by the time their Faculty be prepared to teach in a wide variety of settings, hiring-faculty-positions/ appointment would begin, must have including large undergraduate lecture courses and demonstrated an ability to conduct outstanding classes delivered in non-traditional formats. Please send email to [email protected]. research, and must show promise for excellence in Candidates with a commitment toward build- edu with any questions. teaching. Although we anticipate filling positions ing an equitable and diverse scholarly commu- Carnegie Mellon University shall abide by the at the assistant professor level, applications nity are particularly encouraged to apply. We are requirements of 41 CFR §§ 60-1.4(a), 60-300.5(a) at all levels will be considered; salary and rank very interested in applications from candidates and 60-741.5(a). These regulations prohibit dis- will be commensurate with qualifications and who have a demonstrated track record in men- crimination against qualified individuals based experience. toring and nurturing women and students from on their status as protected veterans or individu- Applicants should submit a curriculum vita, groups traditionally underrepresented in com- als with disabilities, and prohibit discrimination a research statement, a teaching statement, puter science. against all individuals based on their race, color, three recent publications, and complete con- To ensure full consideration of your applica- religion, sex, or national origin. Moreover, these tact information for at least three references. tion, please submit all materials no later than regulations require that covered prime contrac- Applications must be submitted on-line at https:// December 15, 2017. In your cover letter, please in- tors and subcontractors take affirmative action to academicjobsonline.org/ajo/jobs/9871. Review of dicate clearly the department(s) you are applying employ and advance in employment individuals applications will begin immediately. Applica- to. You can learn more about our hiring plans and without regard to race, color, religion, sex, nation- tions received by December 4, 2017, will receive application instructions by visiting http://www. al origin, protected veteran status or disability. full consideration. cs.cmu.edu/employment-scs. The School of Electrical and Computer For more information about the hiring priori- Engineering is especially interested in qualified ties in a particular department, please visit a de- Cornell University candidates who can contribute, through their partment site below: Multiple Tenure-Track Faculty Positions research, teaching, and/or service, to the diversity Computational Biology Department: http:// of the academic community and to creating www.cbd.cmu.edu/tenure-track-faculty- The SCHOOL OF ELECTRICAL AND COMPUTER a climate that attracts students of all races, positions/ ENGINEERING at CORNELL UNIVERSITY, genders and nationalities. We strongly encourage Computer Science Department: https://www. Ithaca, New York, invites applications for multiple underrepresented minority and women csd.cs.cmu.edu/careers/faculty-hiring tenure-track Faculty positions in all areas of candidates to apply. Cornell University actively

Professor or Assistant Professor (Tenure Track) of Embedded Information Systems

→ The Department of Information Technology and Electrical Engineering (www.ee.ethz.ch) at ETH Zurich invites applications for the above-mentioned position.

→ The successful candidate is expected to develop a strong and visible research program in the area of embedded information systems. He or she has a strong background in areas such as embedded software, edge computing, embedded operating systems, real-time systems, biomedical embedded systems, security of embedded systems, as well as corresponding theoretical concepts. Candidates should hold a PhD and have an excellent record of accomplishments. In addition, commitment to teaching and the ability to lead a research group are expected. Generally, at ETH Zurich undergraduate level courses are taught in German or English and graduate level courses are taught in English.

→ Assistant professorships have been established to promote the careers of younger scientists. ETH Zurich implements a tenure track system equivalent to other top international universities. The level of the appointment will depend on the successful candidate’s qualifications.

→ Please apply online: www.facultyaffairs.ethz.ch

→ Applications should include a curriculum vitae, a list of publications, a statement of future research and teaching interests, and a description of the three most important achievements. The letter of application should be addressed to the President of ETH Zurich, Prof. Dr. Lino Guzzella. The closing date for applications is 15 December 2017. ETH Zurich is an equal opportunity and family friendly employer and is responsive to the needs of dual career couples. We specifically encourage women to apply.

102 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 seeks to meet the needs of dual career couples, Durham, Chapel Hill, and the Research Tri- consideration, applications should be submitted has a Dual Career program, and is a member of the angle of North Carolina are frequently ranked by December 15, 2017. Applicants are encouraged Upstate New York Higher Education Recruitment among the best places in the country to live and to clearly identify in their cover letter the area(s) Consortium to assist with dual career searches. work. Duke and the many other universities in the that best describe their research interests. All Diversity and Inclusion are a part of Cornell area offer a wealth of education and employment applications must be submitted online at https:// University’s heritage. We are a recognized opportunities for spouses and families. www.scs.gatech.edu/content/cs-faculty-hiring. employer and educator valuing AA/EEO, Protected More information about the School of Computer Veterans and Individuals with Disabilities. Science is available at http://scs.gatech.edu. In Georgia Institute of Technology addition to collaborations within our school and Tenured-Tenure Track Faculty Positions the two other schools in the college, our faculty Duke University works closely with other world-class faculty in Tenure-Track Faculty Positions in Computing The School of Computer Science in the College of the university, including those in the Colleges of Computing at the Georgia Institute of Technol- Engineering, and Sciences. Duke University invites applications and nomi- ogy invites applications for tenure-track faculty is located in the heart of the nations for five tenure-track or tenured faculty positions. We are primarily seeking junior-level city of Atlanta, the cultural and economic cen- positions in computing, at all ranks, to begin July candidates at the Assistant Professor level, but ter of the Southeastern U.S. The greater Atlanta 2018. This search is a joint initiative between the truly exceptional candidates at all levels may area is very cosmopolitan, home to many large Department of Computer Science and the Depart- also be considered. Applicants should have corporations including Coca Cola, Delta, Home ment of Electrical and Computer Engineering to completed or be near completion of a Ph.D. in Depot, and UPS, and is listed among the top 10 rapidly expand the university’s existing strengths computer science or a related field, and should metropolitan areas in the United States. Georgia and to support exciting new initiatives in com- demonstrate potential for excellence in research Tech is an Affirmative Action/Equal Opportunity puting. We encourage applications in all areas of and teaching. The School of Computer Science, Employer. Applications from women and under- computer science and engineering, with special one of three schools in the College of Computing, represented minorities are strongly encouraged. interest in the following themes: focuses on research that makes computing and ˲˲Two positions in all disciplinary areas of com- communication smart, fast, reliable, and secure, puter science, including but not limited to secu- with research groups in computer architecture, Georgia Institute of Technology, rity and privacy, distributed systems and network- databases, machine learning, networking, pro- School of Computational Science and ing, mobile and embedded systems, machine gramming languages, security, software engi- Engineering learning, algorithms, as well as interdisciplinary neering, systems, and theory. We seek candidates Tenure-Track Faculty work that relates to social sciences or biological that can add to or enhance the current research sciences. areas of the school, and we put more emphasis Computational Science and Engineering solves ˲˲Two joint positions between computer science on excellence than on candidates’ specific area of real-world problems in science, engineering, and electrical and computer engineering in the expertise. health, and social domains, by using high-per- area of machine learning (with an applied and Applications will be considered until open formance computing, modeling and simulation, methodological focus). positions are filled. However, to receive full large-scale “big data” analytics, and machine ˲˲One position in all disciplinary areas of learning. The School of Computational Science computer engineering, including but not limited to security and privacy, distributed systems and networking, as well as mobile and embedded systems.

Candidates are expected to have a doctoral de- gree in computer science, computer engineering, TENURE-TRACK AND TENURED POSITIONS or a related discipline. A successful candidate must have a solid disciplinary foundation and ShanghaiTech University invites highly qualified candidates to fill multiple tenure-track/tenured demonstrate promise of outstanding scholarship faculty positions as its core founding team in the School of Information Science and in every regard, including research and teaching. Technology (SIST). We seek candidates with exceptional academic records or demonstrated The Duke faculty and students communi- strong potentials in all cutting-edge research areas of information science and technology. They must be fluent in English. English-based overseas academic training or background ties are currently very diverse and are strongly is highly desired. committed to further enhancing this diversity. ShanghaiTech is founded as a world-class research university for training future generations We seek faculty members who are committed of scientists, entrepreneurs, and technical leaders. Boasting a new modern campus in to building a diverse and inclusive community, Zhangjiang Hightech Park of cosmopolitan Shanghai, ShanghaiTech shall trail-blaze a new education system in China. Besides establishing and maintaining a world-class research which fosters excellence in research and teach- profile, faculty candidates are also expected to contribute substantially to both graduate ing. We strongly encourage applications from and undergraduate educations. women and underrepresented minorities in com- Academic Disciplines: Candidates in all areas of information science and technology shall puting. Please see www.cs.duke.edu, www.ece. be considered. Our recruitment focus includes, but is not limited to: computer architecture, software engineering, database, computer security, VLSI, solid state and nano electronics, RF duke.edu, and www.provost.duke.edu/faculty/ for electronics, information and signal processing, networking, security, computational foundations, information about the departments and advan- big data analytics, data mining, visualization, computer vision, bio-inspired computing systems, tages that Duke offers to faculty. power electronics, power systems, machine and motor drive, power management IC as well as inter-disciplinary areas involving information science and technology. Applicants should submit their materials (cover letter, research statement, teaching state- Compensation and Benefits: Salary and startup funds are highly competitive, commensurate with experience and academic accomplishment. We also offer a ment, contacts for at least three references) elec- comprehensive benefit package to employees and eligible dependents, including on- tronically through AcademicJobsOnline https:// campus housing. All regular ShanghaiTech faculty members will join its new tenure-track academicjobsonline.org/ajo/jobs/9908 system in accordance with international practice for progress evaluation and promotion. For full consideration, applications and refer- Qualifications: • Strong research productivity and demonstrated potentials; ences should be received by December 15, 2017. • Ph.D. (Electrical Engineering, Computer Engineering, Computer Science, Statistics, Duke University is an Affirmative Action/ Applied Math, or related field); Equal Opportunity Employer committed to pro- • A minimum relevant (including PhD) research experience of 4 years. viding employment opportunity without regard Applications: Submit (in English, PDF version) a cover letter, a 2-page research plan, a CV plus copies of 3 most significant publications, and names to an individual’s age, color, disability, genetic of three referees to: [email protected]. For more information, visit information, gender, gender identity, national http://sist.shanghaitech.edu.cn/NewsDetail.asp?id=373 origin, race, religion, sexual orientation or vet- Deadline: The positions will be open until they are filled by appropriate candidates. eran status.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 103 CAREERS

and Engineering of the College of Computing at en and under-represented minorities are strongly cessful candidate will be to develop and teach the Georgia Institute of Technology seeks tenure- encouraged. courses that support the Information & System track faculty at all levels. Our school seeks candi- For more information about Georgia Tech’s Security concentration, as well as collaborating dates who may specialize in a broad range of ap- School of Computational Science and Engineer- with faculty in computer science, political sci- plication areas including biomedical and health; ing please visit: http://www.cse.gatech.edu/. ence, criminology, and sociology on introductory urban systems and smart cities; social good and courses that explain the connections between sustainable development; materials and manu- these disciplines. facturing; and national security. Applicants must Le Moyne College Le Moyne College strives for academic have an outstanding record of research, a sincere Assistant Professor / Professor of Practice – excellence through its comprehensive programs commitment to teaching, and interest in engag- Cybersecurity rooted in the liberal arts and sciences. The ing in substantive interdisciplinary research with College supports a quality undergraduate learning collaborators in other disciplines. The College of Arts and Sciences at Le Moyne Col- experience with small class sizes, and faculty that Georgia Tech is located in the heart of metro lege invites applications for a full-time faculty are approachable and attentive to student advising. Atlanta, a home to more than 5.5 million people position in its new interdisciplinary cybersecurity For more information and how to apply visit and nearly 150,000 businesses, a world-class program to begin in August, 2018. The College is our website at www.lemoyne.edu/employment. airport, lush parks and green spaces, competi- seeking a candidate with a Ph.D. in cybersecurity Preference given to applications received by tive schools and numerous amenities for enter- or cognate field, to fill a tenure-track Assistant December 8, 2017. Review will continue until tainment, sports and restaurants that all offer a Professor position (other ranks will be consid- position is filled. Le Moyne College is an equal top-tier quality of life. From its diverse economy, ered). Candidates with significant industry ex- opportunity employer and encourages women, global access, abundant talent and low costs of perience in cybersecurity are also encouraged to persons of color, and Jesuits to apply for business and lifestyle, metro Atlanta is a great apply as a Professor of Practice. employment. place to call “home.” Residents have easy access The College’s new cybersecurity program to arts, culture, sports and nightlife, and can ex- presents a holistic approach to thinking about cy- perience all four seasons, with mild winters that bersecurity issues. It is designed to give students Marist College rarely require a snow shovel. foundational knowledge regarding the varied Assistant/Associate Professor, Computing Applications should be submitted online cybersecurity challenges that individuals and or- Technology through https://academicjobsonline.org/ajo/ ganizations face on a daily basis. It will prepare jobs/9687. For best consideration, applications students for graduate programs and careers that The Marist College Department of Computing are due by December 15, 2017. The application include international relations, legal studies, Technology currently seeks applications for ten- material should include a full academic CV, a per- government (federal, state and local), criminol- ure-track positions in Computer Science and In- sonal narrative on teaching and research, a list of ogy, military, security compliance, and cybersecu- formation Technology & Systems. at least three references and up to three sample rity technology specialist. This new cybersecurity We welcome candidates who have the follow- publications. program includes three concentrations: Crime, ing teaching and research interests: software de- Georgia Tech is an Affirmative Action/Equal Society & Culture; Information & System Security; velopment and systems analysis/design, data and Opportunity Employer. Applications from wom- and, Policy & Law. The responsibilities of the suc- information security, cloud computing and net- works, or data science and analytics. Applicants with the ability to teach in multiple areas across several disciplines will receive preference. Appli- cants must be willing to teach undergraduate and graduate courses in both traditional on-ground and on-line environments. Required duties out- side the classroom include scholarly activities that result in peer-reviewed publications as well as engagement in college and professional servic- es such as advising/mentoring students, serving on department, school, or college committees. Candidates must have a doctoral degree in ADVERTISING IN CAREER OPPORTUNITIES Computer Science, Information Systems, or a closely related field. We will consider ABDs in an How to Submit a Classified Line Ad: Send an e-mail to appropriate field who will complete their disser- [email protected]. Please include text, and indicate the issue/or tation within one year of hire. Evidence of excel- lence in teaching and scholarly work is required. issues where the ad will appear, and a contact name and number. Excellent written and oral communication skills Estimates: An insertion order will then be e-mailed back to you. The ad are required. Industry and/or consulting experi- will by typeset according to CACM guidelines. NO PROOFS can be sent. ence is highly desirable. As our programs host a Classified line ads are NOT commissionable. diverse population, the proven ability to work ef- fectively in a multicultural environment is highly Deadlines: 20th of the month/2 months prior to issue date. For latest regarded. deadline info, please contact: Marist College is an independent and com- [email protected] prehensive liberal arts institution located in New York’s historic Hudson River Valley. Situated on Career Opportunities Online: Classified and recruitment display ads 210 acres overlooking the Hudson River, it en- receive a free duplicate listing on our website at: rolls 4,962 traditional undergraduate, 953 full http://jobs.acm.org and part-time graduate and 412 continuing edu- cation students. Marist has a branch campus in Ads are listed for a period of 30 days. Florence, Italy, and extension sites throughout New York. Marist is recognized for excellence by For More Information Contact: U.S. News & World Report, TIME Magazine, The ACM Media Sales Princeton Review’s The Best 376 Colleges, and at 212-626-0686 or Barron’s Best Buys in College Education and is [email protected] noted for being a pioneer in the area of on-line degree programs. To learn more or to apply, please visit http://

104 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 jobs.marist.edu. Only on-line applications are ac- will need to participate in teaching, research and San José State University cepted. Review of applications will begin immedi- service. Missouri State University (MSU) is located Assistant or Associate Professor ately and will continue until the position is filled. in Springfield, Missouri. More information about Marist College is strongly committed to the princi- MSU can be found at: http://www.missouristate. The Computer Engineering Department at San ple of diversity and is especially interested in receiv- edu/. Information about the department, its José State University (SJSU) invites applications ing applications from members of ethnic and racial programs and research endeavors can be found for two tenure-track faculty positions at the rank minority groups, women, persons with disabilities, at: http://computerscience.missouristate.edu/ of Assistant or Associate Professor (10 mo. Apt.) and persons from other under-represented groups. undergraduate/. Areas of interest include virtual and augmented AN EQUAL OPPORTUNITY/AFFIRMATIVE Review of applications will begin November reality, machine learning and artificial intelli- ACTION EMPLOYER 15, 2017 and continue until finalists are identified. gence, robotics, cloud computing and virtualiza- To see the required qualifications, complete tion, big data, networking, mobile systems, cyber list of duties and online application go to: https:// security, computer systems architecture, FPGA, Max Planck Institute for Software Sys- jobs.missouristate.edu/. and embedded systems, other areas in computer tems (MPI-SWS) Employment will require a criminal back- and software engineering will be considered. Ap- Tenure-Track Openings ground check at University expense. plicants must have a doctorate in Computer/Soft- Missouri State University is an equal oppor- ware Engineering, Computer Science, Electrical Applications are invited for faculty positions at all tunity/affirmative action/minority/female/veter- Engineering or a closely related field by the start career stages in computer science, with a particular ans/disability/sexual orientation/gender identity of the appointment. For more information about emphasis on systems (broadly construed). We employer and institution. We encourage appli- the position and to apply, go to: http://apptrkr. expect multiple positions to be filled in systems, but cations from all interested minorities, females, com/1081339. exceptional candidates in other areas of computer veterans, individuals with disabilities, and sexual SJSU is an Affirmative Action/ Equal Oppor- science are also strongly encouraged to apply. orientation/gender identity. tunity committed to the core values of inclu- A doctoral degree in computer science or re- sion, civility, and respect for each individual. A lated areas and an outstanding research record background check (including a criminal records (commensurate for the applicant’s career stage) Purdue University check) must be completed satisfactorily before are required. Successful candidates are expected Tenure-Track and Tenured Positions at the As- any candidate can be offered a position with the to build a team and pursue a highly visible re- sistant, Associate and Full Professor Levels CSU. search agenda, both independently and in col- laboration with other groups. The Department of Computer Science at Purdue MPI-SWS is part of a network of over 80 Max University is in a phase of significant growth. San José State University Planck Institutes, Germany’s premier basic-re- Applications are being solicited for nine tenure- Four Tenure-Track Positions search organisations. MPIs have an established track and tenured positions at the Assistant, As- Rank: Assistant Professor (tenure-track)Job record of world-class, foundational research in sociate and Full Professor levels. Outstanding Opening ID (JOID): 24229 the sciences, technology, and the humanities. candidates in all areas of computer science will The institute offers a unique environment that be considered. Review of applications and candi- Qualifications: combines the best aspects of a university depart- date interviews will begin in September 2017, and ˲˲Applicants must have earned their Ph.D. in ment and a research laboratory: Faculty enjoy will continue until the positions are filled. Computer Science or a closely related field and full academic freedom, lead a team of doctoral The Department of Computer Science offers a have demonstrated excellence in teaching and students and post-docs, and have the opportunity stimulating academic environment. Information scholarship. to teach university courses; at the same time, they about the department and a description of open po- ˲˲Applicants should have awareness of and sen- enjoy ongoing institutional funding in addition sitions are available at http://www.cs.purdue.edu. sitivity to the educational goals of a multicultural to third-party funds, a technical infrastructure Applicants should hold a PhD in Computer population as might have been gained in cross- unrivaled for an academic institution, as well as Science or a related discipline, have demonstrated cultural study, training, teaching and other com- internationally competitive compensation. excellence in research, and strong commitment to parable experience. The institute is located in the German cities of teaching. Successful candidates will be expected to ˲˲Preference will be given to candidates with Saarbruecken and Kaiserslautern, in the tri-border conduct research in their fields of expertise, teach teaching, research, and/or industry experience area of Germany, France, and Luxembourg. We courses in computer science, and participate in in the following areas: Data Structures and Al- maintain an international and diverse work envi- other department and university activities. Purdue gorithms, Data Science, Databases, Computer ronment and seek applications from outstanding University’s Department of Computer Science is Graphics, Human Computer Interaction, Cyber- researchers worldwide. The working language is committed to advancing diversity in all areas of security, and Computer Science Education. English; knowledge of the German language is not faculty effort, including scholarship, instruction, ˲˲Special attention will be given to candidates required for a successful career at the institute. and engagement. Candidates should address at with teaching and research experience in inter- Qualified candidates should apply on our ap- least one of these areas in their cover letter, indi- disciplinary fields. plication website (apply.mpi-sws.org). To receive cating their past experiences, current interests or full consideration, applications should be re- activities, and/or future goals to promote a climate Responsibilities: ceived by December 1st, 2017. that values diversity, and inclusion. Salary and The department’s faculty members are expected The institute is committed to increasing the benefits are competitive, and Purdue is a dual-ca- to teach, supervise, and advise students in both representation of minorities, women, and indi- reer friendly employer. Applicants are strongly en- the undergraduate and graduate programs, and viduals with physical disabilities. We particularly couraged to apply online at https://hiring.science. to establish a successful research program re- encourage such individuals to apply. The initial purdue.edu. Alternatively, hardcopy applications lated to his/her field of interest. Additionally, tenure-track appointment is for five years; it can can be sent to: Faculty Search Chair, Department faculty members will participate in department, be extended to seven years based on a midterm of Computer Science, 305 N. University Street, Pur- college, and university committee and other ser- evaluation in the fourth year. A permanent con- due University, West Lafayette, IN 47907. vice assignments. Candidate must address the tract can be awarded upon a successful tenure A background check will be required for em- needs of a student population of great diversity evaluation in the sixth year. ployment in this position. Purdue University is – in age, cultural background, ethnicity, primary an EOE/AA employer. All individuals, including language, and academic preparation – through minorities, women, individuals with disabilities, course materials, teaching strategies, and advise- Missouri State University and veterans are encouraged to apply. ment. Computer Science Department Head Requirements: PhD in Computer Science, or a Salary Range: Commensurate with qualifica- closely related discipline, be committed to excel- tions and experience. The Department of Computer Science at Missouri lence in teaching, and have demonstrated poten- Starting Date: August 17, 2018. State University seeks a Department Head. As well tial for excellence in research. Eligibility: Employment is contingent upon as administrative duties, the department head proof of eligibility to work in the United States.

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 105 CAREERS

Application Procedures: All materials are due be instrumental in the continuing development Applications for both positions will continue by Friday, December 15, 2017. For full consider- of a world-renowned interdisciplinary research to be accepted after these dates until the posi- ation upload a letter of application, curriculum program synergistic with the ongoing research tions are filled. vitae, statement of teaching interests/philosophy, in the Department and the Lyle School of Engi- The Computer Science Department currently description of research plans, and at least three neering. He/she will possess strong administra- has eight tenure-track faculty and four visiting letters of reference with contact information at tive skills and will be an outstanding communi- faculty. Faculty teach introductory courses as well apply.interfolio.com/44419. cator representing the Computer Science and as advanced courses in their research areas. We Please include Job Opening ID (JOID) on all Engineering Department and the Lyle School have grown significantly in both faculty and stu- communications. on- and off-campus. The anticipated starting dents in the last five years. Presently, we are one Search Committee Chair, date is on or before August 1, 2018. Candidates of the most popular majors at the College and Department of Computer Science must have a Ph.D. degree in Computer Science expect to have over 70 Computer Science majors San José State University or a closely related field and must be qualified graduating this year. One Washington Square for a tenured appointment at the full Professor San José, CA 95192-0249 level. Candidates may apply for both positions. The Computer Science and Engineering QUALIFICATIONS: Applicants must have teach- San José State University is California’s oldest (CSE) department resides within the Lyle School ing experience and should be comfortable teach- institution of public higher learning. The campus of Engineering and is located in the Caruth Hall ing a wide range of courses at the introductory is located on the southern end of San Francisco Engineering Building, a LEED Gold designed and intermediate level. Candidates should ad- Bay in downtown San José (Pop. 967,000), hub of facility. The CSE department offers BS, MS, and ditionally have a strong commitment to involv- the world famous Silicon Valley high-technology Ph.D. degrees in Computer Science and in Com- ing undergraduates in their research. A Ph.D. in research and development center. SJSU is an puter Engineering, BA in Computer Science, MS Computer Science at or near the time of appoint- Equal Opportunity/Affirmative Action Employer in Software Engineering, MS in Security Engi- ment is required. The strongest candidates will committed to the core values of inclusion, civility, neering, and D.Eng in Software Engineering. To be expected to demonstrate a commitment to and respect for each individual. learn more about the rich cultural environment creative teaching and an active research program For further details, please visit www.cs.sjsu.edu. of SMU, please see: http://smu.edu. that speaks to and motivates undergraduates While applications and nominations will from diverse backgrounds. be accepted until a new Chair and Professor of APPLICATION INSTRUCTIONS: Applications Santa Clara University Computer Science and Engineering is selected, should include a cover letter, vita, teaching state- Two Tenure-Track Assistant Professors in interested parties are encouraged to submit ment, research statement, and three letters of Computer Science their application by November 30, 2017 elec- reference, at least one (preferably two) of which tronically to [email protected]. Applications should speak to the candidate’s teaching ability. The Department of Mathematics and Computer should include a cover letter, curriculum vitae, In your cover letter, please briefly describe your Science at Santa Clara University invites appli- list of five references and a statement of interest current research agenda; what would be attractive cations for two tenure-track assistant professor and capabilities related to academic leadership, to you about teaching diverse students in a liberal positions in computer science. Our highest pref- education and research. The anticipated start arts college environment; and what background, erence is in candidates with research interests in date for this position is on or before August 1, experience, or interests are likely to make you a an area related to cybersecurity for the first posi- 2018. Hiring is contingent upon the satisfactory strong teacher of Swarthmore College students. tion and an area related to algorithms for the sec- completion of a background check. SMU will not This institution is using Interfolio’s Faculty ond position. Strong candidates with research discriminate in any employment practice, educa- Search to conduct this search. Applicants to this interests in artificial intelligence and software tion program, or educational activity on the basis position receive a free Dossier account and can aspect of data science will be considered as well. of race, color, religion, national origin, sex, age, send all application materials, including confi- The successful candidates will demonstrate not disability, genetic information, or veteran sta- dential letters of recommendation, free of charge. only potential for excellent undergraduate teach- tus. SMU will not discriminate in any program To apply, visit https://apply.interfolio.com/45234 ing, but also promise in sustained research with or activity on the basis of race, color, religion, Swarthmore College actively seeks and wel- opportunities to involve undergraduates, men- national origin, sex, age, disability, genetic in- comes applications from candidates with ex- toring or recruiting underrepresented groups formation, veteran status, sexual orientation, or ceptional qualifications, particularly those with in computer science, and service to the depart- gender identity and expression. The Executive demonstrable commitments to a more inclusive ment, College or University. Positions available Director for Access and Equity/Title IX Coordina- society and world. Swarthmore College is an starting in September 2018. Ph.D. or equivalent tor is designated to handle inquiries regarding Equal Opportunity Employer. Women and mi- required by September 2018. nondiscrimination policies and may be reached norities are encouraged to apply. The closing date for applications is Decem- at the Perkins Administration Building, Room ber 1, 2017 at 3 pm Pacific time. Undergraduate 204, 6425 Boaz Lane, Dallas, TX 75205, 214-768- teaching only. 3601, [email protected]. University of Alabama Santa Clara University, located in California’s Computer Science Faculty Position Silicon Valley, is a comprehensive, Jesuit, Catho- lic university, and an AA/EEO employer. Swarthmore College The Department of Computer Science at the For more information or to apply, visit https:// Tenure-Track and Visiting Faculty Positions in University of Alabama invites applications for a jobs.scu.edu/postings/6211. Computer Science tenure-track faculty position at the Assistant or Associate level to begin either January or August The Computer Science Department invites ap- 2018. Candidates will be expected to engage with Southern Methodist University plications for one tenure-track position and mul- faculty and researchers in the Alabama Water Chair and Professor, Computer Science and tiple visiting positions at the rank of Assistant Institute (http://awi.ua.edu/). Research areas of Engineering Professor to begin Fall semester 2018. interest include, but are not limited to, manage- For the tenure-track position, we are inter- ment and manipulation of large sensor data sets, Southern Methodist University (SMU) invites ested in applicants whose areas fit broadly into real-time (and near real-time) data processing, nominations and applications for the position systems (including but not limited to operating resource constrained data collection and analy- of Chair and Professor, Computer Science and systems, security, or high-performance comput- sis, high performance computing, big data, and Engineering (Position No: 5781). It is expected ing) or programming languages. Priority will be robotics command-and-control. that the appointment will be at the tenured Full given to complete applications received by No- Located in Tuscaloosa, Alabama, the Univer- Professor level. The Chair will be the intellectual vember 17, 2017. For the visiting position, strong sity of Alabama enrolls over 37,000 students and leader of the Computer Science and Engineering applicants in any area will be considered. Priority is the capstone of higher education in the State. Department with strong interest in educational will be given to complete applications received by Housed in the College of Engineering, the Com- programs at the BS, MS, and PhD levels, and will February 2, 2018. puter Science Department has 24 faculty mem-

106 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 bers (16 tenured/tenure-track faculty), roughly and identify the department or departments for didates should address their background, and 700 undergraduates and approximately 50 grad- their potential tenure home and the joint ap- identify the department for their potential tenure uate students. The Department has produced pointments they would desire. When applying, home and any desired joint appointments. 33 doctoral graduates in the last five years and have all documents ready so they can be attached An equal opportunity/affirmative action em- is funded by agencies such as NSF, Google, De- at that time, as the system does not allow resub- ployer, UCF encourages all qualified applicants partments of Education and Commerce, various mittal to update applications. to apply, including women, veterans, individuals Defense agencies, multiple State agencies, and As an equal opportunity/affirmative action with disabilities, and members of traditionally other sponsors. In 2013, the College completed employer, UCF encourages all qualified appli- underrepresented populations. construction of a $300M Shelby Engineering and cants to apply, including women, veterans, indi- Questions can be directed to the search com- Science Complex. viduals with disabilities, and members of tradi- mittee chair, Gary T. Leavens, at Leavens@ucf. Applicants should apply online at http:// tionally underrepresented populations. edu. facultyjobs.ua.edu requisition # 0810845. Ap- For questions, please contact the Cluster’s plicants must have an earned doctorate (Ph.D.) Search Committee Chair, Gary T. Leavens, at in computer science or a closely related field. [email protected]. University of Central Florida (UCF) The application package should include a cover Assistant or Associate Professor, Computer letter, curriculum vitae, and the names of three Science references. The University of Alabama is an equal University of Central Florida opportunity/affirmative action employer. Women Cluster Lead, Cyber Security and Privacy The Department of Computer Science (CS) at the and minority applicants are particularly encour- Cluster University of Central Florida (UCF) is seeking ap- aged to apply. plicants for two faculty positions with an antici- The University of Central Florida (UCF) is recruit- pated start date of August 8, 2018. The positions ing a lead for its cluster on cyber security and will carry the rank of assistant or associate profes- University of Central Florida privacy. This position has a start date of August 8, sor. Rank (and tenure for associate professors) Assistant or Associate Professor in Faculty 2018. The position will carry a rank of associate will be based on the candidate’s prior experience Cluster for Cyber Security and Privacy or full professor, commensurate with the candi- and record. date’s prior experience and record. The lead is ex- The department is particularly interested in The University of Central Florida (UCF) is recruit- pected to have credentials and qualifications like candidates with experience in the areas of human ing a tenure-track assistant or associate professor those expected of a tenured associate or full pro- computer interaction, virtual reality, robotics, for its cyber security and privacy cluster. This po- fessor. To obtain tenure, the selected candidate data science, algorithms, theory of computing, sition has a start date of August 8, 2018. must have a demonstrated record of teaching, financial technology, and software engineering This will be an interdisciplinary position that research and service commensurate with rank. and systems. However, all relevant technical ar- will be expected to strengthen both the cluster This will be an interdisciplinary position that eas will be considered. The ideal candidate will and a chosen tenure home department, as well will be expected to strengthen both the cluster have a strong research background and be on an as a possible combination of joint appointments. and a chosen tenure home department, as well upward leadership trajectory in their research The candidate can choose a combination of units as a possible combination of joint appointments. area. They will have research impact, as reflected from the cluster for their appointment (see http:// The candidate can choose a combination of units in high-quality publications and the ability to www.ucf.edu/faculty/cluster/cyber-security-and- from the cluster for their appointment. (See build a well-funded research program. privacy/). http://www.ucf.edu/faculty/cluster/cyber-securi- The CS Department is home to the first com- The ideal junior candidates will have a strong ty-and-privacy/.) Both individual and interdisci- puter science Ph.D. program in the state. Its 38 background in cyber security and privacy, and be plinary infrastructure and startup support will be tenured and tenure-track faculty are engaged on an upward leadership trajectory in these ar- provided. in world-class research in Computer Vision, AI eas. They will have research impact, as reflected The ideal candidate will have a strong back- and Machine Learning, Virtual Reality, HCI, data in high-quality publications and the ability to ground in cyber security and privacy and out- analytics, cyber security and privacy, and several build a well-funded research program. All rel- standing research credentials and research im- other areas. The department has both CS and IT evant technical areas will be considered. We are pact, as reflected in a sustained record of high undergraduate degrees accredited by ABET, M.S. looking for a team player who can help bring to- quality publications and external funding. All degrees in CS, Digital Forensics, and Data Analyt- gether current campus efforts in cyber security or relevant technical areas will be considered in- ics and, a Ph.D. in CS. To learn more about the privacy. In particular, we are looking for someone cluding: network security, cryptography, block- department see http://www.cs.ucf.edu/. who will work at the intersection of several areas, chains, hardware security, trusted computing UCF is one of the nation’s largest universi- such as: (a) hardware and IoT security, (b) explain- bases, cloud computing, human factors, anomaly ties. As an economic engine, UCF attracts and ing and predicting human behavior, creating detection, forensics, privacy, and software secu- supports vital industry to Orlando. UCF is lo- policies, studying ethics, and ensuring privacy, rity, as well as applications of security and privacy cated at the center of the Florida High Tech Cor- (c) cryptography and theory of security or privacy, to areas such as IoT, cyber-physical systems, fi- ridor where industries include software, defense, or (d) tools, methods, training, and evaluation of nance, and insider threats. A history of working space, simulation and training, and entertain- human behavior. with teams, especially teams that span multiple ment. Next to UCF is a thriving research park Minimum qualifications include a Ph.D., ter- disciplines, is a strongly preferred qualification. that conducts over $2 billion in funded research. minal degree, or foreign degree equivalent from A record of demonstrated leadership is highly Great weather, easy access to the seashore, one of an accredited institution in an area appropriate desired, as we are looking for a leader to bring the largest convention centers in the nation, and to the cluster, and a record of high impact re- together all the current campus efforts in cyber one of the world’s best airports are just a few fea- search related to cyber security and privacy, dem- security and privacy. This includes three cluster tures that make Orlando an ideal location. Learn onstrated by a strong scholarly and/or funding re- members already hired, as well as a pending hire more about UCF at http://www.ucf.edu/faculty. cord. A history of working with teams, especially for the 2017-18 academic year. As an equal opportunity/affirmative action teams that span multiple disciplines, is a strongly Minimum qualifications include a Ph.D. from employer, UCF encourages all qualified applicants preferred qualification. The position will carry a an accredited institution in an appropriate area, to apply, including women, veterans, individuals rank commensurate with the candidate’s prior and a record of high impact research related to cy- with disabilities, and members of traditionally un- experience and record. ber security and privacy demonstrated by a strong derrepresented populations. UCF’s Equal Oppor- Candidates must apply online at https://www. scholarly publication record and a significant tunity Statement can be viewed at: http://www.eeo. jobswithucf.com/postings/50404 and attach the amount of sustained funding. ucf.edu/documents/PresidentsStatement.pdf. As a following materials: a cover letter, curriculum vi- Candidates must apply online at http://www. Florida public university, UCF makes all applica- tae, teaching statement, research statement, and jobswithucf.com/postings/50044 and upload the tion materials and selection procedures available contact information for three professional refer- following materials: cover letter, CV, teaching and to the public upon request. ences. In the cover letter candidates must address research statements, and contact information for Candidates must apply online at www.job- their background in cyber security and privacy, 3 professional references. In the cover letter, can- swithucf.com and attach the following materials:

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 107 CAREERS

a cover letter, curriculum vitae, teaching state- will be requested for candidates invited for on- Applications received by December 15, 2017 ment, research statement, and contact informa- campus interview. For more information, contact: will be given priority consideration. tion for three professional references. Dr. Songlin Tian, Search Committee Chair The University of Chicago has the highest NOTE: Please have all documents ready when School of Computer Science and standards for scholarship and faculty quality, is applying so they can be attached at that time. Mathematics dedicated to fundamental research, and encour- Once the online submission process is finalized, University of Central Missouri ages collaboration across disciplines. We encour- the system does not allow applicants to submit Warrensburg, MO 64093 age connections with researchers across campus additional documents at a later date. (660) 543-4930 in such areas as bioinformatics, mathematics, For questions regarding this opportunity, [email protected] molecular engineering, natural language pro- please contact the department via email at cs- cessing, statistics, public policy, and social sci- [email protected]. Initial screening of applications begins No- ence to mention just a few. Applicants must have a Ph.D. from an ac- vember 15, 2017, and continues until position is The Department of Computer Science (cs. credited institution in an area appropriate to the filled. uchicago.edu) is the hub of a large, diverse com- department, including Computer Science, Com- AA/EEO/ADA. Women and minorities are en- puting community of two hundred researchers puter Engineering, or Mathematics by the time of couraged to apply. focused on advancing foundations of computing the appointment. UCM is located in Warrensburg, MO, which is and driving its most advanced applications. The In order to obtain tenure, the selected candi- 35 miles southeast of the Kansas City metropoli- larger computing and data science community date must have a demonstrated record of teach- tan area. It is a public comprehensive university at the University of Chicago includes the Depart- ing, research and service commensurate with with about 13,000 students. The School of Com- ment of Statistics, the Computation Institute, the rank in the tenure department. puter Science and Mathematics offers undergrad- Toyota Technological Institute at Chicago (TTIC), uate and graduate programs in both Computer the Polsky Center for Entrepreneurship and Inno- Science and Cybersecurity. vation, the Mansueto Institute for Urban Innova- University of Central Missouri tion and the Argonne National Laboratory. Assistant Professor of Computer Science - The Chicago metropolitan area provides a di- Tenure Track University of Chicago verse and exciting environment. The local econ- Assistant Professor/Associate Professor/ omy is vigorous, with international stature in The School of Computer Science and Mathemat- Professor, Computer Science banking, trade, commerce, manufacturing, and ics at the University of Central Missouri is accept- transportation, while the cultural scene includes ing applications for one tenure-track position in The Department of Computer Science at the diverse cultures, vibrant theater, world-renowned Computer Science at the rank of Assistant Profes- University of Chicago invites applications from symphony, opera, jazz, and blues. The University sor. The appointment will begin August 2018. We qualified candidates for faculty positions at the is located in Hyde Park, a Chicago neighborhood are looking for faculty excited by the prospect of ranks of Assistant Professor, Associate Professor, on the Lake Michigan shore just a few minutes shaping our school’s future and contributing to and Professor. The University of Chicago has em- from downtown. its sustained excellence. barked on an ambitious, multi-year effort to sig- The University of Chicago is an Affirmative Ac- The Position: Duties will include teaching un- nificantly expand its computing and data science tion/Equal Opportunity/Disabled/Veterans Em- dergraduate and graduate courses in computer activities. Candidates with research interests in ployer and does not discriminate on the basis of science and cybersecurity and developing new all areas of computer science will be considered. race, color, religion, sex, sexual orientation, gen- courses depending upon the expertise of the ap- However, applications are especially encouraged der identity, national or ethnic origin, age, status plicant and school needs, conducting research in the areas of AI and Machine Learning, Robot- as an individual with a disability, protected veter- which leads toward peer-reviewed publications ics, Data Analytics, Human-Computer Interac- an status, genetic information, or other protected and/or externally funded grants, and program ac- tion, and Visual Computing. classes under the law. For additional information creditation/assessment. Faculty are expected to Candidates must have demonstrated excel- please see the University’s Notice of Nondiscrimi- assist with school and university committee work lence in research and a strong commitment to nation at http://www.uchicago.edu/about/non_ and service activities, and advising majors. teaching. Completion of all requirements for a discrimination_statement/. Job seekers in need Required Qualifications: Ph.D. in Computer Science or a related field is of a reasonable accommodation to complete the ˲˲Ph.D. in Computer Science by August 2018 required at the time of appointment. Candidates application process should call 773-702-0287 or ˲˲Research expertise and/or industrial experi- for Associate Professor and Professor positions email [email protected] with ences in Cybersecurity must have demonstrated leadership in their field, their request. ˲˲Demonstrated ability to teach existing courses have established an outstanding independent re- at the undergraduate and graduate levels search program and have a record of excellence in ˲˲Ability to develop a quality research program teaching and student mentorship. University of Illinois at Urbana- and secure external funding Applications must be submitted through the Champaign ˲˲Supporting the school’s plans to establish a University’s Academic Jobs website. Positions in Computing National Security Agency designated Center for To apply for Assistant Professor, go to http:// Academic Excellence in Information Assurance/ tinyurl.com/ya46ybql The Department of Electrical and Computer En- Cyber Defense To apply for Associate Professor, go to http:// gineering (ECE ILLINOIS) at the University of Il- ˲˲Commitment to engage in curricular devel- tinyurl.com/ydgx33eu linois at Urbana-Champaign invites applications opment/assessment at the undergraduate and To apply for Professor, go to http://tinyurl. for faculty positions at all areas and levels in com- graduate levels com/yaqpar49 puting, broadly defined, with particular empha- ˲˲A strong commitment to excellence in teach- sis on cybersecurity and reliability; embedded ing, research, and continued professional growth To be considered as an applicant, the follow- systems, cyber-physical systems, and the internet ˲˲Excellent verbal and written communication ing materials are required: of things; networked and distributed comput- skills ˲˲cover letter ing; data-centric computing systems and stor- ˲˲curriculum vitae including a list of publications age; quantum computing; robotics and machine The Application Process: To apply online, go ˲˲statement describing past and current research vision; machine learning and AI; and bio com- to https://jobs.ucmo.edu. Apply to position accomplishments and outlining future research putation, and health, among other areas. Appli- #997374. The following items should be plans cations are encouraged from candidates whose attached: a letter of interest, a curriculum vitae, ˲˲description of teaching philosophy and experi- research programs specialize in core as well as a teaching and research statement, copies of ence interdisciplinary areas of electrical and computer transcripts, and a list of at least three professional ˲˲the names of at least three references engineering. From the transistor and the first references including their names, addresses, computer implementation based on von Neu- telephone numbers and email addresses. Official Reference letter submission information will mann’s architecture to the Blue Waters petascale transcripts and three letters of recommendation be provided during the application process. computer (the fastest computer on any university

108 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 campus), ECE ILLINOIS has always been at the https://www.eecs.umich.edu/eecs/etc/csejobs/. including the NSF CAREER award, the MIT TR35 forefront of computing research and innovation. The University of Michigan is a Non-Discrimina- award, honorable mention in the ACM disser- ECE ILLINOIS is in a period of intense demand tory/Affirmative Action Employer. tation competition, multiple Google research and growth, serving over 3000 students and aver- awards, and best paper designations at top-tier aging 7 new tenure-track faculty hires per year in conferences. In 2015 we were one of only two recent years. It is housed in its new 235,000 sq. ft. University of Notre Dame CS departments nationwide to secure three NSF net-zero energy design building, which is a major Multiple Tenure-Track Faculty Positions at all CRII awards for junior faculty. campus addition with maximum space and mini- Ranks The department is deeply committed to build- mal carbon footprint. ing a more diverse and representative faculty, and Qualified senior candidates may also be con- The Department of Computer Science and strongly encourages applications from groups sidered for tenured full Professor positions as Engineering at the University of Notre Dame underrepresented in higher education. We have part of the Grainger Engineering Breakthroughs invites applications for multiple tenure-track a vibrant community, and Initiative (graingerinitiative.engineering.illinois. faculty positions at all ranks, with one position are a charter member of the ABI/HMC BRAID edu), which is backed by a $100-million gift from specifically in circuits, architecture, or related Initiative. With funding from the NSF, the CRA, the Grainger Foundation. areas. We seek to attract, develop, and retain and major industrial sponsors, BRAID works to Please visit http://jobs.illinois.edu to view the excellent faculty members with strong records increase diversity and inclusivity in the under- complete position announcement and applica- and future promise. The Department is especially graduate program and to rigorously evaluate fac- tion instructions. Full consideration will be given interested in candidates who will contribute to tors that contribute to change. In 2017, women to applications received by November 15, 2017, the diversity and excellence of the University’s constituted 33% of our BA/BS graduates, and we but applications will continue to be accepted un- academic community through their research, are actively working to improve the environment til all positions are filled. teaching, and service. for other underrepresented groups. Illinois is an EEO Employer/Vet/Disabled www. The Department offers the Ph.D. degree and The University of Rochester is a private, Tier inclusiveillinois.illinois.edu. undergraduate Computer Science and Computer I research institution with approximately 5,000 The University of Illinois conducts criminal Engineering degrees. Faculty are expected to undergraduates and a comparable number of background checks on all job candidates upon excel in classroom teaching and to lead highly- graduate students. It has recently committed acceptance of a contingent offer. visible research projects that attract substantial $50M to the multidisciplinary Goergen Insti- external funding. More information about the tute for Data Science (GIDS), of which Computer department can be found at: http://cse.nd.edu/. Science is the leading departmental member The University of Michigan, Ann Arbor Applicants must submit a CV, a teaching state- — and with which it shares a newly constructed Department of Electrical Engineering and ment, a research statement, and contact informa- state-of-the-art facility. Ongoing hiring in GIDS Computer Science tion for three professional references at http:// provides exciting opportunities for collabora- Computer Science and Engineering Division apply.interfolio.com/41330. To guarantee full tion between computing and other disciplines. Faculty Positions consideration, applications must be received by Anchoring the Finger Lakes region of west- December 1, 2017, however, review of applications ern New York State, the greater Rochester area is The University of Michigan Computer Science will continue until the positions have been filled. home to over a million people, and offers unsur- and Engineering (CSE) Division expects strong The University is an Equal Opportunity passed quality of life, with a thriving arts scene, growth in the coming years and invites applica- and Affirmative Action employer; we strongly outstanding public schools, affordable housing, tions for multiple tenure-track positions at all encourage applications from women, minori- and a huge range of cultural and recreational levels. These positions include, but are not lim- ties, veterans, individuals with a disability and opportunities. Traditionally strong in optics re- ited to, cross-disciplinary areas and an endowed those candidates attracted to a university with a search and manufacturing, the area was recently professorship (the Fischer Chair) in theoretical Catholic identity. selected by the Department of Defense as the hub computer science. Exceptional candidates from of a $600M Integrated Photonics Institute for all areas of computer science and computer engi- Manufacturing Innovation. neering will be considered. Qualifications include University of Rochester The University of Rochester is an Equal Op- an outstanding academic record, a doctorate or Faculty Positions in Computer Science portunity Employer: equivalent in computer science or computer en- EOE Minorities/Females/Protected Veterans/ gineering, and a strong commitment to teaching The Computer Science Department at the Univer- Disabled and research. The college is especially interested sity of Rochester seeks applicants for two tenure- The University of Rochester, an Equal Op- in candidates who can contribute, through their track positions. Outstanding candidates will be portunity Employer, has a strong commitment research, teaching, and/or service, to the diversity considered in any area of computer science and to diversity and actively encourages applications and excellence of the academic community. at any level of seniority. We are particularly eager from candidates from groups underrepresented The University of Michigan is one of the to grow our strength in human-computer interac- in higher education. world’s leading research universities with annual tion and in the theory and practice of security and EOE Minorities/Females/Protected Veterans/ research funding of well over $1 billion. It con- privacy. Disabled sists of highly-ranked departments and colleges Candidates must have (or be about to receive) across engineering, sciences, medicine, law, a doctorate in computer science or a related dis- business, and the arts. More than a quarter of cipline. Applications should be submitted online University of Toronto CSE faculty have sponsored collaborations with (at https://www.rochester.edu/faculty-recruiting/ Assistant Professor - Electrical and Computer faculty in other units. The CSE Division is vibrant login) no later than January 1, 2018, for full con- Engineering and innovative, with over 50 world-class faculty sideration; submissions beyond this date risk be- members, over 300 graduate students, several ing overlooked due to limited interview slots. The Edward S. Rogers Sr. Department of Electri- Research Centers, and a large and illustrious net- The Department of Computer Science cal and Computer Engineering (ECE) at the Uni- work of alumni. Ann Arbor is well known as one (https://www.cs.rochester.edu) has a distin- versity of Toronto invites applications for up to of the best college towns in the country. The Uni- guished history of research in artificial intelli- four tenure-stream faculty appointments at the versity of Michigan has a strong dual-career assis- gence, HCI, systems, and theory. We nurture a rank of Assistant Professor. The appointments tance program. highly collaborative and interdisciplinary cul- will commence on July 1, 2018. We encourage candidates to apply as soon ture, with exceptionally strong external funding Within the general field of electrical and com- as possible. For best consideration for Fall 2018, and with active ties to numerous allied depart- puter engineering, we seek applications from please apply by December 1, 2017. Positions re- ments, including brain and cognitive science, candidates with expertise in one or more of the main open until filled and applications can be linguistics, biomedical engineering, electrical following strategic research areas: 1. Computer submitted throughout the year. and computer engineering, and several depart- or Communications Engineering, with prefer- For more details on these positions and to ments in the medical center. Recent faculty ence for a focus on machine learning, computer apply, please visit the Application Web Page at hires have received a host of national honors, security and privacy, or data science; 2. Electrical

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 109 CAREERS

Power Systems, with preference for a focus on versity of Toronto invites applications for up to committees or human resources staff. Results power systems protection; 3. Systems Control, four tenure-stream faculty appointments at the will be aggregated for institutional planning pur- with preference for a focus on robotics. rank of Associate Professor. The appointments poses. For more information, please see http:// Applicants are expected to have a Ph.D. in will commence on July 1, 2018. uoft.me/UP. Electrical and Computer Engineering, or a relat- Within the general field of electrical and com- All qualified candidates are encouraged to ed field, at the time of appointment or soon after. puter engineering, we seek applications from apply; however, Canadians and permanent resi- Successful candidates will be expected to ini- candidates with expertise in one or more of the dents will be given priority. tiate and lead an independent research program following strategic research areas: 1. Computer of international calibre, and to teach at both the or Communications Engineering, with prefer- undergraduate and graduate levels. Candidates ence for a focus on machine learning, computer US Air Force Academy should have demonstrated excellence in research security and privacy, or data science; 2. Electrical Assistant Professor of Computer Science and teaching. Excellence in research is evidenced Power Systems, with preference for a focus on primarily by publications in leading journals or power systems protection; 3. Systems Control, The Department of Computer Science at the US conferences in the field, presentations at signifi- with preference for a focus on robotics. Air Force Academy seeks to fill a full-time faculty cant conferences and strong endorsements by Applicants are expected to have a Ph.D. in position at the Assistant Professor level. The de- referees of high international standing. Evidence Electrical and Computer Engineering, or a relat- partment is particularly interested in candidates of excellence in teaching will be demonstrated ed field, and have at least five years of academic or with backgrounds in artificial intelligence, com- by strong communication skills, a compelling relevant industrial experience. puter and network security, operations research, statement of teaching submitted as part of the Successful candidates will be expected to or unmanned aerial systems, but all candidates application highlighting areas of interest and initiate and lead an independent, competi- with a passion for undergraduate computer sci- accomplishments, as well as strong letters of rec- tive and innovative research program of inter- ence teaching are encouraged to apply. ommendation. national calibre, and to teach at both the un- The Academy is a national service institu- Eligibility and willingness to register as a Pro- dergraduate and graduate levels. Candidates tion, charged with producing lieutenants for fessional Engineer in Ontario is highly desirable. should have demonstrated excellence in re- the US Air Force. Faculty members are expected The Edward S. Rogers Sr. Department of Elec- search and teaching. Excellence in research is to exemplify the highest ideals of professional- trical and Computer Engineering at the Univer- evidenced primarily by publications in leading ism and character. USAFA is located in Colo- sity of Toronto ranks among the best in North journals or conferences in the field, presenta- rado Springs, an area known for its exceptional America. It attracts outstanding students, has ex- tions at significant conferences and a high pro- natural beauty and quality of life. The United cellent facilities, and is ideally located in the mid- file in the field with strong endorsements by ref- States Air Force Academy values the benefits of dle of a vibrant, artistic, diverse and cosmopoli- erees of high international standing. Evidence diversity among the faculty to include a variety tan city. Additional information may be found at of excellence in teaching will be demonstrated of educational backgrounds, professional and http://www.ece.utoronto.ca. by strong communication skills, a compelling life experiences. Review of applications will begin after Oc- statement of teaching submitted as part of the For information on how to apply, go to tober 3, 2017, however, the position will remain application highlighting areas of interest and usajobs.gov and search with the keyword open until December 11, 2017. You must submit accomplishments, as well as strong letters of 478670300. your application online, by following the submis- recommendation. sion guidelines given at http://uoft.me/how-to- Eligibility and willingness to register as a Pro- apply. Applications submitted in any other way fessional Engineer in Ontario is highly desirable. York University will not be considered. The Edward S. Rogers Sr. Department of Elec- Assistant Professor As part of your online application, please in- trical and Computer Engineering at the Univer- clude a curriculum vitae, a summary of your previ- sity of Toronto ranks among the best in North The Department of Electrical Engineering and ous research and future research plans, as well as America. It attracts outstanding students, has ex- Computer Science, York University, is seeking an a statement of teaching experience and interests. cellent facilities, and is ideally located in the mid- outstanding candidate at the rank of Assistant Applicants must arrange for three letters of refer- dle of a vibrant, artistic, diverse and cosmopoli- Professor in the area of Computer Systems al- ence to be sent directly by the referees (on letter- tan city. Additional information may be found at though exceptional applicants from other areas head, signed and scanned), by email to the ECE http://www.ece.utoronto.ca. in Computer Science may also be considered. The department at [email protected]. Review of applications will begin after Oc- successful candidate will have a PhD in Computer The University of Toronto is strongly com- tober 3, 2017, however, the position will remain Science, or a closely related field, and a research mitted to diversity within its community and es- open until December 11, 2017. You must submit record commensurate with rank. The appoint- pecially welcomes applications from racialized your application online, by following the submis- ment will commence on July 1, 2018, subject to persons / persons of colour, women, Indigenous / sion guidelines given at http://uoft.me/how-to- budgetary approval. For full position details, see Aboriginal People of North America, persons with apply. Applications submitted in any other way http://www.yorku.ca/acadjobs. disabilities, LGBTQ persons, and others who may will not be considered. Applicants should complete the on-line pro- contribute to the further diversification of ideas. As part of your online application, please in- cess at http://lassonde.yorku.ca/new-faculty/. A As part of your application, you will be asked clude a curriculum vitae, a summary of your previ- complete application includes a cover letter indi- to complete a brief Diversity Survey. This survey is ous research and future research plans, as well as cating the rank for which the candidate wishes to voluntary. Any information directly related to you a statement of teaching experience and interests. be considered, a detailed CV, statement of contri- is confidential and cannot be accessed by search Applicants must arrange for three letters of refer- bution to research, teaching and curriculum de- committees or human resources staff. Results ence to be sent directly by the referees (on letter- velopment, three sample research publications will be aggregated for institutional planning pur- head, signed and scanned), by email to the ECE and contact information for three referees. Com- poses. For more information, please see http:// department at [email protected]. plete applications must be received by November uoft.me/UP. The University of Toronto is strongly com- 30, 2017. All qualified candidates are encouraged to mitted to diversity within its community and es- York University is an Affirmative Action em- apply; however, Canadians and permanent resi- pecially welcomes applications from racialized ployer and strongly values diversity, including dents will be given priority. persons / persons of colour, women, Indigenous gender and sexual diversity, within its communi- / Aboriginal People of North America, persons ty. The Affirmative Action Program, which applies with disabilities, LGBTQ persons, and others to women, Aboriginal people, visible minorities University of Toronto who may contribute to the further diversifica- and people with disabilities, can be found at Associate Professor - Electrical and Computer tion of ideas. http://acadjobs.info.yorku.ca/affirmative-action/ Engineering As part of your application, you will be asked or by calling the AA office at 416.736.5713. All to complete a brief Diversity Survey. This survey is qualified candidates are encouraged to apply; The Edward S. Rogers Sr. Department of Electri- voluntary. Any information directly related to you however, Canadian Citizens and Permanent Resi- cal and Computer Engineering (ECE) at the Uni- is confidential and cannot be accessed by search dents will be given priority.

110 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 last byte

[CONTINUED FROM P. 101] its effects was impossible. INTERACTIONS But now he considered a slightly But there was greater perturbation—not a butterfly’s one spot that was flutter but a disturbance as large as a tennis court. A change in the parameters yellow . . . where of one cell in that mother of all spread- the local wind speed sheets. Could such a perturbation have a significant effect anywhere else? had increased Eliot decided on an experiment. First, by a factor of two. he ran a model based on the latest grid data—nothing new there—predicting A factor of two! the weather 20 minutes into the future. Then he went back into the input matri- ces and manipulated the parameters in ACM’s Interactions magazine one cell, the one that covered his back- explores critical relationships yard. He then lowered the temperature “You mean like the shooting of Arch- between people and by five degrees and ran the model again. duke Franz Ferdinand?” technology, showcasing A simple program subtracted the “Who?” emerging innovations and two results, color-coding any differ- “1914. Franz Ferdinand was assassi- industry leaders from around ences. Blue indicated cells with no nated by a fervent Yugoslavian national- the world across important change, green very slight differences, ist. Small event. But within months, they applications of design thinking and yellow and red locations that were were digging trenches in France. The and the broadening eld of significantly altered. Because weather, First World War. interaction design. carried by the wind, doesn’t move fast- “No trenches here, Lily. And no arch- Our readers represent a growing er than the speed of sound, he looked dukes either.” He pulled the gate on the community of practice that is at the results only within 250 miles. pickup, and three tons of water spilled of increasing and vital global Beyond that, Eliot would need a pre- out onto the parched earth. importance. diction further into the future, and In 20 minutes, Eliot was looking at couldn’t be sure if any effects were due his difference plots. Despite the fact that to his manipulated data or not. he had lowered the surface temperature The difference plot was mostly just a of his yard considerably more than the noisy sea of blue and green pixels. A few five degrees of his numerical test, there were red, indicating, for example, the was still no apparent effect within 250 barometric pressure was 1% greater or miles. This butterfly had flapped its lesser than it would have been without wings for naught. the backyard temperature change. His What Eliot couldn’t see, and wouldn’t theoretical butterfly hadn’t had much understand until much later, was that six effect. But there was one spot that was hours after he flooded the yard, a Russian yellow . . . where the local wind speed military transport carrying diplomats had increased by a factor of two. A factor would run into trouble on its approach of two! Eliot pushed back his chair and to Warsaw’s Bemowo Airfield. The let out a low whistle. flight forecast had predicted smooth Lily watched her husband inch the skies all the way. And they were, except rented pickup down the driveway, stop- at 3,000 feet short of the runway where an ping just short of the desiccated waste- unusually forceful dust devil lifted one land that was their yard. He had lined its wing and tossed the plane on its back. To learn more about us, bed with plastic and spent several hours Was it an accident? The Russians cer- visit our award-winning website filling it with water from the kitchen tainly didn’t think so. This was subtle http://interactions.acm.org sink. sabotage, and no degree of pleading “Are you going to tell me what this is could convince them otherwise. In the Follow us on Facebook and Twitter about, Eliot? I mean, is this your idea of predawn skies the following day, an un- making rain?” manned drone headed west, hugging To subscribe: “I’m going for a real live butterfly the flat landscapes of Northern Europe. http://www.acm.org/subscribe effect.” Someone had finally done something “Like when you feel nervous? When about the weather. your tummy turns upside down?” Association for “Not at all. It’s when a small stimu- Seth Shostak ([email protected]) is the senior astronomer at Computing Machinery the SETI Institute in Mountain View, CA. lus produces a big effect. Consequences that greatly outweigh causes.” © 2017 ACM 0001-0782/17/11 $15.00

NOVEMBER 2017 | VOL. 60 | NO. 11 | COMMUNICATIONS OF THE ACM 111

IX_XRDS_ThirdVertical_V01.indd 1 3/18/15 3:35 PM last byte

From the intersection of computational science and technological speculation, with boundaries limited only by our ability to imagine what could be.

DOI:10.1145/3140960 Seth Shostak Future Tense Butterfly Effect But, like the weather, what can anyone do about it?

LILY’S EYES SCANNED the yard, an ex- pansive tract of suburban real estate she called the back 40. She was not pleased. “Eliot, this is embarrassing. Our property makes the Dust Bowl look lush. Is it ever going to rain again?” Her husband turned toward the sky as if seeking an answer. But of course he already knew what his response would be, and so did she. “October, Lily . . . It rains in October.” “I don’t care about seasonal behav- ior or what’s normal for the state. If it doesn’t rain in Longmont or Loveland, well, tough for them. I just care about this backyard patch. Make it rain here, will you? You’re the meteorologist.” “Yes, dear, I am,” Eliot replied, flash- ing a slight smile in the interests of do- mestic tranquility, and went inside. This development was inevitable, re- So Lily’s remarks bothered him. Despite the fact that Eliot had a quiring only the ability to build a finer Was he truly helpless? All his life sheepskin testifying to his meteorologi- grid of weather data—wind, tempera- people had teased him with Charles cal chops Lily’s gibe was a reminder he ture, barometric pressure—and the Dudley Warner’s bromide that every- wasn’t the man he had once hoped to be. compute power to crunch it all. Both one complains about the weather, but He wasn’t really a weather forecaster. In were now at hand. no one does anything about it. But he high school, he could look at the clouds, Improved satellites had refined was a scientist. He knew that weather sense the temperature, and know a front the grid by a factor of 20 in all direc- involved immense highly energetic sys- was moving in. Some of his friends ad- tions. The whole planet—continents, tems. The output of 100 power plants mired his ability to predict afternoon ocean surface, the entire atmosphere was nothing compared to a hurricane’s thunderstorms, while others thought he below six miles—was now sampled terawatts. How could anyone do any- was just obsessive. Eliot didn’t care much. on a scale of 300 feet. Every few min- thing about the weather? It was akin to He enjoyed the fact that his skills could be utes, the weather was measured and moving the Rocky Mountains. tested every day, his predictions verified binned into five trillion cells, the He also knew the weather was a sys- or disproved within hours. It was like mother of all spreadsheets. tem that was chaotic and close to equi- being a day trader, but without the risk. It was Eliot’s job to feed this vast librium. A small change could have big Eliot followed his meteorological in- anthill of numbers into the models sev- consequences. The butterfly effect. terests through grad school, but by the eral times a day and bring to bear the The idea that a butterfly could late 2030s technology was rendering his compute power available in the Extend- precipitate a storm was a popular idea, skills obsolete. Weather forecasts were ed Cloud. Yes, he had to understand recognized centuries ago. Eliot had increasingly the province of computer what he was doing, and, yes, he had thought about it in school but reck- models, massive calculations that spit to be careful. But it wasn’t traditional oned the flapping wings of a single out accurate predictions for any place weather forecasting, and definitely not insect really couldn’t do much. And

on the planet, down to an acre or less. weather manipulation. predicting [CONTINUED ON P. 111] ASSOCIATES/SHUTTERSTOCK ANDRIJ BORYS BY IMAGE

112 COMMUNICATIONS OF THE ACM | NOVEMBER 2017 | VOL. 60 | NO. 11 CONFERENCE 27 – 30 November 2017 EXHIBITION 28 – 30 November 2017 BITEC, Bangkok, Thailand

THE CELEBRATION OF LIFE & TECHNOLOGY

The 10th ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia

Register online by 15 October 2017, & enjoy early bird discounts of up to 20% SA2017.SIGGRAPH.ORG/REGISTRATION

Sponsored by Organized by Middleware 2017 Dec 11 – 15, Las Vegas We invite you to attend the 18th ACM/IFIP/USENIX International Middleware Conference 2017 covering recent scientific advances in middleware systems. The conference will showcase an exciting agenda including a single-track technical program, invited speakers, workshops, tutorials, demos & posters, doctoral symposium, and social events with researchers from academia and industry. Keynotes Prof. Magdalena Balazinska, Jean Loup Baer Professor of Computer Science & Engineering Performance SLAs for Cloud Data Analytics

Dr. Ricardo Bianchini, MSR Redmond Chief Efficiency Strategist & Manager of the Systems Research Group Toward intelligent cloud platforms: the Resource Central experience

Tutorials • High Performance Network Middleware with Intel DPDK and OpenNetVM • Istio Service mesh for more robust, secure and easy to manage microservices • SMACK stack 101: Building Fast Data stacks • Trusted Execution of Software using Intel SGX

Workshops • Active: International Workshop on Active Middleware on Modern Hardware • ARM: Adaptive and Reflective Middleware • DIDL: Workshop on Distributed Infrastructures for Deep Learning • M4IoT: Middleware and Applications for the Internet of Things • MECC: Middleware for Edge Clouds & Cloudlets • SERIAL: ScalablE and Resilient InfrAstructures for distributed Ledgers • WoSC: Workshop on Serverless Computing 2017.middleware-conference.org @middleware2017