Professional Documents
Culture Documents
Communications of ACM 2019 November Vol 62 No. 11
Communications of ACM 2019 November Vol 62 No. 11
ACM
CACM.ACM.ORG OF THE 11/2019 VOL.62 NO.11
Special Section
on India Region
An Interview with
Leonard Kleinrock
DeepXplore: Automated Whitebox
Testing of Deep Learning Systems
When Drones Fly Association for
Computing Machinery
The 13th ACM SIGGRAPH Conference and
Exhibition on Computer Graphics and Interactive
Techniques in Asia
Driving Diversity
SA2020.SIGGRAPH.ORG
#SIGGRAPHAsia | #SIGGRAPHAsia2020
Sponsored by Organized by
TCS Research
Positions for PhDs:
fresh and mid-career
Tata Consultancy Services (TCS) is a part of the Tata group, a multinational business group with interests across
industry segments. TCS is an IT services, consulting and business solutions provider that has been partnering with the
world’s largest businesses in their transformation journeys for the last fifty years.
TCS operates on a global scale, with diverse talent base representing 149 nationalities, across 46 countries. TCS has been
recognized as a Global Top Employer for the fourth consecutive year (2018-19) by the Top Employers Institute, and the
Number One Top Employer in four regions – North America, Europe, Asia Pacific, and the Middle East.
TCS has been a pioneer in Software Research and has continued to systematically invest in Research over decades.
TCS Researchers seek to make an impact and solve real-world problems for global Fortune 1000 companies.
Towards this goal, TCS Research conducts applied industrial research in the following areas: Behavioral, Business,
and Social Sciences, Computing Systems, Cybersecurity and Privacy, Data and Decision Sciences, Deep Learning and
AI, Embedded Systems and Robotics, Foundations of Computing, Life Sciences, Media and Advertising, Physical
Sciences, Software Systems and Services.
We seek researchers who will advance our capabilities in our core research areas, contribute to global thought
leadership and create an intellectual foundation to address current and future business and technology opportunities.
Who can apply?
Applicants should have a PhD from a premier University/Institute related to the Research Areas mentioned above.
Fresh PhDs as well as mid-career researchers are invited to apply.
How to apply:
Write to us at careers.research@tcs.com with a detailed CV highlighting your interest area of research, education
and publications. Please mark “Research Careers” in the subject line.
5 Cerf’s Up
Hazards of the
Information Superhighway
By Vinton G. Cerf
7 Vardi’s Insights
The Winner-Takes-All
Tech Corporation
By Moshe Y. Vardi
10 BLOG@CACM
16 28
The Benefits of Indolence
Yegor Bugayenko explains
his realization that software 13 Information Is Physics 24 Legally Speaking
developers should go neither Individual bits of information can Europe’s Controversial Digital
above nor beyond. have direct physical consequences. Copyright Directive Finalized
By Don Monroe Considering the new liability risks
27 Calendar for ISPs, search engines, and news
16 When Drones Fly aggregators under recent EU-wide
146 Careers Drone technology is poised to enter mandatory rules.
the mainstream of business By Pamela Samuelson
and society, but engineering robust
Last Byte controls remains a challenge. 28 Education
By Samuel Greengard Computational Thinking Should
160 Future Tense Just be Good Thinking
Cantando con la Corrente 19 Real-World Applications for Drones Seeking to change computing teaching
(Singing with Current) Unmanned vehicles have a number to improve computer science.
An augmented singer gets of compelling real-world use cases. By Mark Guzdial, Alan Kay,
some unexpected feedback By Logan Kugler Cathie Norris, and Elliot Soloway
from his audience.
By P-Ray 31 Interview
An Interview with Leonard Kleinrock
The UCLA professor and networking
pioneer reflects on his career in
industry and academia.
By George Varghese
37 Viewpoint
Algorithms, Platforms,
and Ethnic Bias
How computing platforms
IMAGES BY: ( L) RIKO BEST; (R) OLLYY
Review Articles
40 104 122 An Elementary Introduction
to Kalman Filtering
40 The nations within the India 104 The Effects of Mixing Machine Demystifying the uses of a powerful
region are key players in the global Learning and Human Judgment tool for uncertain information.
computing vista. This special Collaboration between humans By Yan Pei, Swarnendu Biswas,
section presents a collection of and machines does not necessarily Donald S. Fussell, and Keshav Pingali
articles written by leading voices lead to better outcomes.
from the region, sharing stories of By Michelle Vaccaro and Jim Waldo
innovative technologies, software Research Highlights
advancements, educational goals, 111 Write Amplification
and much more. vs. Read Perspiration 136 Technical Perspective
The trade-offs between A Whitebox Solution
Watch the co-organizers write and read. for Blackbox-Like Behaviors
discuss this section
in the exclusive
By Pat Helland By David G. Andersen
Communications video.
https://cacm.acm.org/ Articles’ development led by
videos/india-region-special- 137 DeepXplore: Automated Whitebox
queue.acm.org
section Testing of Deep Learning Systems
IMAGES BY: ( L) KONSTANTIN FA RA KT INOV; (R) M IK HA IL LEO NOV, COLORIZED BY ANDRIJ BORYS ASSOCIAT ES
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 computer science, 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 STA F F EDITORIAL BOARD ACM Copyright Notice
and scientific computing society, delivers DIRECTOR OF PU BL ICATIONS E DITOR- IN- C HIE F Copyright © 2019 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 cacm-publisher@cacm.acm.org eic@cacm.acm.org Permission to make digital or hard copies
computing field’s premier Digital Library Deputy to the Editor-in-Chief of part or all of this work for personal
and serves its members and the computing Executive Editor Lihan Chen or classroom use is granted without
profession with leading-edge publications, Diane Crawford cacm.deputy.to.eic@gmail.com fee provided that copies are not made
conferences, and career resources. Managing Editor S E NIOR E DITOR or distributed for profit or commercial
Thomas E. Lambert Moshe Y. Vardi advantage and that copies bear this
Executive Director and CEO Senior Editor notice and full citation on the first
Vicki L. Hanson Andrew Rosenbloom NE W S page. Copyright for components of this
Deputy Executive Director and COO Senior Editor/News Co-Chairs work owned by others than ACM must
Patricia Ryan Lawrence M. Fisher Marc Snir and Alain Chesnais be honored. Abstracting with credit is
Director, Office of Information Systems Web Editor Board Members permitted. To copy otherwise, to republish,
Wayne Graves David Roman Monica Divitini; Mei Kobayashi; to post on servers, or to redistribute to
Director, Office of Financial Services Editorial Assistant Rajeev Rastogi; François Sillion lists, requires prior specific permission
Darren Ramdin Danbi Yu and/or fee. Request permission to publish
Director, Office of SIG Services VIE W P OINTS from permissions@hq.acm.org or fax
Donna Cappo Art Director Co-Chairs (212) 869-0481.
Director, Office of Publications Andrij Borys Tim Finin; Susanne E. Hambrusch;
Scott E. Delman Associate Art Director John Leslie King; Paul Rosenbloom For other copying of articles that carry a
Margaret Gray Board Members code at the bottom of the first or last page
Assistant Art Director Terry Benzel; Michael L. Best; Judith Bishop; or screen display, copying is permitted
ACM CO U N C I L
Mia Angelica Balaquiot Lorrie Cranor; Boi Falting; James Grimmelmann; provided that the per-copy fee indicated
President
Production Manager Mark Guzdial; Haym B. Hirsch; in the code is paid through the Copyright
Cherri M. Pancake
Bernadette Shade Richard Ladner; Carl Landwehr; Beng Chin Ooi; Clearance Center; www.copyright.com.
Vice-President
Intellectual Property Rights Coordinator Francesca Rossi; Len Shustek; Loren Terveen;
Elizabeth Churchill
Barbara Ryan Marshall Van Alstyne; Jeannette Wing; Subscriptions
Secretary/Treasurer
Advertising Sales Account Manager Susan J. Winter An annual subscription cost is included
Yannis Ioannidis
Ilia Rodriguez in ACM member dues of $99 ($40 of
Past President
P R AC TIC E which is allocated to a subscription to
Alexander L. Wolf
Chair, SGB Board Columnists Co-Chairs Communications); for students, cost
Jeff Jortner David Anderson; Michael Cusumano; Stephen Bourne and Theo Schlossnagle is included in $42 dues ($20 of which
Co-Chairs, Publications Board Peter J. Denning; Mark Guzdial; Board Members is allocated to a Communications
Jack Davidson and Joseph Konstan Thomas Haigh; Leah Hoffmann; Mari Sako; Eric Allman; Samy Bahra; Peter Bailis; subscription). A nonmember annual
Members-at-Large Pamela Samuelson; Marshall Van Alstyne Betsy Beyer; Terry Coatta; Stuart Feldman; subscription is $269.
Gabriele Anderst-Kotsis; Susan Dumais; Nicole Forsgren; Camille Fournier;
Renée McCauley; Claudia Bauzer Mederios; C O N TAC T P O IN TS Jessie Frazelle; Benjamin Fried; Tom Killalea; ACM Media Advertising Policy
Elizabeth D. Mynatt; Pamela Samuelson; Copyright permission Tom Limoncelli; Kate Matsudaira; Communications of the ACM and other
Theo Schlossnagle; Eugene H. Spafford permissions@hq.acm.org Marshall Kirk McKusick; Erik Meijer; ACM Media publications accept advertising
SGB Council Representatives Calendar items George Neville-Neil; Jim Waldo; in both print and electronic formats. All
Sarita Adve and Jeanna Neefe Matthews calendar@cacm.acm.org Meredith Whittaker advertising in ACM Media publications is
Change of address at the discretion of ACM and is intended
C ONTR IB U TE D A RTIC LES
BOARD C HA I R S acmhelp@acm.org to provide financial support for the various
Letters to the Editor Co-Chairs activities and services for ACM members.
Education Board James Larus and Gail Murphy
letters@cacm.acm.org Current advertising rates can be found
Mehran Sahami and Jane Chu Prey Board Members by visiting http://www.acm-media.org or
Practitioners Board William Aiello; Robert Austin; Kim Bruce;
W E B S IT E by contacting ACM Media Sales at
Terry Coatta Alan Bundy; Peter Buneman; Jeff Chase;
http://cacm.acm.org (212) 626-0686.
Andrew W. Cross; Yannis Ioannidis;
REGIONA L C O U N C I L C HA I R S WEB BOARD Gal A. Kaminka; Ben C. Lee; Igor Markov; Single Copies
ACM Europe Council Chair Lionel M. Ni; Adrian Perrig; Doina Precup; Single copies of Communications of the
Chris Hankin James Landay Marie-Christine Rousset; Shankar Sastry; ACM are available for purchase. Please
ACM India Council Board Members m.c. schraefel; Ron Shamir; Sebastian Uchitel; contact acmhelp@acm.org.
Abhiram Ranade Marti Hearst; Jason I. Hong; Hannes Werthner; Reinhard Wilhelm
ACM China Council Jeff Johnson; Wendy E. MacKay COMMUN ICATION S OF THE ACM
RES E A R C H HIGHLIGHTS
Wenguang Chen (ISSN 0001-0782) is published monthly
Co-Chairs
AU T H O R G U ID E L IN ES by ACM Media, 1601 Broadway, 10th Floor
Azer Bestavros, Shriram Krishnamurthi,
PUB LICATI O N S BOA R D http://cacm.acm.org/about- New York, NY 10019-7434 USA. Periodicals
and Orna Kupferman
Co-Chairs communications/author-center postage paid at New York, NY 10001,
Board Members
Jack Davidson and Joseph Konstan Martin Abadi; Amr El Abbadi; and other mailing offices.
Board Members ACM ADVERTISIN G DEPARTM E NT Animashree Anandkumar; Sanjeev Arora;
Phoebe Ayers; Chris Hankin; Mike Heroux; 1601 Broadway, 10th Floor Michael Backes; Maria-Florina Balcan; POSTMASTER
Nenad Medvidovic; Tulika Mitra; New York, NY 10019-7434 USA David Brooks; Stuart K. Card; Jon Crowcroft; Please send address changes to
Michael L. Nelson; Sharon Oviatt; T (212) 626-0686 Alexei Efros; Bryan Ford; Alon Halevy; Communications of the ACM
Eugene H. Spafford; Stephen N. Spencer; F (212) 869-0481 Gernot Heiser; Takeo Igarashi; 1601 Broadway, 10th Floor
Divesh Srivastava; Robert Walker; Srinivasan Keshav; Sven Koenig; New York, NY 10019-7434 USA
Julie R. Williamson Advertising Sales Account Manager Ran Libeskind-Hadas; Karen Liu; Greg Morrisett;
Ilia Rodriguez Tim Roughgarden; Guy Steele, Jr.;
ACM U.S. Technology Policy Office ilia.rodriguez@hq.acm.org Printed in the USA.
Robert Williamson; Margaret H. Wright;
Adam Eisgrau, Nicholai Zeldovich; Andreas Zeller
Director of Global Policy and Public Affairs Media Kit acmmediasales@acm.org
1701 Pennsylvania Ave NW, Suite 200,
S P EC IA L S EC TIONS
Washington, DC 20006 USA
T (202) 580-6555; acmpo@acm.org Association for Computing Machinery Co-Chairs
(ACM) Sriram Rajamani, Jakob Rehof,
Computer Science Teachers Association 1601 Broadway, 10th Floor and Haibo Chen A
SE
REC
Y
CL
PL
Executive Director T (212) 869-7440; F (212) 869-0481 Tao Xie; Kenjiro Taura; David Padua
NE
TH
S
I
Z
I
M AGA
Hazards of the
Information Superhighway
I
N THE 1990S, U.S. Vice President groups against one another for the be repeatedly validated. In the global
Al Gore characterized the purpose of disrupting democracy. Internet, there is a universe of sources
Internet as an “Information The propagandists were not inter- and finding quality brands is made all
Superhighway.” This meta- ested in one group or another prevail- the more difficult by the scale of the
phor has some utility as we try ing as much as they wanted to sow problem. Given the critical nature of
to understand emerging properties distrust of democratic institutions, the Internet’s search engines as tools
of the global Internet. More recently, disrupt rational and civil discourse, for discovery of World Wide Web con-
an old friend, Judith Estrin, touted and generally increase domestic ten- tent, it seems inescapable that the
the importance of friction in the on- sions among groups with potentially presentation of search results not
line environment. She had two things conflicting agendas. only must be prioritized by some mea-
in mind, I believe. The first is that It is tempting to think such mis- sure of quality but also that the rank-
friction slows things down and some- chief would be obvious to those ex- ing criteria must be clear and well un-
times that is exactly what is needed to posed to these campaigns but we are derstood. Transparency is our friend
give time to think about the content human and being human we are sub- in this endeavor. This also applies to
found on the Internet, especially in ject to effects such as group think and sources of information. Unvalidated
social media. Friction also keeps you confirmation bias. We grow comfort- sources or anonymous sources should
on the road and not spinning off at able with our beliefs and those of like- be considered less trustworthy than
every turn. As reports of the deliberate minded people, so much so that even strongly authenticated ones. This
injection of misinformation and dis- in the face of clear evidence, we may does not mean, however, that even a
information into the Internet contin- be more likely to reject factual refuta- well-known source should be taken
ue to escalate, my attention has been tion of our positions than to change at face value. Just because a source is
drawn to efforts to counter this trend. our minds and our positions. Indeed, well identified does not mean it car-
I went back and re-read the May 2019 there is some evidence that factual ries valid information.
report about the Finnish response to rebuttals may generate increased in- Ultimately, this takes us back to
information pollution,a which has tolerance of views opposing our own, critical thinking and the need for
garnered attention from other coun- despite their factual basis. multiple reinforcing sources. There
tries and organizations concerned The Finnish antidote is to train its may be serious disagreements among
about this phenomenon. citizens to think critically about what legitimate sources of information as
The Finnish response centers on they see and hear; to ask questions is often the case in scientific disputes.
critical thinking and teaching citi- about corroborating evidence; to ex- The solution to those problems al-
zens of all ages to ask probing ques- plore and uncover the sources of con- most always relies on obtaining more
tions about information they gather troversial statements. That this takes factual information and better inter-
whether online or offline. Propagan- real work is evident. Students report pretive theories. This should be the
da is intended to steer the recipient’s the effort is sometimes onerous. None- essence of democratic discourse and
thinking into the directions intended theless, it strikes me that such effort should not be replaced by fabricated
by its source. Interestingly, the so- is an obligation derived from living information intended to mislead and
called weaponization of informa- in a democratic society. The price we derail genuine search for truth.
tion need not be unidirectional. The pay for the freedom of access to in-
disinformation campaigns allegedly formation that we enjoy on the open Vinton G. Cerf is vice president and Chief Internet
Evangelist at Google. He served as ACM president
conducted by Russia against the U.S., Internet is the need for due diligence from 2012–2014.
France, and the U.K., for example, applied to the sources of information
were often designed to pit opposing we rely upon.
Not surprisingly, brand can be-
a https://www.cnn.com/interactive/2019/05/ come a key indicator of quality of in-
europe/finland-fake-news-intl/ formation if the branded source can Copyright held by author.
The 2020 IEEE World Congress on Services (SERVICES) will be held on July 6-11, 2020 in Beijing, China. The
Congress is solely sponsored by the IEEE Computer Society under the auspice of the Technical Committee
on Services Computing (TCSVC). The scope of the Congress will cover all aspects of services computing and
applications, current or emerging. It covers various systems and networking research pertaining to cloud,
edge and Internet-of-Things (IoT), as well as technologies for intelligent computing, learning, big data
and blockchain applications, while addressing critical issues such as high performance, security, privacy,
dependability, trustworthiness, and cost-effectiveness. The Congress will also include symposia and workshops
supporting deep-dive discussions on emerging important topics, and complement the Congress program with
industry and application presentations and panels. Authors are invited to prepare early and submit original
papers to any of these conferences at www.easychair.org. All submitted manuscripts will be peer-reviewed by
at least 3 reviewers. Accepted and presented papers will appear in the conference proceedings published by
the IEEE Computer Society Press. The Congress will be organized with the following affiliated conferences and
symposia:
IEEE International Conference on Cloud Computing (CLOUD): The flagship theme-topic conference
for modeling, developing, publishing, monitoring, managing, delivering XaaS (everything as a service) in the
context of various types of cloud environments.
IEEE International Conference on Web Services (ICWS): The flagship theme-topic conference for Web-
based services, featuring Web services modeling, development, publishing, discovery, composition, testing,
adaptation, and delivery, and Web services technologies as well as standards.
IEEE International Conference on Services Computing (SCC): The flagship theme-topic conference
for services innovation lifecycle that includes enterprise and vertical services modeling, microservices based
solution creation, services orchestration, services optimization, services management, services marketing,
and business process integration and management.
IEEE International Conference on Smart Data Services (SmartDataServices): The flagship theme-topic
conference for data driven applications and solutions under the as-a-service model, including analytic services,
smart data foundation, big data services, blockchain, and data computing at the edge and in IoT systems.
Symposium on Digital Health as a Service (DHAASS): DHAASS represents an emerging and critical
direction for SERVICES covering the application of digital health as a service in transforming health and social
care. Key themes for 2020 include definitions and safe implementations of health/medical microservices
(dubbed the Uber of digital health), crowd sensing/sourcing, microservice integration, health service
economics, among others.
Key Dates: Early paper submission due December 2, 2019; Review comments for early-submission papers
provided January 17, 2020; Normal paper submission due February 13, 2020; Final notification to authors
provided April 6, 2020; Camera-ready manuscripts due April 20, 2020
Send inquiries to: ieeecs.services@gmail.com
See: http://conferences.computer.org/services/2020/ for more information.
vardi’s insights
T
of corporate personhood, which gives
HE FIVE LARGEST U.S. corpo- must enforce anti-trust laws against corporations the same legal rights
rations—Alphabet, Amazon, such corporations. enjoyed by natural persons. In fact,
Apple, Facebook, and Micro- Public concerns about overly domi- under U.S. law, some essential rights
soft—are all tech companies nant corporations have been aggra- of the 14th Amendment, which ad-
with combined market capi- vated by what has become a dogma in dresses equal protection of the laws,
talization of over four trillion dollars. the U.S. business community over the belong not only to U.S. citizens but
Tech is often called “Big Tech” these past generation, which is the Share- also to corporations. This has far-
days. Furthermore, a small number of holder-Primacy Principle, asserting reaching implications. For example,
corporations have come to dominate the that shareholders should be assigned the U.S. Supreme Court ruled in 2010
IT industry, as within each industry seg- a priority relative to all other corpo- that corporate funding of indepen-
ment one corporation often dominates. rate stakeholders, such as employees, dent political broadcasts in candidate
The phenomenon whereby cor- customers, and the like. According to elections cannot be limited under the
porate dominance seems to be en- this view, the goal of a corporation is First Amendment because of corpo-
trenched is often referred to as “win- just to generate profits, period! This rate personhood. This had led to a sig-
ner takes all.” In the context of tech, approach, which has emerged in the nificant flow of corporate funds into
such a phenomenon can be partly ex- 1970s and became dogmatic in the U.S. political campaigns—and money
plained by two “laws:” Metcalfe’s Law 1980s, has replaced the earlier ap- buys influence in politics.
asserts that the effect of a communi- proach of “corporate responsibility,” But the 14th Amendment was passed
cations network is proportional to the which made corporations accountable in response to issues related to former
square of the number of connected to multiple stakeholders. slaves following the American Civil
users. This makes Facebook, with over Sensing public frustration with the War. How it came to be interpreted to
1.5B daily users, dominant as a social narrow profit motive of U.S. corpora- grant personhood to corporations is a
network. Kai-Fu Lee’s Virtuous Cycle tions, the Business Roundtable, an long and convoluted tale. Many argue
asserts “More data begets more users association of close to 200 influential that corporations should not have the
and profit, which begets more usage U.S. CEOs, recently abandoned its 1997 same rights as natural persons. As IBM
and data.” This explains, for example, shareholder-primacy position and CEO Rometty said, society offers corpo-
the dominance of the Google search declared that “the paramount duty of rations a license to operate, so it makes
engine. Metcalfe’s Law and the Virtu- management and boards of directors sense for society to define the terms of
ous Cycle make tech companies into is to the corporation’s stakeholders.” that license, including rights and re-
natural monopolies, some claim. “Society gives each of us a license to op- sponsibilities, the issue of corporate
As I argued earlier this year, we erate,” declared Ginni Rometty, IBM’s personhood, and the relationship be-
need laws and regulations, instead of CEO. “It’s a question of whether society tween shareholders and other stake-
an ethics outrage, to deal with unde- trusts you or not.” holders. Perhaps the time has come to
sired business models and conduct But doubts have been expressed formally define the terms of the rela-
of tech corporations. What may have about whether corporations can be tionship between society and corpora-
been a radical position less than a trusted to regulate themselves, even tions via a constitutional amendment
year ago has become a conventional after their stakeholder-primacy decla- that explicitly addresses the rights and
wisdom now. There are several initia- ration. In a recent book, The Anarchy, responsibilities of corporations.
tives to regulate tech; the question historian William Darlymple describes Follow me on Facebook and Twitter.
now is how rather than if. The biggest the history of the East India Company,
regulatory issue on the table is how to the most successful and most ruthless Moshe Y. Vardi (vardi@cs.rice.edu) is the Karen Ostrum
George Distinguished Service Professor in Computational
deal with overly dominant corpora- start-up in history. “Yet if history shows Engineering and Director of the Ken Kennedy Institute for
tions. In a 2018 book, The Curse of Big- Information Technology at Rice University, Houston, TX, USA.
anything,” write Darlymple, “it is that He is the former Editor-in-Chief of Communications.
ness: Antitrust in the New Gilded Age, in the intricate dance between the pow-
legal scholar Tim Wu argues the U.S. er of the state and that of the corpora- Copyright held by author.
DOI:10.1145/3361696
T
H E V I E W P O I N T C O L U M N “On- uting digital credentials to all voters, Author’s Response:
line Voting: We Can Do It! which Ms. Orman ignores entirely. No research proves that online voting a
(We Have To)” in the Sep- Even if one regards her View- priori defies security principles. The growing
tember 2019 issue is naïve point as a guide to the many diffi- set of innovative tools and techniques
and unscientific. Although cult scientific challenges that must for software verification, trustworthy
the column is explicitly framed as a be overcome before it’s safe to pro- identity credentials, and publicly verified
response to the scientific community ceed with online voting, the con- computation argues the contrary. As in all
of experts who explain the dangers of cluding paragraphs are completely practical solutions, there will be a trade-off
Internet voting, it does not actually pie-in-the-sky. She presumes that between cost and security.
cite any of the scientific literature Ms. we could have secure smartphones My perspective is that the balance point
Orman is claiming to refute. with trusted hardware and software is rapidly shifting, and security researchers
The scientific community (the “9 if only the government would subsi- and professionals need to produce, critique,
out of 10 experts” she mentions) have dize them; as if well-resourced, tech- analyze, and verify high-assurance voting
published many articles and reports nically savvy corporations such as systems. The volatility surrounding these
laying out the scientific basis for why Apple and Google were not already issues should not deter progress.
online voting is inherently insecure busting their butts to make their Hilarie Orman, Woodland Hills, UT, USA
(given any known or imminently fore- phones secure and failing in any
seeable technology).1–7 Yet Ms. Orman case. And Ms. Orman suggests, in
does not cite any of these scientific the very last paragraph, that secure Editor-in-Chief’s response:
papers among the bibliographic cita- TPM+TCB+PKI+(new-standardized- In an era of active election interference by
tions in the References section of her markup-language) could all happen foreign powers in the U.S. and many other
column. Given that Communication’s within five years, by 2024, and be countries, the importance of careful design,
Viewpoint format does not permit an widespread by 2028. That claim is vetting, and validation of online voting
extensive bibliography, she did not where the essential unreality of this systems can’t be overstated. At the same
have room to cite all of references whole scheme becomes clear. With time, U.S. voter participation in national
listed here,1–7 but in a response to the so many intractable scientific prob- elections (the presidential elections every
scientific community it would have lems unresolved—as they are even four years) has been mired in the 50%–60%
been appropriate to cite (and explicitly by Ms. Orman’s own analysis—it is range for past 50 years, so the need for
respond to the science in) at least some irresponsible to suggest pilot proj- technology that could increase participation
of them. ects in elections for public office in democracy are also desirable! This is
There are gaping technical holes within such a short timeframe. an important issue where the experts
at the core of Ms. Orman’s proposal. of the ACM have contributed greatly to
She proposes to rely on Trusted Plat- References understanding and public policy, and there
form Modules (TPM) to secure the 1. The Myth of “Secure” Blockchain Voting. D. is much more to be done.
Jefferson, Oct. 2018; www.verifiedvoting.org/
end-user devices; but TPM cannot jefferson_themythof_secure_blockchainvoting/. Andrew A. Chien, Editor-in-Chief
possibly do that within any foresee- 2. Securing the Vote: Protecting American
Democracy. National Academies of Science,
able future, for two reasons. First, Engineering, and Medicine, Sept. 2018; https://doi.
org/10.17226/25120.
TPM replaces your trust in the de- 3. Email and Internet Voting, The Overlooked Threat
ACM Must Maintain
vice with your trust in the holder of to Election Security. S. Greenhalgh, S. Goodman, P. Profession Neutrality
Rosenzweig, and J. Epstein, Oct. 2018.
the signing key. Intel or Google or 4. The Future of Voting: End-to-End Verifiable Internet
Companies like Google are strong
Samsung or Apple holds the signing Voting—Specification and Feasibility Study. Report supporters of ACM, sponsoring
of the U.S. Vote Foundation, 2015; https://www.
key of your device; shall we let them usvotefoundation.org/sites/default/files/E2EVIV_ ACM’s A.M. Turing Award and en-
choose who wins our elections? full_report.pdf. couraging its employees to become
5. If I Can Shop and Bank Online, Why Can’t I
And even if we did—TPM has been Vote Online? D. Jefferson, 2011; https://www. ACM members. But that support gives
around for 20 years and we still keep verifiedvoting.org/resources/internet-voting/vote- ACM a greater, not lesser, responsi-
online/.
finding security holes in it; it’s simply 6. Recommendations Report to the Legislative bility to maintain objectivity and neu-
not trustworthy. Assembly of British Columbia. The Independent trality. Consequently, I was dismayed
Panel on Internet Voting, 2014; http://bit.
I won’t even begin to explain why ly/2lHEDYS. to read Vinton Cerf’s editorial “Poly-
Blockchain doesn’t solve online vot- 7. Security Analysis of the Estonian Internet Voting glot!” (Sept. 2019), a thinly veiled
System. J.A., Halderman, H. Hursti, et al., 2014;
ing, since that is so well explained in http://bit.ly/2lUlzXf laundry list of all the wonderful things
the scientific literature.1,2 So too is the Google can do: “Google speaks 106 lan-
immensely thorny problem of distrib- Andrew W. Appel, Princeton, NJ, USA guages ... Google’s language ability vastly
ACM
ACM Transactions on Computing
Computing
to serve Google as its “chief evangelist,”
as his byline notes. ACM should not
allow itself to be used as its platform.
Transactions on
Jonathan Grier, Pikesville, MD, USA for
for Healthcare (HEALTH)
Healthcare (HEALTH)
AAmultidisciplinary journal for
multidisciplinary journal for high-quality
high-qualityoriginal
original
Editor-in-Chief’s response:
It’s a good point that ACM aspires
work
work on
on how
how computing
computing isis improving
improvinghealthcare
healthcare
to balance coverage of advanced
technologies from leading academic
researchers, government researchers, Computing
Computing forfor Healthcare has emerged
Healthcare has emergedas asan
animportant
important
companies, and other leaders around
and
and growing
growing research area. By
research area. By using
usingsmart
smartdevices,
devices,the
the
the world. This case was a failure of
expediency and familiarity. Vinton Cerf’s Internet of Things for health, mobile computing, machine
Internet of Things for health, mobile computing, machine
employer certainly has no monopoly learning,
learning, cloud computing
computing andand other
othercomputing
computingbased
based
on advanced technology in language technologies,
technologies, computing for for healthcare
healthcarecancanimprove
improve
translation (for example, Microsoft
Translator, Amazon Translate, Baidu
the effectiveness, efficiency, privacy, safety, and security
the effectiveness, efficiency, privacy, safety, and security
Translate) and image recognition of
ofhealthcare
healthcare (e.g., personalized
personalized healthcare,
healthcare,preventive
preventive
(for example, SenseTime, Amazon healthcare,
healthcare, ICU without walls,
walls, and
andhome
homehospitals).
hospitals).
Rekognition, Bing Visual search). We will
continue to strive to do better!
ACMTransactions
ACM Transactions on Computing
Computingfor forHealthcare
Healthcare(HEALTH)
(HEALTH)
Andrew A. Chien, Editor-in-Chief
isisthe
the premier
premier journal for thethe publication
publicationof ofhigh-quality
high-quality
© 2019 ACM 0001-0782/19/11 $15.00 original research
original research papers, survey papers, and challenge
survey papers, and challenge
papers
papers that that have scientific and
scientific and technological
technologicalresults
results
pertaining to how computing
computing isis improving
improvinghealthcare.
Coming Next Month in COMMUNICATIONS
pertaining healthcare.
The Rise of Serverless This journal
This journal isis multidisciplinary,
multidisciplinary, intersecting
intersectingCS,
CS,ECE,
ECE,
Computing mechanical engineering, bio-medical
mechanical engineering, bio-medical engineering, engineering,
behavioral and
behavioral and social
social science,
science, psychology,
psychology,and andthethe
Automated Program Repair health field, in general. All submissions
health field, in general. All submissions must showmust show
evidence of
evidence of their
their contributions
contributions to tothe
thecomputing
computingfield
field
Rethinking as informed by healthcare.
as informed by healthcare. We do We do
Search Engines and not publish
not publish papers
papers on
on large
large pilot
pilot
studies, diseases, or other
studies, diseases, or other medicalmedical
Recommendation Systems
assessments/results that do not
assessments/results that do not
have novel computing research
Q&A with Garth Gibson have novel computing research
results. Datasets and other artifacts
results. Datasets and other artifacts
needed to support reproducibility
OpenPiton: An Open Source needed to support reproducibility
of results are highly encouraged.
Hardware Platform of results are highly encouraged.
Proposals for special issues are
for Your Research Proposals
encouraged. for special issues are
encouraged.
Hack for Hire
DOI:10.1145/3360907 http://cacm.acm.org/blogs/blog-cacm
The Benefits
of Indolence
Yegor Bugayenko explains his realization that
software developers should go neither above nor beyond.
Yegor Bugayenko Clients want to keep costs low, and code itself. And the clients overseeing
Lazy Developers Are if they can, they will pass costs onto out- the code are also at fault.
the Best Developers side companies. That’s why we decided Once you accept that, you can put
http://bit.ly/2lEC9KE to “get lazy” and only do what we are together a basic report by creating
July 15, 2019 paid to do. We won’t go out of our way new tickets. This report could be
We are taught from a to improve a project, refactor, or fix code lazy-simple:
young age that the hardest workers en- unless we are getting paid for it. ˲˲ There is no documentation for
joy the most success. Hard work pays And when we find ourselves with a Class Y, can’t figure out how it works.
off, or so we are told. But “hard work” task in front of us and we don’t under- ˲˲ Library Z is in use but why aren’t
can be a bit problematic for software stand how to solve it, we usually don’t you using library B?
developers, because it often means go- blame ourselves. This is especially true ˲˲ This algorithm is a complex mess,
ing well above and beyond the original if the problem has something to do can you explain what it does?
scope of the project. with legacy code. See, here’s the thing: ˲˲ The class naming rules are incoher-
This is especially true when it comes we weren’t paid to understand the leg- ent, can you provide documentation?
to understanding legacy code. When acy code. We were paid to add a feature, Suddenly, your initial “report” is in-
you deal with legacy code, you often find solve a bug, or whatever. stead a list of questions. You can’t provide
yourself having to engage in so-called Suddenly becoming experts in a the answers because you don’t honestly
“deep thinking.” You are expected to un- project’s legacy code would be outside know them and you are too lazy to figure
derstand large problem scopes before the scope of our work, and since we’re it out. Answering these questions falls
you even begin trying to fix the small lazy, we’re not going to venture outside outside of the scope of work you were
bugs. For a long time, this stressed me of our assignment unless we’re paid to hired for, so it is reasonable to expect
out. Then I got an idea: be lazy. do so. A project shouldn’t expect you to the client to provide documentation.
At my company, Zerocracy, we prac- be intelligent or tech-savvy, as far as the Now, you might have noticed a com-
tice a #NoAltruism policy. We, quite lit- legacy code is concerned. Instead, you mon thread in the questions here. I
erally, think only about ourselves and need to focus on closing tickets. didn’t ask for help. I didn’t ask some-
our personal profit. This might sound It’s not your fault if the code is a one to create something for me. Pro-
a bit harsh. Isn’t it better to play nice complete mess, or the bug is serious, grammers will often reach out for help,
and try to appease your clients? In an or you can’t estimate how much time it saying something like “which library
ideal world, maybe. But here’s what we will take to understand the legacy code, should I use for this task?”
have learned about clients: they also let alone how to fix the bug. So whose Here’s the thing: your clients aren’t
practice #NoAltruism. fault is it? The first guilty party is the hiring you so they can do your work for
you. They aren’t hiring you so they can There’s a difference between purposeful
be your teacher, either. They don’t real- altruism as a means to improve the system,
ly want to explain anything to you. For “It’s not your fault and blind altruism as a fanatic ideology.
them, it’s money and time they would if the code is The thing we need to keep in mind is, the
rather not spend. human psychology is never without its flaws,
So your goal, then, is to get your cli- a complete mess, no matter how hardcore a saint you would
ents to fix the code base so that the code or the bug is serious, be trying to play here. I myself have seen
itself becomes more obvious and easier numerous examples of a biased altruist
to read. This will help not only you, but or you can’t estimate doing much more harm than a selfish but
everyone else. As such, focus on asking how much time rational person in a similar situation.
for documentation and code source fixes. Zerocracy is about regulating those
Okay, so you’ve got the tickets out it will take to psychological flaws, not trying to abolish
and you’ve asked the client to fix their understand the them, which would most certainly end in
source code and address other prob- (yet another) wasted effort. Being truthful
lems. So what now? Sit back and relax! legacy code, let alone with oneself, first and foremost, is the
You wait for the tickets to be resolved how to fix the bug.” key in building all sorts of constructive
and don’t sweat who is resolving the is- professional relationships. Ignorance of
sues; that’s not really our business. that is bound to amplify guilt and fear in
Now, your employer may decide to performers many times in the end, which
kick the problem back to you, asking might be appealing to certain moral
you to solve it on your own. That’s fine, fundamentalists who believe a scared
so long as you’re getting paid for it and and advancement of computing systems. programmer’s guilt complex is like some
the employer expands the scope of your Zerocracy is a disgraceful movement for sort of a virtue. The truth is, it just doesn’t
work. Instead of fixing bugs, you’re computing profession. work out like that.
now documenting some functionality —Mehmet Suzen —Ilyas Gasanov
or refactoring this and that.
As you create tickets and blame every- Mehmet, can you please elaborate on how This occurs whether one is a consultant
one else around you, you will continue to exactly “contribute to society” leads to the or contractor, or a salaried employee
create smaller and smaller scopes. Even- conclusion that we are obliged to behave in of the organization that owns the software.
tually, you may find that the tickets can an altruistic manner? I have been both.
be fixed in a half hour or less. And keep —Yegor Bugayenko Even within the organization that owns
in mind, when I say “blaming everyone the software, the deep thinking required
else,” that doesn’t mean shouting at oth- I think this policy is created to end the to document otherwise undocumented
er people. It simply means not beating abuse on the client’s behalf. #NoAltruism systems or to fix underlying design problems
yourself up for problems you didn’t cre- does not mean that in Zerocracy people is discouraged, and the attitude of “fix the
ate, and shifting responsibility for poor- would create software to support terrorism. immediate problem” prevails. This causes
ly written code to the original source. Engineering is not altruistic, is precise. The the organization’s maintenance costs to
Being lazy can take a lot of effort Zerocracy policies are meant to create an increase steadily over time as technical debt
(seriously). We are programmed not efficient culture, not people without values. piles up unaddressed, deeper and deeper.
to be lazy. Some people will resist the I think Mehmet misunderstood what This works similarly to the principle
call. They might feel ashamed (stop #NoAltruism means. of conservation of energy, which pops up
it!). They want to be perfectionists (only —Eduardo Portal Geroy in infinitely varied guises whenever one
perfect what you’re paid to!). Or maybe attempts to create a perpetual motion
you lack the passion needed to be lazy “Computing professionals have the machine: it is always thus regardless of
(get a new job!). obligation to behave in altruistic manner which trendy or modern “methodology”
and help each other for both advancement is used in an attempt to manage the
Comments of business productivity, human well-being, problem solved without doing the actual,
This is unacceptable practice from and advancement of computing systems.” necessary work.
ACM’s professional ethics guidelines. As much as they have an obligation to not In the end, one is doing one’s client
Zerocracy promotes no altruism and waste their time for free, increasing the or one’s employer a disservice by not
no help. This practice violates the core engineering level in the company, helping warning them that a failure to solve the
mission of ACM as an organization, which others do their job, and saving time to help deeper problems will cost far more in the
is “Contribute to society and to human others and contribute to society in their free long run than any immediate savings they
well-being, acknowledging that all people time, doing really altruistic things, not what will realize by ignoring those problems for
are stakeholders in computing.” I request you are talking about. the present.
ACM to retract this article. Computing —Nikita Puzankov —Robert Watkins
professionals have the obligation to
behave in an altruistic manner and help “Computing professionals have the Yegor Bugayenko is founder and CEO of software
engineering and management platform Zerocracy.
each other for both advancement of obligation to behave in an altruistic manner
business productivity, human well-being, and help each other.” © 2019 ACM 0001-0782/19/11 $15.00
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 11
This text/reference is an in-depth introduction to the systematic, universal software
engineering kernel known as “Essence.” This kernel was envisioned and originally created by
Ivar Jacobson and his colleagues, developed by Software Engineering Method and Theory
(SEMAT) and approved by The Object Management Group (OMG) as a standard in 2014.
Essence is a practice-independent framework for thinking and reasoning about the practices
we have and the practices we need. It establishes a shared and standard understanding
of what is at the heart of software development. Essence is agnostic to any particular
methods, lifecycle independent, programming language independent, concise,
scalable, extensible, and formally specified. Essence frees the practices from their
method prisons.
Information Is Physics
Individual bits of information can have direct physical consequences.
E
FFI CI E N T E RROR-CORRE CT I N G
CO D E Sfor quantum comput-
ing recently emerged from
mathematical models used
to study black holes. This sur-
prising finding joins to a long list of
profound connections between infor-
mation and physics.
The most intriguing examples
began as paradoxes or “thought ex-
periments” that are hard to test ex-
perimentally. Physicists take them
seriously because they challenge core
concepts and may require revolution-
ary theoretical changes that could
have practical consequences.
IMAGE BY HID EO KOSAKA ; CO MMU NICAT IONS PH YSICS, 2 , 74 ( 201 9) . D OI: 10.103 8/ S42 005 - 019 - 01 58 - 0
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 13
news
of one bit requires an energy expendi- Actually, “classical” (non-quan- embedded in the infalling material is
ture of a few billionths of a picojoule tum) physics also follows microscop- forever inaccessible. However, in the
(at room temperature). This puts a ic equations that pay no attention to 1970s, Stephen Hawking of the U.K.’s
lower bound on the energy needed the direction of time. Indeed, physi- University of Cambridge suggested
for computation, because the output cists have long struggled to describe that normally ephemeral pairs of
of a logic gate usually compresses how such deterministic processes particles that appear in the quantum-
the information in its inputs. Fortu- lead to the apparent loss of informa- mechanical vacuum could be ripped
nately, current electronics devices tion embodied in increasing entropy, apart at the horizon, with one sucked
use millions of times more energy per since the final state contains all of inside and the other escaping.
operation, so the limit is not (yet) the details needed to reconstruct the One consequence of this escaping
important practically. initial state. “Information is just as “Hawking radiation” is that the black
preserved classically as in quantum,” hole will eventually evaporate com-
Quantum Information said Sean Carroll, a theoretical physi- pletely. At that time (usually ridicu-
Even this fundamental limit could cist at the California Institute of Tech- lously far in the future), information
in principle be avoided, however, by nology (Caltech) in Pasadena, CA. that had been carried in would not
making all computations reversible, just be inaccessible, but gone forever,
retaining enough information to re- Beyond the Horizon violating the quantum rule that it is
construct the original input. For re- It is in the quantum realm, however, always preserved. Physicists argued
searchers working on candidate com- that information has raised the most about how to resolve this “black-hole
ponents for quantum computers, this profound conceptual challenges. This information paradox” for decades, but
turns out to be immediately relevant, is most apparent in the field of quantum they largely came to accept that the in-
because these devices always operate gravity, which aims to reconcile quan- formation was somehow carried away
reversibly, and this needs to be incor- tum mechanics and general relativity. in quantum “entanglement” between
porated in circuit design. Traditionally, quantum mechanics different radiated particles. In 2004,
Indeed, according to quantum me- plays out on a “stage” of unchanging Hawking famously agreed, conceding
chanics, the mathematical evolution spacetime, Swingle said. “If you try to a bet with the California Institute of
of any system is restricted to “unitary” make that stage dynamical, as happens Technology’s John Preskill.
transformations, which “basically in general relativity, where the geome-
means that whatever information you try of spacetime is changing as a func- The Universe as a Hologram
have, it always is there in some form,” tion of time, then combining those two Important support for this consensus
said theoretical physicist Brian Swing- things is hard.” came from a tool proposed in 1997 by
le of the University of Maryland. “May- The gravitational and quantum Juan Maldacena of the Institute for
be it’s very hard to read it out in some frameworks can usually agree to dis- Advanced Study in Princeton, NJ. It is
sense, but it’s there.” agree, since they apply to very large called the AdS/CFT correspondence
Quantum information should still and small scales respectively. How- because it allows a mathematical
be conserved even when it is scrambled ever, their conflict becomes unavoid- mapping between a particular mod-
by interactions with the environment, able for physicists studying black el of spacetime (AdS) and a class of
which can be viewed as a larger quan- holes, which are both extremely mas- quantum models (CFT).
tum system. Although such interactions sive and relatively compact. In this Intriguingly, although the gravita-
are often viewed as uncontrolled noise reconciliation effort, information tional and quantum systems are equiva-
that causes “decoherence” that scram- plays a central role. lent, the quantum system has one fewer
bles quantum information, quantum Once anything falls within a black spatial dimension, somewhat like the
error-correction schemes exploit the hole’s “event horizon,” from which surface of the gravitational system. This
overall reversibility of the combined sys- even light cannot escape, it should is an example of a “holographic” uni-
tem to ensure the desired information have no more influence on outside verse, so called because it resembles the
is preserved where it is needed. space. In particular, any information way that a flat holographic film can en-
capsulate a three-dimensional image.
Black holes constructed in this ideal-
AAA BBB AAA BBB ized universe can evaporate while con-
serving information. Carroll said the
hope is that looking at such explicit
SCH EMATIC BY HT KYM/ WIKIPEDIA ( CC BY- SA 3 .0)
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 15
news
A
matured
S D R O N E S H AV E
into smarter and more
practical machines, they
have hummed, buzzed,
and whirred their way
into industries as diverse as movie
production, agriculture, civil engi-
neering, and insurance. It is entirely
clear that autonomous drones will
play a prominent role in business
in the coming years. Firms such as
Amazon, FedEx, and Uber have ex-
perimented with the technology to
deliver packages, food, and more,
while military agencies, emergency
responders, gaming companies, en-
tertainment firms, and others have
explored other possibilities.
“Drones introduce far more effi-
cient ways to accomplish some tasks,”
says Todd Curtis, president of Airsafe.
com, a site that tracks drone and other
aeronautic technologies.
Powering more advanced drones
are more sophisticated on-board sen- are known. Fog, snow, smoke, and dust Drones Take Flight
sors and processors, better artificial present additional challenges. Engineering a fully autonomous drone
intelligence (AI) algorithms, and more It is one thing to showcase a drone is rife with challenges—particularly in
advanced controllers and communica- in a controlled environment; it is quite busy and complex urban areas.
tion systems. In addition, engineers another to have it operate flawlessly in First, they are not like the autono-
are packing greater numbers of sen- the wild. UAVs must have near-perfect mous vehicles that operate on land.
sors into drones—and using them in vision and sensing, as well as the abil- UAVs have extreme space and weight
different combinations—to create ity to navigate areas where satellite and restrictions. Whereas a car can poten-
greater “awareness” of the surround- communications signals cannot reach tially have dozens, even hundreds, of
ing environment. This sensing, when and need backup and fail-safe systems sensors mounted across its surface, a
combined with GPS and other naviga- that can take control of the drone if/ drone can accommodate the weight
tion capabilities, allows drones to tack- when something goes astray. of only a few.
le more advanced autonomous tasks, “We are seeing remarkable advances Second, UAVs move in almost every
including devices that explore caverns in onboard sensing and processing, direction in a three-dimensional (3D)
or other hard-to-reach spaces, as well but also the use of far more sophisti- space, while a motor vehicle operates on
as underwater drones that conduct re- cated AI (artificial intelligence) algo- a two-dimensional plane. This makes
search by scanning oceans. rithms in drones,” says Nathan Michael, designing software and algorithms for
Yet, despite rapidly evolving capa- associate research professor at the UAVs exponentially more complex.
bilities, it also is clear that autono- Robotics Institute of Carnegie Mellon Finally, the simple fact these ma-
mous drones have not completely University. “These navigation and con- chines are suspended in the air and
mastered the art and science of navi- trol systems are moving drones be- constantly moving introduces addi-
gating and accomplishing their des- yond the basic ability to fly from tional challenges and risks.
IMAGE BY RIKOBEST
ignated task. Buildings, birds, power Point A to Point B. They’re making it Today, most UAVs operate on a
lines, trees and people remain formi- possible for drones to understand line-of-sight basis. Essentially, a per-
dable obstacles for autonomous Un- the world around them and make son uses a transmitter, typically oper-
manned Aerial Vehicles (UAVs), as they complex decisions in real time.” ating in the 2.4GHz frequency band,
to communicate with and control the and latency problem. Moreover, bet-
drone’s onboard computer. However, ter algorithms cannot anticipate every
for drones to become truly autono- What comes possible scenario or obstacle the world
mous, operate at high speeds, and naturally to pilots can toss at a drone. Ultimately, a UAV
ultimately become a commercially must be able to react to external events
viable tool, onboard systems need to when watching and avoid collisions while staying on
operate independently of humans (at a UAV video stream course and accomplishing its intended
least the vast majority of the time). task. Says Amoroso, “While a drone
This requires a dozen or more on- is considerably more will likely not have the understanding
board sensors, such as cameras that difficult for today’s a pilot has of the behavior of everyday
work in both the visible and infrared objects, it nonetheless must react ap-
spectra, LIDAR (light detection and smartest UAVs. propriately and quickly to avoid situa-
ranging), or multi-spectral cameras; tions where it can cause harm to oth-
more advanced algorithms for un- ers or itself. Maybe the drone doesn’t
derstanding a wide range of environ- understand that branches can fall, or
mental conditions; and sophisticated (medium-wave infrared), LWIR (long- doors can open, but if given a robust
navigation systems that allow UAVs to wave infrared), LIDAR (light detection enough SLAM system, it will still be
sense their position more precisely. and ranging), and radar (radio detec- able to navigate itself safely under such
There is also a need for improved tion and ranging)—to robustly “see” environmental disturbances.”
safety systems—particularly in crowd- what is going on around the drone.
ed urban areas. “Currently, drone What comes naturally to pilots Gaining Direction
companies add redundant propellers when watching a UAV video stream— Although GPS technology allows most
to avoid crashing. More advanced tech- depth of field and localization of both drones to operate effectively most of
nology is necessary,” says Davide Scar- static and dynamic objects—is consid- the time, a dependence on satellites is
amuzza, director of the Robotics and erably more difficult for today’s smart- not ideal—or even adequate—for com-
Perception Group at the University of est UAVs. Consequently, researchers panies looking to use UAVs for special-
Zurich in Switzerland. are continuing to experiment with dif- ized commercial purposes. Objects
At drone manufacturing firms and ferent combinations of sensors and such as buildings, trees or mountains
in research labs, the next generation SLAM algorithms to guarantee sight might temporarily block signals. GPS
of drone controls and navigation in cluttered environments. This in- also doesn’t deliver the level of per-
systems is taking shape. Engineers cludes stereoscopic vision and associ- formance and precision needed when
and computer scientists are taking ated algorithms that help a drone gain many drones operate autonomously
aim at various challenges, including depth-of-field and better understand close together. Without additional vi-
how to process visual information at relationships between and among ob- sion sensors and on-board navigation
speeds reaching near 100 mph (160 jects—including other moving drones. systems, collisions could occur, or
kph), how to teach UAVs to react to Equipping drones with vision and drones might simply cease doing what
unknown obstacles, what to do if the sensing capabilities that operate at the they are supposed to do.
drone does not know how to respond speed of flight is only part of the navi- More advanced UAVs now incorpo-
to a given situation, and how to take gation challenge, however. There is rate a technology called Visual Inertial
over the controls for malfunctioning, also a need to ensure that a drone can Navigation System (VINS) assistance.
rogue, or dangerous drones that may process visual images quickly enough These systems rely on onboard cam-
pose a threat. Not surprisingly, many and make intelligent decisions in real eras and inertial measurement units
of the decisions involve trade-offs. time. Microprocessor and component (IMUs) to track a drone’s location when
For example, it is already possible to manufacturers have introduced highly GPS signals are weak or nonexistent,
fly an autonomous drone that has a specialized chips that use increasingly such as in caves or deep valleys. Essen-
very low probability of colliding with powerful graphics processing units tially, they work by detecting and track-
objects or crashing—as long as it flies (GPUs) and accelerator chips to reduce ing interest points across images and
at a very slow speed. visual processing time to milliseconds. using them as anchor points for the
At the center of the challenge is si- Yet, further improvements are needed. robot to orientate itself, Scaramuzza
multaneous location and mapping For now, pilots can detect operational says. In a certain sense, it’s the drones
(SLAM). Eric Amoroso, cofounder of anomalies and react more quickly than mapping territory and using the map
KEF Robotics, a drone company that an autonomous UAV. The ultimate ob- as they move over land, within caves,
captured first place in a qualification jective for drone manufacturers is to or underwater. However, this, too, has
round for a 2019 Lockheed-Martin UAV push the devices’ reaction time to the limitations since some environments
challenge, says inaccuracies in sens- level of professional pilots so they can change quickly.
ing and processing algorithms neces- perform on par with humans, or per- Completely autonomous drones
sitates multiple onboard systems—as haps even exceed them. would require a combination of sen-
many as a dozen conventional cameras, Machine learning will certainly sors, navigational capabilities, and
vision sensors using such technologies make UAVs smarter and more agile, but communications links that push be-
as SWIR (short-wave infrared), MWIR it cannot completely solve the speed yond current technology. They may
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 17
news
Real-World Applications
for Drones
Unmanned vehicles have a number
of compelling real-world use cases.
I
announced it
N JU N E , A M A Z ON
was close to being able to offer
for package deliveries by drone
for its Prime Air service. That
same month, Uber said it plans
to test food delivery by aerial drone
in crowded cities. And drone delivery
company Flytrex already touts the abil-
ity to deliver drinks via unmanned ve-
hicle on the golf course.
Despite such announcements,
drones are not crowding the skies over
major cities and population centers just
yet. But that may be about to change.
After several years of hype, wide-
spread drone usage may be close to
ready for primetime.
Drones increasingly are being de-
ployed in a number of compelling
real-world use cases. These use cases
have drone companies and enthusiasts
bullish that, no matter what happens,
there are serious real-world applica-
tions for drone technology today and
in the near future that will disrupt life
and business as we know it.
Drone-Assisted Aerial footage shot by drones is less insurance companies an easy way to
Photography/Surveying expensive than manually taking aerial assess damage related to claims. For
“Traditionally, we’ve seen drones be- footage from a helicopter. One drone insurance and surveying purposes,
ing used for photography and survey- photographer interviewed by The Bal- aerial drones offer the ability to cover
ing,” says Eric Peck, CEO of Swoop timore Sun noted the cost differences: more ground while traversing more
Aero, an Australian company that “I can drive up to my destination, plug areas and angles than might be pos-
delivers medical supplies via aerial my equipment in, and be done [pho- sible (or affordable) with traditional
drone. “It’s all about data capture, tographing] in five or 10 minutes,” manned aircraft.
because data really is driving the abil- said Jack Hardway, owner of a drone Other use cases include surveying
ity to generate economic growth at the photography firm. “It doesn’t cost me and monitoring progress at construc-
moment.” $5,000. It costs me pennies to put that tion sites, and performing simple regu-
From construction to insurance to thing in the air.” latory inspections for commercial real
real estate to agriculture, the ability to The cost is one benefit. The ability estate properties. Aerial drones are
survey and photograph wide swaths of to collect more visual data from more even used to fly around warehouses
land and hard-to-reach locations with angles than from a traditional camera and find supplies or products faster
IMAGE F RO M SH UTT ERSTOCK.CO M
aerial drones is valuable to companies. also is important. and more accurately than humans do.
For instance, high-quality photos and A Santa Monica, CA-based compa- Aerial drones also come in handy
videos from different aerial angles can ny called DroneBase uses unmanned in agricultural applications. They of-
better showcase residential properties aerial vehicles (UAVs, or aerial drones) fer a dual benefit in this context. First,
up for sale, more effectively highlight- to offer, among other services, aerial drones are used to survey fields. In-
ing elements that appeal to buyers. surveying of building rooftops to give stead of having to traverse hundreds
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 19
news
or thousands of acres on foot or by “While the market for drone pho- in countries like Pacific island-nation
vehicle, farmers have the ability to tography and data capture is massive, Vanuatu, which is composed of doz-
fly drones faster and more efficiently it’s close to saturated right now, both ens of islands. Often, the deliveries
over large areas. That helps reduce in terms of platforms and operators,” take a fraction of the time they would
the time it takes to monitor fields, as says Peck of Swoop Aero. “We’ll see a if conducted by boat. Vanuatu is a
well as reducing the amount of fertil- lot of movement as big players gain ad- country in which, the United Nations
izer and pesticides they must use to vanced regulatory approvals [for more Children’s Fund (UNICEF) estimates,
maintain crops. extensive drone applications], which a full 20% of children under five do
“We identify diseases and pests allow them to gain a cost advantage not receive all the vaccines they need
and fungus and weeds in the crop at based on economies of scale and drive because of the logistical challenges
an earlier stage,” U.K. farmer Colin smaller operators out of the market.” around medicine delivery.
Rayner told German broadcaster DW. That dynamic has led companies Last December, Joy Nowai of Vanu-
Some drones are even used to spray like Swoop Aero to look at use cases atu was, according to the company,
fields with pesticides. According to for drones that involve delivering high- the “first child in the world to be
DW, Chinese drone company DJI sold value commodities. vaccinated with a vaccine delivered by
20,000 pesticide-spraying drones in a drone under a commercial contract,”
2018 alone. Drone Doctors thanks to Swoop Aero. The drone trav-
In all of these examples, the drones One high-value commodity that makes elled 30-plus miles to deliver the vac-
are being piloted remotely by experi- a lot of economic sense to deliver via cine, while keeping it at the optimum
enced professionals. For instance, drone is medicine. temperature during the entire trip.
DroneBase claims it has the “largest The market need is clear: when it After the successful delivery, the
network of professional drone pilots comes to perishable medical samples company conducted a further four-
in the world,” but they are all still or life-saving vaccines, time is of the es- month trial in Vanuatu, which Swoop
human beings. sence, and few technologies are better Aero says led to continued work with
Right now, this gives an advantage at traversing crowded or hard-to-reach the country’s Ministry of Health.
to bigger companies that can scale and areas than aerial drones. Swoop Aero is now preparing to
capture cost advantages that offset the Swoop Aero operates drone net- deploy additional drone networks to
expense of human pilots. works that deliver medicine quickly countries that lack easy logistical ac-
ACM News
cess to life-saving medicines and vac- to real-world obstacles with a high de-
cines. One of the company’s networks gree of accuracy and speed.
is being deployed in the Democratic Companies That could take some time, Keith
Republic of the Congo, in collabora- are desperately Lynn, a program manager for Lock-
tion with that country’s Ministry of heed Martin’s autonomous drone
Health and local non-governmental trying to make racing competition, told The New
organizations. it possible for York Times. “Right now, autonomous
Another company, California-based drones are a thing you’d only find
Zipline, just launched a series of drone consumers to receive in labs, being pioneered by a small,
distribution centers in Ghana designed deliveries by drone niche audience.”
to deliver vaccines and medications A big reason for that is because au-
to the country’s population 24 hours a of products that tonomous drones struggle to make
day. The company says health workers would normally be sense of visual information, particu-
placing orders for medications via text larly at high speeds, in part because
can expect the requested medicine to delivered by mail of shortcomings in the sensors they
be delivered within 30 minutes. or manned vehicle. utilize. Also, the faster the drones fly,
This type of aerial delivery at scale is dive, or drive, the more difficult it is for
not only well suited for geographically today’s algorithms and cameras to pro-
inaccessible areas. Earlier this year, cess images at the speed required to
logistics company UPS, in partnership recognize (and avoid) obstacles.
with California-based drone startup Aero Swoop’s Peck is optimistic ad-
Matternet, launched a pilot program go deeper underwater and stay under vancements will lead to drones having
to deliver medical samples around the longer than human-manned vessels greater autonomy in the near future.
campus of WakeMed, a not-for-profit could, transmitting back valuable data “Over the next five to 10 years, we are
health care system in Raleigh, NC. all the while. going to see increasing levels of full
According to Business Insider, such The unmanned underwater drone autonomy used across all aspects of
samples used to take up to 30 minutes approach has been so effective that aviation,” he predicts.
to deliver due to traffic congestion, one company tasked with finding
but using drones, the deliveries now Malaysia Airlines Flight MH370, the
Further Reading
take just over three minutes. Dutch geosciences company called
“Transport is a clear opportunity Fugro NV, reportedly “plans to do away Hamilton, I.
Amazon drone deliveries are coming, but
commercially,” says Peck. “We are fo- with some [human] crews entirely.”
Jeff Bezos still missed his own deadline
cused on last-mile logistics for high- for airborne logistics, Business Insider,
value commodities, predominantly The Future of Drone Delivery Jun. 6, 2019, http://bit.ly/2kJWNcv
in healthcare. It’s a market which is There are plenty of intriguing real-
forecast to grow in size from close to world applications for drones at pres- Holley, P.
Uber plans to start delivering fast food via
zero right now, to be worth over $10 ent, but that has not stopped compa- drone this summer, The Washington Post,
billion in the next seven years.” nies from salivating over the holy grail Jun. 13, 2019, https://wapo.st/2m46EK9
Drones are not just flying to areas of drones: last-mile logistics. Compa-
Thompson, F.
where people find it hard or time- nies such as Amazon are desperately Next generation farming: How drones are
consuming to go; they are swimming trying to make it possible for consum- changing the face of British agriculture, DW,
there, too. ers to receive deliveries by drone of Jul. 19, 2019, http://bit.ly/2kviDAk
In 2018, submersible drones built products that would normally be de- Waseem, F.
by Texas-based Ocean Infinity worked livered by mail or manned vehicle. Howard drone users search for opportunity
together to survey parts of the ocean While some companies focus on as ‘the skies open’, The Baltimore Sun,
delivering high-value goods via un- July 7, 2016, http://bit.ly/2mit76q
floor inaccessible to humans. The
goal was to find the remains of ships manned or semiautonomous drones, Wilke, J.
that had gone missing. The company last-mile logistics at scale requires A drone program taking flight, Amazon,
Jun. 5, 2019, https://blog.aboutamazon.
succeeded in discovering the wrecks near-full autonomy. To deliver most com/transportation/a-drone-program-
of an Argentinian submarine and a or all products at scale, drones from taking-flight
South Korean commercial vessel, long a company like Amazon will need
Wise, J.
after hope was lost that concerned have the ability to fly themselves short Underwater Drones Nearly Triple Data
parties would learn the fate of the dis- distances with little to no human in- From the Ocean Floor, Bloomberg
appeared craft. volvement. Businessweek, Jun. 7, 2019,
Like their aerial counterparts, un- That means building highly so- https://bloom.bg/2kobrpq
derwater drones are packed with sen- phisticated, reliable types of artificial
Logan Kugler is a freelance technology writer based
sors that collect data and share it with intelligence (AI) and machine learn- in Tampa, FL, USA. He has written for over 60 major
company control centers. Instead of ing into delivery drones. These AI- publications.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 21
ACM ON A MISSION TO SOLVE TOMORROW.
Dear Colleague,
Without computing professionals like you, the world might not know the modern
operating system, digital cryptography, or smartphone technology to name an obvious few.
For over 70 years, ACM has helped computing professionals be their most creative, connect
to peers, and see what’s next, and inspired them to advance the profession and make a
positive impact.
ACM offers the resources, access and tools to invent the future. No one has a larger
global network of professional peers. No one has more exclusive content. No one
presents more forward-looking events. Or confers more prestigious awards. Or provides
a more comprehensive learning center.
Here are just some of the ways ACM Membership will support your professional growth
and keep you informed of emerging trends and technologies:
Joining ACM means you dare to be the best computing professional you can be. It means
you believe in advancing the computing profession as a force for good. And it means
joining your peers in your commitment to solving tomorrow’s challenges.
Sincerely,
Cherri M. Pancake
President
Association for Computing Machinery
q Join ACM-W: ACM-W supports, celebrates, and advocates internationally for the full engagement of women
in computing. Membership in ACM-W is open to all ACM members and is free of charge.
PAYMENT INFORMATION
Name
Purposes of ACM
ACM is dedicated to:
Mailing Address 1) Advancing the art, science, engineering, and application
of information technology
2) Fostering the open interchange of information to serve
both professionals and the public
City/State/Province
3) Promoting the highest professional and ethics standards
ZIP/Postal Code/Country
By joining ACM, I agree to abide by ACM’s Code of Ethics
q Please do not release my postal address to third parties (www.acm.org/code-of-ethics) and ACM’s Policy Against
Harassment (www.acm.org/about-acm/policy-against-
harassment).
Email Address
I acknowledge ACM’s Policy Against Harassment and agree
q Yes, please send me ACM Announcements via email that behavior such as the following will constitute
q No, please do not send me ACM Announcements via email grounds for actions against me:
Legally Speaking
Europe’s Controversial
Digital Copyright
Directive Finalized
Considering the new liability risks for ISPs, search engines,
and news aggregators under recent EU-wide mandatory rules.
I
NTERNET G OV E R N A N C E RULES tions on some aspirations that underlie for online content-sharing platforms,
in the EU are about to change this Directive. such as YouTube and Facebook.
radically. The final version of Under laws in place in the EU and
its Directive on Copyright and Repeal of the Safe Harbor for U.S. since 1998, Internet service pro-
Related Rights in the Digital ISP Storage of User Contents viders (ISPs) have enjoyed a safe harbor
Single Market (DSM), which has been The most significant and controversial from liability for infringing acts of their
under consideration for the past three of the new DSM rules is the stiffer li- users of which the ISPs were unaware.
years, was promulgated on April 17, ability rules the Directive established ISPs faced liability only if they failed to
2019. EU member states now have two investigate and take down infringing
years to transpose the Directive’s rules materials after receiving notice from
into their national laws. There is still reason copyright owners about where such
In some respects, the DSM Direc- materials were located.
tive is better than previous drafts to worry the new Article 17 of the DSM Directive
(of which more anon). There is still rules will be harmful (Article 13 under previous drafts) im-
reason to worry the new rules will be poses strict liability on online content-
harmful for freedom of expression for freedom sharing sites for user infringements
and information privacy interests of of expression and obliges them to use “best efforts
individual creators and users. How to ensure the unavailability of specific
much harm will depend on how mem- and information works.” Because EU member states
ber states implement the Directive privacy requests. may decide that “best efforts” requires
and how courts interpret it, as many platforms to use filtering technologies,
of its terms are ambiguous. this provision has often been called the
This column discusses key differ- “upload filter” rule. (Previous drafts of
ences between earlier drafts of the the Directive were more pointed about
DSM Directive and the final version the need to use filtering technologies.)
and makes some general observa- There are two exceptions to the DSM
Directive’s strict liability rules for on- exceptions for “(a) quotation, criti- in Article 3 that member states adopt a
line content-sharing platforms. One is cism, review; (b) use for the purpose of mandatory new exception to copyright
for nonprofit services, such as online caricature, parody, or pastiche.” and database rules to allow nonprofit
encyclopedias, educational and scien- Whether this effort to ensure privi- research and cultural heritage institu-
tific repositories, and open source soft- leged uses can be uploaded to content- tions to engage in text and data mining
ware developing platforms. A second sharing sites will meaningfully limit the for scientific research purposes.
is for startup online content-sharing Directive’s scope or serve only as aspira- While this exception was good so far
services that have been available to the tional window dressing remains to be as it went, earlier versions of the DSM
EU public for less than three years and seen. It seems unlikely, though, that EU Directive would have left independent
that have annual revenues of 10 mil- member states can require developers researchers and profit-making text and
COMPOSITION BY A ND RIJ BORYS ASSO CIAT ES, USING IMAGE BY OLIVIER LE M OAL
lion euros or less. The liability of these of filtering technologies to refine their data miners out in the cold. Because
two types of services for user infringe- algorithms so that all parodies, critical EU policymakers aspire to foster the
ments are subject to compliance with comments, and other privileged uses growth of artificial intelligence and
the existing notice and takedown rules. will remain available to the public. Yet, other data-intensive businesses, they
Critics have charged that the DSM this seems to be the only way to ensure came to recognize restricting text and
strict liability rules will interfere with privileged uses can be preserved. data mining to nonprofit scientific re-
user freedoms to make lawful uses of search was shortsighted, especially
copyrighted works, such as parodies or Text and Data Mining Exceptions given that other countries, notably the
critical commentaries, because filter- Under existing EU law, text and data U.S. and Japan, have adopted broader
ing technologies cannot distinguish mining on digital repositories of copy- text and data mining privileges.
between outright infringements and righted works and databases had an un- While Article 3 retains the text and
privileged uses. certain status. The drafters of the DSM data mining exemption for nonprofit
Seemingly in response to this criti- Directive decided this activity should scientific research, the final DSM Di-
cism, Article 17 now states that mem- be lawful because of the important in- rective sets forth a new Article 4 requir-
ber states “shall ensure that users in sights the use of such research tools ing member states to create a more
each Member State are able to rely on can enable. To this end, they proposed general mandatory exception to copy-
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 25
viewpoints
right and database rules to allow text Council and Parliament decided to ap-
and data mining by independent re- prove the grant of this new set of exclu-
searchers and profit-making establish- sive rights.
ments without restriction on purpose. Seemingly to counter the charge
Although Article 4 is broader than that Article 15 would create a “link tax,”
Article 3 in the users and uses to which Article 15(1) explicitly provides that the
it would apply, Article 4 is more limited press publisher right does not apply
than Article 3 in two respects: First, to hyperlinking. In an effort to further
the Article 4 exception does not apply narrow its reach, Article 15(1) says it
to the extent that rights holders have would not apply to “private or non-
expressly reserved the right to control commercial uses of press publications
text and data mining. Second, the Ar- by individual users.” Nor would it apply
ticle 4 exception can be overridden by to use of individual words or very short
contract, whereas the Article 3 excep- extracts of a press publication.”
tion is nonwaivable by contract. But what exactly constitutes a “very
short extract” of a press publication is
UC MERCED
Press Publishers Right unclear. Ambiguities about this and
The final version of the DSM Directive other terms in Article 15 makes it un-
directs member states in Article 15 likely that member states of the EU will
(previously Article 11) to grant press implement this new right in a harmo-
publishers two years of exclusive rights nious way.
to control reproductions and commu-
nications to the public by information Licensing as a Goal
society service providers. of the DSM Directive
Earlier versions of the DSM Direc- Proponents of the DSM Directive told
tive’s press publisher right attracted in- European policymakers a powerful
tense criticism. Opponents charged it story in support of the new liability
would impede the free flow of news and and exclusive rights rules that the Di-
other information vital to a democratic rective has now established. They as-
society, harm journalists who often rely sert there is a “value gap” the Direc-
on search engines and aggregators, and tive could correct.
create uncertainty about its coverage The short version of that story is that
and scope. Critics also thought this new U.S. technology companies are making
right was unnecessary, unlikely to pro- huge revenues from their uses of Euro-
duce significant licensing revenues, and pean rights holders’ contents and too
likely to further entrench powerful me- little of these revenues are flowing to
dia conglomerates and global platforms European content providers. (Both Eu-
to the detriment of smaller players. ropean and American commentators
Critics also expressed concern have expressed considerable skepti-
about how the new publisher right cism about the “value gap” story, but it
would interact with existing copyright was an influential part of the rationale
laws, which typically allow for fair quo- for adopting Articles 15 and 17.)
tation rights, as well as with database To narrow, if not close, this gap, the
rights, which allow users to extract in- DSM Directive aims to induce tech-
substantial parts of databases. nology companies to negotiate for li-
Notwithstanding serious concerns censes. If such licensing occurs, then
about the press publisher right, the EU the stricter rules will not need to be ap-
plied, and worries about harms to free-
dom of expression and other social val-
Earlier versions ues expressed by critics of the stricter
liability rules will not come to pass.
of the DSM Directive’s Consider, for instance, Article 17(1).
press publisher After providing that online content-
sharing sites will be strictly liable for
right attracted giving the public access to infringing
intense criticism. copyright-protected contents upload-
ed by users, that provision goes on to
say that to avoid this liability, such sites
“shall obtain an authorization” from
rights holders “by concluding a licens-
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 27
V
viewpoints
DOI:10.1145/3363181 Mark Guzdial, Alan Kay, Cathie Norris, and Elliot Soloway
Education
Computational
Thinking Should
Just be Good Thinking
Seeking to change computing teaching to improve computer science.
J
EA N N E T T E WIN G ’ S 2006 Com-
munications Viewpoint on
computational thinking5 ig-
nited a worldwide movement
to give students new knowl-
edge and skills to solve problems in
their daily lives. Quickly, teachers, cur-
riculum and standards writers, and
other education specialists were pro-
posing what children needed to know
about computation and how to devel-
op a computational mindset. There is
still little evidence that knowing about
computation improves everyday prob-
lem-solving, but there is no doubt that
Wing’s call to action led to a broad and
dramatic response.
The computational thinking move-
ment puts the onus on the student and
on the education system. They argue
that if we change humans to think in
ways that are informed by how we now
work with computers, that will have prob-
lem-solving advantages for the humans.
Maybe. back on the computer scientists and the tools of carpentry, the way a car-
If a city does not work for the resi- other computationalists. It is our job to penter thinks.
dents, we could change the residents. design better. Closer to home, the “kids these
Alternatively, we could redesign the city. days” use all manner of digital—read:
The best urban redesign has citizens For Our Children, Computational computational—tools. Before drawing
understanding the purpose and actively Thinking Is Just Thinking the obvious conclusion, consider the
participating, so there is parallel devel- Tool use shapes thinking. While we following vignettes.
opment of both the city and the citizens. might not think like a carpenter when Vignette 1: Consider the following
Children today already think with we start using carpentry tools, if we two problems, drawn from a research
computation. If we want better think- apply ourselves (for example, reflect study1:
IMAGE BY OLLYY
ing and problem-solving, we have to on our doing, as Dewey suggests2), we ˲˲ (Algebraic Context): Given the fol-
improve the computing and use that to can develop carpentry thinking. We lowing statement: “There are six times
change our teaching. We put the onus can learn to see what is possible with as many students as professors at this
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 29
viewpoints
manuscripts must are taught to all to provide a “richness” Journal for Research in Mathematics Education 13, 1
(Jan. 1982), 16–30; doi:10.2307/748434
of thought about both civilizations
demonstrate scientific and how to be a citizen who supports
2. Dewey, J. How We Think. A Restatement of the
Relation of Reflective Thinking to the Educative Process
rigor and present results civilization. Understanding civiliza-
(Revised edition); D.C. Heath, Boston, MA, 1933.
3. Montessori, M. The Montessori Method. Frederick A.
that are reproducible. tion as a system is a powerful idea for Stokes Company, New York, 1912.
4. Norris, C. and Soloway, E. Students write more, write
all citizens. In our metaphor, we want better on the computer: Rigorously supported! T.H.E.
Journal, (Nov. 11, 2017); https://bit.ly/2KQToTg
citizens to participate in the redesign 5. Wing, J.M. Computational thinking. Commun.
of the city and understand the rationale ACM 49, 3 (Mar. 2006), 33–35; DOI: https://doi.
org/10.1145/1118178.1118215
for its design. Students need fluency in
order to be able to understand models
Mark Guzdial (mjguz@umich.edu) is Professor of
and systems. Important thresholds of Electrical Engineering and Computer Science, College
understanding must be reached before of Engineering and Professor of Information, School of
Information, University of Michigan, Ann Arbor, MI, USA.
they can be part of one’s thinking tools.
Alan Kay (alan.viewpoints@yahoo.com) is Adjunct
Finding and inventing these thresholds Professor, Computer Science, University of California, Los
for the general population of children, Angeles, USA. He is the recipient of the 2003 ACM A.M.
Turing Award.
and how to teach to them, is the critical
Cathie Norris (cathie.norris@unt.edu) is Regents
need of our time! Professor, Learning Technologies, College of Information,
Representations to help thinking— University of North Texas, Denton, TX, USA.
Interview
An Interview with
Leonard Kleinrock
The UCLA professor and networking pioneer
reflects on his career in industry and academia.
L
EONARD KLEINROCK, DEVELOP-
ER of the mathematical theory
behind packet switching, has
the unique distinction of hav-
ing supervised the transmis-
sion of the first message between two
computers. As a doctoral student at
MIT in the early 1960s, Kleinrock ex-
tended the mathematical discipline of
queuing theory to networks, providing
a mathematical description of packet
switching, in which a data stream is
packetized by breaking it into a se-
quence of fixed-length segments (pack-
ets). ACM Fellow Kleinrock has received
many awards for his work, including
the National Medal of Science, the
highest honor for achievement in sci-
ence bestowed by a U.S. president.
UCLA Professor and ACM Fellow
George Varghese conducted a wide-
ranging interview of Kleinrock, an ed-
PHOTO BY DAN A NDERSO N, COURTESY OF ELON UNIVERSIT Y/F LICK R (CC BY- ND 2 .0)
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 31
viewpoints
industrial electronics. So instead of Ken Olsen, who later went on to planned to come in the final summer
attending CCNY (City College of New found Digital Equipment and build just before I finished my MS, at which
York) as a daytime student, I attended the line of PDP computers. I worked point I would be working full time at
at night. My day work was, however, in a group at Lincoln run by Wes MIT Lincoln Lab and we needed the
wonderfully interesting: we were in- Clark who built arguably one of the money since my wife would have to
volved in designing and using photo- first PCs (the Linc computer). So care for our newborn. So, I was not at
electric devices in many applications. there were a lot of brilliant people at all interested in pursuing a Ph.D. But
The people in night school were an Lincoln; and of course MIT profes- Frank Reintjes was insistent and,
interesting bunch—after all, who at- sors would often visit. amazingly, Lincoln Labs decided to
tends night school: crazies, dropouts, offer me a follow-up Ph.D. fellowship
motivated students who had to work What did you do your MS thesis on? to MIT just as they had done for my MS
during the day, and GIs coming back When I first got to MIT, I was inter- fellowship; this was a first for Lincoln.
from World War II (this was 1951) ested in servomechanism systems So I succumbed to the pressure and
who were disciplined and very deter- and automatic control. Yet, my mas- accepted the Ph.D. fellowship. Two
mined. The professors at night school ter’s thesis at MIT was on optical others were also offered the Ph.D. fel-
worked in industry during the day so readout of thin magnetic films for lowship: Larry Roberts [one of the
they had insight into practical mat- storage and processing. I made use of founders of the Internet, see later—
ters. I remember a professor bring- the Kerr magneto-optic effect where- Ed.] and Ivan Sutherland [one of the
ing a germanium transistor he worked by polarized light rotates differently founders of graphics and an ACM
on during the day to class saying “this when it reflects off a magnetized sur- A.M. Turing Award recipient—Ed.]
is a better thermometer than an am- face depending on the direction of who both became lifelong friends.
plifier,” and began to discuss ways to magnetization. As a result, one could
eliminate the temperature-dependent use polarized light to non-destruc- What were the first years of your MIT
variations. This combination of com- tively “read” the bits on thin magnet- Ph.D. experience like?
bining practical issues with mathe- ic films (this was before disks). My Our Ph.D. qualifier was legendary
matical approaches has always sup- job was to improve the reading pro- for its difficulty with 50% of the appli-
ported my seeking to find intuition cess by amplification and coding. cants failing out like flies. My MS at
and insight behind theory. Claude The thesis involved experiments and MIT made it easier since the qualify-
Shannon, who was then—and still is— models and I even constructed a spe- ing exam was largely based on the
my role model, similarly had great in- cial digital logic using light bouncing MIT MS curriculum, but full of trick
sight and physical intuition into why off a sequence of thin films. My thesis questions. Interestingly Ivan (Suther-
things happened alongside his mathe- must have impressed my MS supervi- land) came in directly from Caltech
matical approach to problems. sor—Frank Reintjes—at MIT because (that is, without the benefit of expo-
he insisted that I apply for a Ph.D. sure to the MIT MS material directly)
You probably were thinking of getting and came out on top with one month
a job after CCNY. How did you go to But the idea was that after a Lincoln to study; he is one heck of a smart guy.
MIT instead? Labs fellowship you should work at Lin- When I agreed to continue on with a
I learned one day that an MIT pro- coln Labs as an engineer, right? And Ph.D. program, I decided I wanted to
fessor was coming to CCNY at 4 P.M. you had a first child coming by then? work with the best professor I knew at
to describe a terrific fellowship that That’s right. Our first child was MIT, and so called up Claude Shan-
would provide considerable financial non (founder of information theory).
support to pursue a master’s at MIT He surprised me (and shocked my
as an MIT Lincoln Labs [a well-known Being surrounded friends) by inviting me to his house in
R&D laboratory associated with MIT— Winchester, MA, USA. I remember the
Ed.] staff associate. I managed to get by computers at MIT scene looking out on Mystic Lake as
off work early that day, but when I and at Lincoln Lab, an automatic lawn mower (rigged up
asked the MIT professor for an appli- by Shannon) mowed the grass and his
cation to the program, he told me they it seemed inevitable son’s swinging hammock narrowly
were available from a CCNY professor to me they would missed my head. Shannon wanted me
sitting at the back. The CCNY profes- to work on a strategy for the middle
sor did not recognize me and when I eventually need game in chess as part of a project that
told him I went to night school he to communicate he and [AI Founder and Turing Award
said “get out of here.” So I had to con- winner—Ed.] John McCarthy were
tact MIT directly to get a form. That I with each other. working on.
did and I was fortunate to be awarded
the fellowship! How did you gravitate to what is con-
sidered your seminal thesis on packet
What was it like doing a master’s at communication?
MIT as a Lincoln Labs associate? I was looking for a fresh field to work
My first supervisor at Lincoln was on. It seemed to me that even Shannon
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 33
viewpoints
partial network damage. His work albeit less optimal, protocol where the
was classified and so I did not see his dynamics of the network were reflect-
papers until my thesis was complet- While we think of the ed in which links of interest out of a
ed. His work was right on! Another Internet today to send node were idle, indicating that the
important thread was from Donald route to the destination using a
Davies who was working indepen- email and support queued link was not desirable and
dently at the National Physical Labo- social networks, that the uncongested links leading
ratory in England and who realized out of a node to that destination were
that the packet switching was good the motivation then currently better choices.
because, as we had articulated, data was to share the
was bursty. He coined the word “pack- Your thesis defense was a remarkable
et” and pointed out that long packets expensive computers event ...
were more likely to contain an error ARPA was funding. Larry Roberts, Ivan Sutherland,
than were small packets; hence he and I were very close friends and did
suggested packets of approximately our thesis defenses at the same time
128 bytes, which was later used in the because we had all heavily used the
ARPANET design. He promoted pack- MIT Lincoln Lab TX2 computer as
etization as having the desirable ad- part of our Ph.D. research. The union
vantage of allowing small messages to of our three Ph.D. committees came
swoop past large messages; interest- out to Lincoln Lab to view our work
ingly, I had shown the exact form of Davies. These include techniques including Claude Shannon, who was
this trade-off mathematically in my such as priority queueing (for example on each of our committees, Marvin
dissertation years earlier. Morover, he VoIP is queued before data packets in Minsky, and Peter Elias. It was a big
recognized that once messages were today’s routers) and splitting packets heyday and a bit stressful given the
packetized, then retransmission of for the same destination across multi- credentials of those committee
packets rather than whole messages ple paths (called ECMP). Is there any- members. The projects were very
would reduce delays in overall trans- thing else you want to mention before different; we were just all using the
mission. Further, he noted that the we move on from your thesis? TX2. Ivan did this great work on
ability to pipeline packets reduced la- Some other aspects include the ef- Sketchpad, Larry did his on machine
tency through the network. fects of scaling. I showed for the first perception of three-dimensional ob-
time that in terms of performance, a jects, and I did mine on communica-
So in some ways, Paul, Donald, and single link of capacity C is better than tion networks. The TX2 allowed me
you explored different facets of the N links each of capacity C/N (this was to run an enormous simulation to
benefits of packet switching. Paul fo- an example of the second principle I verify the accuracy of my mathemati-
cused on routing resiliency, Donald on mentioned previously). I investigated cal approximations.
packet-level error resiliency, and you how to optimally design network to-
on mathematical performance evalua- pologies, which contributed to the You submitted your thesis and clearly
tion and optimization of packet- field of network flows that Howie it was well liked at MIT since they sug-
switching networks using stochastic Frank and Ivan Frisch and others gested you publish it as a book. How
models. Is that accurate? made major contributions to. I also did you end up at UCLA and not at Lin-
That is a fair characterization. I investigated distributed adaptive coln Labs?
would add that Paul and Donald were routing control but I modeled that by Morally, after their fellowship I
looking mainly at critical architectural having each router precompute an or- felt I should work for Lincoln. But
issues whereas I was more focused on dered sequence of favorable routes they were remarkably generous and
extracting the underlying principles for each destination and use the first offered to have me look around the
and developing a mathematical theory route that was not congested locally. academic and industrial circuit to
of packet networks. Among the princi- see what opportunities were there. I
ples were: dynamic resource sharing is Interesting! That’s different from dy- received some great offers of re-
key in an environment of bursty de- namic routing in today’s Internet search positions: Bell Labs, Lincoln
mands; large shared resources sup- where routers use Dijkstra’s algorithm Labs, Hughes, and many more. And
porting lots of traffic are far more effi- to compute shortest paths. However, then there were academic offers, in-
cient than small resources supporting that technique takes longer to respond cluding the one from UCLA for a ten-
less traffic; and distributed adaptive to failure. Some networks today (for ex- ure-track position (at half the salary
control is efficient, stable, robust, ample, MPLS protection) use your idea I would get at Lincoln). But I had
fault-tolerant, and it works. for faster recomputation of (possibly loved the few courses I taught while
less optimal) routes after failure. at MIT and realized I could augment
Your thesis also anticipates and ana- It is true that I did not suggest a Di- my salary by consulting. So with the
lyzes other benefits of packet switch- jkstra-type updating procedure dy- West Coast weather, the Wild West
ing we rely on today that are comple- namically based on networkwide appeal, and a university position, I
mentary to those pointed out by shortest paths. I introduced a simpler, drove my family all the way across
the country. Lincoln Labs was ex- In 1967, Larry brought a bunch of us just one of many networks in an evolv-
tremely gracious and even said that I together to help him specify what this ing “internetwork of networks” where
could come back if I did not like it at network would look like and what per- every network would have a network
UCLA—but it has been 56 years and I formance characteristics it would have. number. The need for a more advanced
am still here! We specified the network and created internetworking protocol became
the spec and then Larry put it out for clear and this was Cerf and Kahn’s
Fast-forward to the birth of the Inter- bid. In December 1968, BBN was grant- great achievement of TCP/IP for which
net in a UCLA office. On October 29, ed the contract. they justly were given the ACM A.M.
1969, you and Charley Kline, one of the Turing Award.
Ph.D. students on your software team, In September 1969, BBN delivered the The rest of the story, the commis-
transmitted the first message between first IMP to you at UCLA. Why UCLA sioning of the NSF backbone, the deci-
computers hundreds of miles apart. and not SRI or Utah? sion to transition to multiple com-
What was the backstory? My role in this ARPANET project mercial backbones who had to
In my software team, besides Char- was performance evaluation, design, cooperate, and so forth, are all well
ley there was Steve Crocker who head- experimentation, and measurement. known. We had no clear idea of how
ed the software group, Vint Cerf, and At UCLA we had specified the meas- the Internet would be used, but we
Jon Postel, all UCLA graduate stu- urement software BBN later imple- caught our first glimpse when Ray
dents at that time and subsequently mented in each switch. It was natural Tomlinson introduced email in 1972
Internet luminaries. The backstory that we would be the first node so that and it very quickly took over a major
starts with Ivan Sutherland who be- we could begin to conduct experimen- portion of the traffic; that was when
came head of IPTO for ARPA in 1964. tation and make measurements of it became clear that a major use
Ivan visited UCLA in 1965 and sug- what was going on. would be to support people-to-people
gested we network the three nearly communication. Put another way, we
identical IBM 7090s on campus. But The first message on the Internet was completely missed social networking
the three administrators didn’t want “Lo” which seems to have Biblical con- as a major use of the Internet. In-
to share their computers, so that net- notations that go along with the Cre- deed, it has been the case over and
work was never implemented. ation Story. Was this deliberate? over again that the Internet commu-
Not at all. We were trying to send the nity has been surprised each time ma-
How did it finally happen? text “Login” to login to the SRI host but jor new applications have exploded in
Bob Taylor (who later led Xerox there was a bug and the software use (for example, the World Wide
PARC) took over IPTO after Ivan. Bob crashed after sending “Lo.” Of course, Web, peer-to-peer file sharing, blogs,
was convinced that IPTO needed a the bug was in SRI’s software, not ours user-generated content, search en-
computer network to link the sites he nor in the network itself! gines, shopping engines, social net-
was supporting so that they could works). What we are good at predict-
share each other’s computers and How did we get from the first ARPANET ing is the underlying infrastructure of
applications. Bob convinced Larry to the Internet we know today? the Internet (networking technology,
Roberts to come to Washington in The first host-to-host protocol was IoT, wireless access, mobility, and so
1966 and head up this idea of deploy- called NCP but soon it became clear forth). One other aspect of today’s In-
ing a computer network. While we that the ARPANET would shortly be ternet we did not foresee was the
think of the Internet today to send emergence of the dark side (in all its
email and support social networks, manifestations) that plagues us today.
the motivation then was to share the One other aspect
expensive computers that ARPA was While the Internet was gaining steam,
funding at sites like Utah (for graph- of today’s Internet you trained several generations of re-
ics), Stanford Research Institute (da- we did not foresee markable students whose Ph.D. the-
tabases, Doug Engelbart was there). ses and papers with you greatly influ-
Larry was familiar with my network- was the emergence enced the Internet and analyses of
ing research and publicly credits my of the dark side time-shared systems. Tell us more ...
thesis for giving him confidence to There is so much to tell, so let me
spend millions of dollars of ARPA (in all its provide a small sample only. My first
money on this crazy idea. Larry was manifestations) that student was Ed Coffman, who worked
also well aware of Baran’s work and on some extensions to priority
that of Davies (who had even built a plagues us today. queueing and time sharing. Most of
single-node packet switch) and in- my students who followed concen-
corporated their ideas in the ARPA- trated on various performance anal-
NET design. yses of aspects of the Internet as it
emerged. For example, the early AR-
How did Larry get everyone together to PANET did synchronous (periodic)
create the ARPANET, the precursor to routing updates but Gary Fultz’s
our Internet? thesis analyzed the benefits of
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 35
viewpoints
Viewpoint
Algorithms, Platforms,
and Ethnic Bias
How computing platforms and algorithms can potentially
either reinforce or identify and address ethnic biases.
E
T H N I C A N D O T H E R biases embedded in the algorithms that make
are increasingly recognized or structure real-world decisions.
as a problem that plagues Without proper We model algorithm development,
software algorithms and da- mitigation, preexisting implementation, and use as having five
tasets.9,12 This is important distinct nodes—input, algorithmic op-
because algorithms and digital plat- societal bias will erations, output, users, and feedback.
forms organize ever-greater areas of be embedded in the Importantly, we incorporate users be-
social, political, and economic life. Al- cause their actions affect outcomes. As
gorithms already sift through expand- algorithms that shown in the accompanying figure, we
ing datasets to provide credit ratings, make or structure identify nine potential biases. They are
serve personalized advertisements, not mutually exclusive, as it is possible
match individuals on dating sites, real-world decisions. for multiple, interacting biases to exist
flag unusual credit-card transactions, in a single algorithmic process.
recommend news articles, determine
mortgage qualification, predict the Types of Bias
locations and perpetrators of future Training Data Bias. Predictive algo-
crimes, parse résumés, rank job can- rithms are trained on datasets, thus
didates, assist in bail or probation accompanying figure) that extends pre- any biases in the training data will be
proceedings, and perform a wide va- vious research to locate where bias may reflected in the algorithm. In principle,
riety of other tasks. Digital platforms occur in an algorithmic process.6 this bias should be easy to detect, but
are comprised of algorithms executed the sources may be difficult to detect.
in software. In performing these func- Interrogating Bias in Presumed gold standard datasets, such
tions, as Lawrence Lessig observed, Algorithmic Decision-Making as government statistics or even judi-
“code” functions like law in structur- Of course, social bias has been long rec- cial conviction rates, frequently con-
ing human activity. Algorithms and on- ognized. Some attribute the introduc- tain bias. For example, if the criminal
line platforms are not neutral; they are tion of bias into algorithms to the fact justice system is biased, then, absent
built to frame and drive actions.8 that software developers are not well corrections, the algorithm will mirror
Algorithmic “machines” are built versed in issues such as civil rights and such bias. Thus, training sets can be
with specific hypotheses about the fairness.3 Others suggest it is far more subtle contributors to bias.
relationship between persons and deeply embedded in society and its Algorithmic Focus Bias. Algorith-
things. As techniques such as machine expressions.4 Inspired by value chain mic focus bias occurs from both the
learning are more generally deployed, research, while our model cannot re- inclusion and exclusion of particular
concerns are becoming more acute. solve bias; it provides a template for variables. For instance, the exclusion
For engineers and policymakers alike, identifying and addressing the sources of gender or race in a health diagnos-
understanding how and where bias of bias—conscious or unconscious— tic algorithm can lead to inaccurate or
can occur in algorithmic processes can that might infect algorithms. What is even harmful conclusions. However,
help address it. Our contribution is the certain is that without proper mitiga- the inclusion of gender, race, or even
introduction of a visual model (see the tion, preexisting societal bias will be ZIP codes in a sentencing algorithm
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 37
viewpoints
Potential biases and where they may be introduced in the algorithmic value chain.
Source: The first six biases were adapted from Danks, D., & London, A.I. (2017).
The visualization and remaining materials are by Silva and Kenney.
can lead to discrimination. This is the Algorithms, particularly artificial intel- bias can occur from either side, or party,
conundrum: in certain cases, such ligence and machine learning, often in a digital interaction. Or, even more
variables must intentionally be used to generate opaque results. The reasons deliberately, anonymous online hack-
produce less-biased outcomes.5 for the results may even be inexplicable ers purposely “taught” Microsoft’s Tay
Algorithmic Processing Bias. Bias can to the algorithm’s creators or the soft- chatbot, which was opened to the public
be embedded in the algorithm itself. ware’s owner. For example, when a ma- for only a few days in 2016, to respond
One source of such bias is the inclusion chine-learning program recommends with racially objectionable statements.
and weighting of particular variables. denial of a loan application, the bank Effectively, the algorithm or platform
Consider the case of a firm’s chief sci- official conveying the decision may not provides users with a new venue within
entist’s finding that “one solid predictor know the exact reasons for denial. The which to express their biases.
of strong coding is an affinity for a par- absence of transparency makes it diffi- Feedback Loop Bias. Algorithmic sys-
ticular Japanese manga site.”10 If this is cult for the subjects of these decisions tems create a data trail. For example,
embodied in job-candidate-sorting soft- to identify discriminatory outcomes or the Google Search algorithm responds
ware, then this seemingly innocuous even the reasons for the outcome. to and records a query that becomes
choice might exclude particular qualified Automation Bias. Automation bias customized input for subsequent
candidates. Effectively, a desired proxy results from the belief the output is searches. The algorithm learns from
trait inadvertently excludes certain fact, rather than a prediction with a con- user behavior. For example, in predic-
groups that could perform the job. fidence level. For instance, credit deci- tive policing, the algorithm relies al-
Transfer Context Bias. Transfer con- sions are now fully automated and use most entirely on historical crime data.
text bias occurs when algorithmic out- group aggregates and personal credit Suppose the algorithm sends police
put is applied to an inappropriate or un- history.13 The algorithm gives certain officers into a neighborhood to pre-
intended context. One example is using people lower scores and limits their vent crime. Not surprisingly, increased
credit scores to make hiring decisions. access to credit. Credit denial means police presence leads to higher crime
Bad credit is equated with inferior future their scores cannot improve. Often, the detection, thereby raising the statisti-
job performance, despite little evidence subjects and decision-makers are un- cal crime rate. This can motivate the
that credit scores are related to work aware of the algorithm’s assumptions dispatch of more police, who make
performance. If the undesirable, but ir- and uncritically accept the decisions. more arrests, thereby initiating a feed-
relevant trait is correlated with ethnicity, The European Union’s GDPR’s Article back loop. In another example, Google
then it might lead to biased outcomes. 22 has attempted to provide some pro- Search can learn that ethnically biased
Interpretation Bias. Interpretation tection by limiting automated algorith- websites are often selected and there-
bias arises when users interpret algo- mic decision processes for legal or the fore recommend them more often,
rithmic outputs according to their in- equivalent life-affecting decisions.11 thereby propagating them. As smart as
ternalized biases. For example, a judge Consumer Bias. The biases that hu- algorithms can be, human monitoring
can receive an algorithmically generat- man beings act upon in everyday life are continues to be necessary.
ed recidivism prediction score and de- expressed in their online activities. Fur-
cide on the punishment or bail amount ther, digital platforms can exacerbate or Benefits of Platforms
for the defendant. Because individual give expression to latent bias in online and Algorithms
judges may be unconsciously biased, behavior. Users may consciously or un- The potential benefits of algorithmic
they may use the score as a “scientific” consciously discriminate on the basis decision-making are less noticed, but
justification for a biased decision. of a user profile that contains ethnically it can also be used to decrease social
Outcome Non-Transparency Bias. identifiable characteristics. Consumer bias. It is well known that members of
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 39
India
Region
Special
Section
ILLUSTRATION BY SPOOKY POOKA AT DEBUT ART.
FOR CREDITS ON IM AGES IN COLL AGE, SEE P.3.
W
to introduce the India Region special section
E ARE PLEASED
to Communications’ readers. The Indian subcontinent has a
population close to 1.8 billion, and is unique due to its diversity
of people, cultures, spoken languages, and wide disparities in
socioeconomic conditions. The region plays an important role
in the global computing landscape with its highly trained manpower, software
companies, and top universities that produce students that not only serve local
needs, but move around the world and have global impact. We developed this
special section to mirror all these facets.
Last year, we publicized the plans for the special section and made an open
call for contributions through ACM member channels and the ACM India
website. We received 45 proposals through this period and augmented the list
by reaching out to others to cover specific topics and areas. We also received
proposals from colleagues in Sri Lanka and Pakistan. A selection of 22 outlines
were identified for consideration. A workshop held at Microsoft Research in
Bangalore on February 23, 2019 converged on the selection of 17 proposals to
pursue as full articles. These articles underwent three rounds of reviews and
modification. The final section presents nine articles covering Hot Topics and
nine articles following Big Trends.
Computing solutions for the India region must deal with the scale of its popu-
lation. We feature India’s attempt at creating digital infrastructure and solutions EDITORIAL BOARD
at that scale, notably the biometric identification through Aadhar. The other big EDITOR-IN-CHIEF
story is the success and reach of India’s software industry; practically every piece Andrew A. Chien
eic@cacm.acm.org
of software sold in the world has components developed in India. The linguistic
diversity of South Asia is a challenge when creating computer-based solutions, DEPUTY TO THE
EDITOR-IN-CHIEF
starting with suitable keyboards to the challenges of multilingual and mixed- Lihan Chen
lingual search. Another vibrant aspect of India is the resurgence of its start-ups. cacm.deputy.to.eic
The articles in the special section attempt cover all these stories and more. The @gmail.com
challenges of the socioeconomic milieu of this region are highlighted through CO-CHAIRS, REGIONAL
articles on empowering women through mobiles, using speech to counter illit- SPECIAL SECTIONS
eracy, and in the problems faced by social media. The section also samples some Sriram Rajamani
Jakob Rehof
of the research advances and challenges from this region. Haibo Chen
We hope this collection of articles gives you a glimpse of the unique problems,
INDIA REGION
opportunities, and exciting work in computing from the Indian subcontinent. SPECIAL SECTION
CO-ORGANIZERS
—P J Narayanan, Pankaj Jalote, and Anand Deshpande P J Narayanan
India Region Special Section Co-Organizers International Institute
of Information Technology
P J Narayanan (pjn@iiit.ac.in) is a professor and director of the International Institute of Information Technology Gachibowli, in
Pankaj Jalote
Hyderabad, India. He is the former president of ACM India and former co-chair of ACM India Council. Indraprastha Institute
of Information Technology
Pankaj Jalote (jalote@iiitd.ac.in) is Distinguished Professor at the Indraprastha Institute of Information Technology
Delhi, India, where he previously served as its founding director. Anand Deshpande
Persistent Systems
Anand Deshpande (anand@persistent.com) is the founder, chairman, and managing director of Persistent Systems, a
global technology services company headquartered in Pune, Maharashtra, India.
44 Extreme Classification
By Manik Varma
50 Digital Transformation
in the Indian Government
By Neeta Verma and Savita Dawar 70 88
54 CSpathshala: Bringing 64 The Growth and Evolution of India’s 88 Highlights of Software R&D in India
Computational Thinking to Schools Software Industry By Supratik Chakraborty and
By Vipul Shah By Pankaj Jalote and Pari Natarajan Vasudeva Varma
By Gautam Shroff
and K. Ananth Krishnan
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 43
H
hot topics
Extreme Classification
BY MANIK VARMA
W
HAT generation of retail recom- out a choice, it would take IIT Kanpur, Max Planck,
WOULD mender systems that have you more than 30 years to go Microsoft, MIT, Montreal,
do
YOU things delivered to your through a billion choices. NEC, NUS, NYU, Stanford,
if you doorstep just as you need In 2012, state-of-the-art Technion, TU Poznan, UC
had the them? Or would you try and multi-label classification Davis, UT Austin, Yahoo,
super- predict the next word about algorithms were struggling and others have developed a
power to accurately answer, to be uttered by U.S. Presi- to pick the correct subset of plethora of algorithms with
in a few milliseconds, a dent Donald Trump? options in questions involv- varying trade-offs between
multiple-choice question The objective in ex- ing thousands of choices. the prediction accuracy, the
I MAGE BY R.CL AS S E N
with a billion choices? treme classification, a new Then, in 2013, a team from prediction time, the training
Would you design the next research area in machine Microsoft Research India time of the classifier and its
generation of Web search learning, is to develop algo- and IIT Delhi developed a size. Most of these algo-
engines, which could pre- rithms with such capabili- classifier1 that could scale to rithms are either based on:
dict which of the billions ties. The difficulty of the task 10 million choices, thereby
of documents might be can be judged from the fact laying the foundations of
relevant to a given query? that, even if it were to take the area. The approach was
Would you build the next you just a second to read based on the realization that
only a handful of choices
would be relevant for any
Extreme classification given question on average.
The trick was therefore to
has found applications in diverse quickly eliminate the mil-
areas ranging from information lions of irrelevant choices.
The classifier could then
retrieval to recommender accurately and efficiently
systems to computational choose from the remaining
hundred or so options.
advertising to natural language Since 2013, extreme clas-
processing and even sification has come to be a
thriving area of research in
computer vision. both academia and industry.
Groups from Amazon, CMU,
Columbia, Facebook, Fudan
University, Google, Hum-
boldt University, IIT Delhi,
References
1. Agrawal, R. et al. Multi-label learning
with millions of labels: Recommending
advertiser bid phrases for Web pages.
In Proceedings of the Intern. World
Wide Web Conference (Rio de Janeiro,
Brazil, May 2013).
2. Bengio, S. et al. Extreme classification
(Dagstuhl Seminar 18291). v, 7 (2019),
62–80.
3. The Extreme Classification Repository;
https://bit.ly/2IDtQbS
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 45
india region hot topics
I
NFORMATION AND ment and adoption of ICTs where only 7%a of women care. More than half a
COMMUNICATION an understanding of the are included in the formal million women, 99%
technology (ICT) specific needs and chal- financial sector (have ac- of them in developing
interventions are lenges faced by women is cess to financial services) countries, die each year
increasingly be- imperative. compared to South Asia’s from pregnancy-related
ing used in devel- Given the patriarchal average of 37% (in contrast, causes, of which Pakistan
oping countries to enable structures that constrain 76% of women in India alone is responsible for an
economic growth, employ- women in Pakistan, tra- own bank accountsb). In estimated 30,000 deaths.c
ment, and empowerment. ditional Western digital India, the main driving Large parts of Africa have
There is, however, growing solutions do not work. De- force behind increases in tackled this using mobile
agreement that the impact tailed ethnographies reveal financial inclusion has phones to run services
of ICTs in the Global South specific cultural, religious, been the ‘Jan Dhan Yojana’ like MoTech and Abiye
is not gender neutral but and social contexts ICT scheme that mandated to deliver maternal and
amplifies the existing interventions must design state-owned banks open at child health informa-
gender inequalities within around. We explore the least one account for every tion.5,6 Similarly, in Paki-
these countries.2,7 This is specific needs and con- unbanked household. stan, we have launched
also true for Pakistan and straints of low-literate, low- Although digital fi- mobile-phone-based
India, where most ICT income women in Pakistan nancial services (DFS) are information systems to
interventions deployed and tackle the gendered presented as a viable alter- give low-income mothers
have largely ignored the design of technologies for native to formal banking access to critical preg-
unique needs of the female financial inclusion, mater- structures for the develop- nancy information. Based
Pakistani (48.63%) and In- nal healthcare, and digital ing world, we show the on qualitative interviews
dian populations (48.53%). social connectivity. concept of DFS, as it stands with doctors and pregnant
Multi-country research on Financial inclusion currently, is unsuited to the mothers, we designed and
the impact of ICTs reveals refers to a process that financial needs of Pakistani launched a combination
their great potential for ensures ease of access and women. Unlike the devel- of SMS text messages and
bringing about positive affordability of financial oped world, mobile bank- interactive voice response
socioeconomic change services for a popula- ing in Pakistan must ac- (IVR) system that provides
and gains in economic tion and is an important count for phones as shared critical information for
growth.9 Similarly, studies means to tackle poverty resources, secret/hidden maternal care. Impact
reveal ICTs are one of the and inequality. Financial financial transactions evaluations of the system
main drivers of economic exclusion is a crucial issue (women hide money from reveal that targeted mes-
growth in Asia, the Middle facing women in Pakistan, family/spouse), flexible saging has the largest,
East, and Sub-Saharan Afri- which is on the list of seven and self-determined sav- statistically significant im-
can.3,8 However, in order to countries that constitute ings, and loan and dowry pact on pregnancy-related
ensure entire populations half of the unbanked popu- dynamics.7 In the Western knowledge.4
benefit from the deploy- lation around the world context, mobile phones are Another key issue faced
considered and designed by women in Pakistan is
for use as personal devices, the lack of digital social
Given the patriarchal structures unlike Pakistan where only connectivity. This is be-
39% of women own mobile cause all social media is
that constrain women in Pakistan, phones.7 based on the assumption
traditional Western digital Another crucial area of literacy and Pakistan is
where ICTs have great a country with an overall
solutions do not work. potential for impacting literacy rate of 58%; the
women’s lives is health- literacy rate of women is
48%. To solve this, we have
a See http://bit.ly/2KE9Ma1
b See http://bit.ly/2F1omVE c See http://bit.ly/2WAlEB5
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 47
india region hot topics
Turbocharging Database
Query Processing and Testing
BY JAYANT R. HARITSA AND S. SUDARSHAN
D
ATA BA S E M AN - contributions on robust of queries. However, a ure 1 for a representative set
AG E M E N T SYS- query processing, holistic common limitation in this of queries from the industry
T E M S (DBMS) optimization of database prior work is the inability standard TPC-DS bench-
constitute the applications, and testing to furnish performance mark, the comparison yard-
backbone of to- strategies for SQL queries guarantees. stick being the PostgreSQL
day’s informa- and database engines. A radically different native optimizer.
tion-rich society. A primary approach that addresses These techniques repre-
reason for the popularity of Robust Query Processing the guarantee issue, called sent an important mile-
database systems is their A crucial input to generat- PlanBouquet,a has been stone in the history of ro-
support for declarative ing efficient query execu- recently developed at IISc bust query processing since
queries, typically in the tion strategies, called plans, Bangalore.3 PlanBouquet they are the first to provide
SQL query language. In this are the statistical estimates completely abandons quantitative performance
programming paradigm, of the output data volumes the classical estimation guarantees, addressing a
the user only specifies the for the algebraic predicates process for error-prone critical need of the database
end objectives, leaving it to present in the query. In selectivities—instead, it community.
the DBMS to automatically practice, these estimates, employs a carefully cali-
identify the optimal execu- called selectivities, are often brated “trial-and-error” Holistic Optimization
tion strategy to achieve significantly in error with sequence of time-budgeted Database-backed applica-
these objectives. Declara- respect to the actual values plan executions that are tions often suffer from
tive specification of queries subsequently encountered progressively capable of poor performance arising
is also central to parallel during query execution. handling more and more from sub-optimal ways in
query execution in modern The unfortunate outcome data until the query is which imperatively writ-
big data platforms. is a poor plan choice, eventually taken to comple- ten application programs
Query processing and resulting in query response tion. An advanced variant access information from
optimization have been times that may be worse by of this approach, called a database. For example,
extensively researched for orders of magnitude relative SpillBound, guarantees that many application programs
close to five decades now, to the optimal plan choice the performance is always issue a long sequence of
and are implemented in with the correct selectivi- within a factor of (D2+3D) queries to a database, each
all contemporary database ties. A considerable body of relative to the ideal, where D of which requires a signifi-
systems. Nevertheless, literature exists on improv- is the number of predicates cant round-trip time due to
important challenges re- ing the statistical qual- whose selectivity estimates latency in the database and
main unsolved, and Indian ity of selectivity estimates may be erroneous.5 network. Such inefficien-
universities have played a through sophisticated sum- Further, empirical evalu- cies cannot be addressed
visible role in addressing mary structures, feedback- ations on industry-standard either by traditional
these issues. As exemplars, based adjustments, and benchmarks have shown database query optimizers
we highlight recent research on-the-fly re-optimization SpillBound to perform, in or by traditional compiler
the worst-case, within a optimizations. The DBridge
factor of 10–20 of the ideal, systemb developed at IIT
The DBridge suite of techniques whereas contemporary Bombay therefore tackles
database systems may suf- this problem by rewriting
brings the powerful benefits fer performance degrada- application code to opti-
of declarative query optimization tion factors running to the mize data access.
1,000s and beyond in such DBridge carries out a
to imperative code. environments. This perfor- set of potent equivalence-
mance robustness of Spill- preserving transforma-
Bound is quantified in Fig- tions on imperative code
a https://dsl.cds.iisc.ac.in/projects/ b https://www.cse.iitb.ac.in/infolab/
QUEST dbridge
Original Database
106 Application TAXI
Program
105
104
3D_Q96
4D_Q7
4D_Q26
4D_Q27
4D_Q91
5D_Q19
5D_Q29
5D_Q84
5D_Q18
Program
Figure 1. Performance robustness profile. Figure 2. Rewrites for optimizing data access.
to speed up data access. garnering excellent reviews is currently operational at expressive power comes
The transformations suc- from users.c multiple universities. great challenges.” In this
cessfully carry out batch- Collectively, the DBridge The testing of Big Data article, we have highlighted a
ing and asynchronous suite of techniques brings platforms is addressed by few recent successes in tack-
submission of queries,6 the powerful benefits of de- the CODD projecte at IISc ling these challenges, but
prefetching of query clarative query optimization Bangalore, using a distinc- there remain rich opportuni-
results, and conversion of to imperative code, open- tive metaphor of “dataless ties for further contributions
procedural code to SQL. ing a new research frontier. databases.”1 Here, data- to the field. Productive future
A metaphorical depiction More details on these tech- bases with a desired set of work areas include extend-
of batching rewrites in niques may be found on the characteristics can be ef- ing the holistic optimization
DBridge is shown in Figure DBridge project home page. ficiently simulated without concept to new domains (for
2, where queries that are explicit creation or persis- example, machine learn-
issued one-at-a-time, sym- Query and Engine Testing tent storage of the contents. ing), and leveraging query
bolized by the individual With the onset of the Big This approach is essential and data characteristics to
“taxis,” are batched into Data world, where data is since traditional testing deliver tighter robustness
a single unified request, the engine driving virtu- techniques, which involve guarantees.
carried by a “bus.” Each ally all aspects of human construction of represen-
transformation caters to endeavor, it is vitally tative databases and re- References
1. Ashoke, S. and Haritsa, J. CODD: A
a restricted scope and is important to ensure both gression query suites, are dataless approach to big data testing.
therefore easy to prove cor- the applications and the completely impractical at PVLDB 8, 12 (Aug. 2015), 2008–2011.
2. Chandra, B., Chawda, B., Kar, B., Reddy,
rect, but in tandem they underlying platforms are Big Data scale due to the K., Shah, S. and Sudarshan, S. Data
generation for testing and grading SQL
can successfully rewrite functionally correct. The time and space overheads queries. VLDB J. 24, 6 (Dec. 2015),
complex application pro- XData systemd developed at involved in their execu- 731–755.
3. Dutt, A. and Haritsa, J. Plan bouquets:
grams. Further, the Cobra IIT Bombay supports test- tion. The CODD tool has A fragrant approach to robust query
component of DBridge4 ef- ing of SQL queries by gen- been successfully used processing. ACM Trans. Database Syst.
41, 2 (June 2016), 11–1:37.
ficiently chooses the least erating datasets designed for testing of database en- 4. Emani, K. V., and Sudarshan, S.
cost program from many to detect many types of gines in the software and Cobra: A framework for cost-based
rewriting of database applications. In
alternative transformed common errors.2 XData can telecom industries. Proceedings of the IEEE Intl. Conf. on
programs, by leveraging be used in database courses Data Engg. (Apr. 2018), 689–700.
5. Karthik, S., Haritsa, J., Kenkre, S.,
concepts from query opti- to help students master Future Research Pandit, V. and Krishnan, L. Platform-
mization based on alge- the nuances of SQL query An important reason for the independent robust query processing.
IEEE Trans. Knowl. Data Eng. 31, 1
braic equivalence rules. formulation and verify their rapid adoption of SQL in (Jan. 2019), 17–31.
Techniques for holistic correctness; further, the the 1970s was its simplicity, 6. Ramachandra, K., Chavan, M.,
Guravannavar, R. and Sudarshan,
optimization of queries con- XData system facilitates au- which lent itself to effective S. Program transformations for
taining imperatively coded tomated grading of incorrect query optimization. Howev- asynchronous and batched query
submission. IEEE Trans. Knowl. Data
user-defined functions queries by assigning partial er, a host of complex features Engg. 27, 2 (Feb. 2015), 531–544.
(UDFs) were developed markings that reflect the have been added over the
Jayant R. Haritsa (haritsa@iisc.ac.in) is a
jointly by IIT Hyderabad severity of the errors. XData years, and today’s query professor at the Indian Institute of Science,
and IIT Bombay; some of processing world can be Bangalore, India.
these mechanisms have paraphrased as “with great S. Sudarshan (sudarsha@cse.iitb.ac.in)
c https://www.microsoft.com/en-us/ is a professor at the Indian Institute of
subsequently been imple- research/project/froid Technology, Bombay, India.
mented and released in d https://www.cse.iitb.ac.in/infolab/ e https://www.cse.iitb.ac.in/infolab/
Microsoft SQL Server 2019, xdata xdata © 2019 ACM 0001-0782/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 49
india region hot topics
Digital Transformation
in the Indian Government
BY NEETA VERMA AND SAVITA DAWAR
D
I G I TA L I N D IA
ISthe flagship
program of the
Government
of India with a
vision to trans-
form India into a digitally
empowered society and
knowledge economy. This
program is centered on the
vision of offering digital
infrastructure as a core
utility to every citizen, pro-
viding governance and ser-
vices on demand, enabling
the digital empowerment
of citizens.2 Besides policy
making facilitation to the
IT industry and start-ups,
the government has also
adopted state-of-the-art
ICT for its own transfor-
mation for efficient and
effective delivery of infor-
mation and services to
citizens at large. A specific tion of the Indian govern- of-the-art infrastructure, ment mail, GIS infrastruc-
focus has been on reach- ment. building solutions, as well ture, the public finance
ing the last mile as digital NIC is the driving force as advising individual de- management system, and
inclusion is at the core of of the Digital India pro- partments on action plans digital payments are key
the Digital India program. gram and has also helped and adoption of appropri- pieces which help provide
National Informatics Cen- the government be in the ate technologies. a foundation for govern-
tre (NIC)5 under the Min- forefront in the use of NIC set up the VSAT- ment departments to build
istry of Electronics and information technology. It based network for IT systems that deliver
Information Technology is has been working with the inter-government com- services to citizens.
an important stakeholder government for over four munication in 1982. The
in the digital transforma- decades, providing state- X400-based electronic mes- Digital Platforms
saging service was used by Government has been us-
government long before the ing ICT-based systems to
Centralized systems Internet was introduced to implement and manage
the country. With the help its programs for over three
have had a huge impact on of NIC, the government has decades. These systems
the work of government led the country in adopt- have evolved with the ad-
IMAGE BY KONSTA NT IN FA RAK TINOV
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 51
india region hot topics
References
1. DBT https://dbtbharat.gov.in/
E-Courts 2. Digital India https://digitalindia.gov.in/
3. E-Way bill https://ewaybill.nic.in/
4. Ecourts https://ecourts.gov.in/
ecourts_home/
5. National Informatics Centre https://
Key Features E-Courts Mobile App www.nic.in/
Common Software across 6. PFMS https://pfms.nic.in/
the Country NewDefaultHome.aspx
7. Sengupta, D.and Shastri, N. Digital
Payments through PFMS—Facilitating
Open Source Technology Kiosks in Court Pendency Dashboards digital inclusion and accelerating
Complexes on NJDG transformation to a ‘Digital Economy.’
In Proceedings of the 12th Intern.
Conf. Theory and Practice of Electronic
Implemented for 10 Years Governance (Apr. 2019).
3000 Court Complexes Citizen services on Video Conferencing 8. Unique Identification Authority of
More then 18000 Courts ecourts.gov.in Facility India; https://uidai.gov.in/
Case Status, Orders,
Automated e-Mails
Cause Lists on portal
Neeta Verma is Director General of the
Key
National Judicial Data Grid National Informatics Centre in New Delhi,
Services
India.
SMS Push / Pull Mobile App
Savita Dawar is Deputy Director General
of the National Informatics Centre in New
Delhi, India.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 53
india region hot topics
Education | DOI:10.1145/3343445
CSpathshala: Bringing
Computational Thinking to Schools
BY VIPUL SHAH
B
H U M I KA A N D Introducing a comput- mendations,9 CAS U.K.
P U S H KA R, ing curriculum has not curriculum,2 code.org
12-year-old been easy and has posed lessons,a Computer Masti,8
students from several unique challenges: and CS unplugged mate-
a government ˲˲ Scale: As per govern- rial,4 and have developed an
school in the ment reports,7 India has unplugged computing cur-
village of Takalkarwadi, over 1.6 million schools riculum5 influenced by the
in Khed, Maharashtra, offering K–12 education to New Jersey discrete math
are playing the “Guess My 300 million students. To curriculum for problem
Birthdate” game. The goal compound the problem, solving.10 It includes top-
of the game is to find the in addition to two national ics like systematic listing,
“Guess my Birthdate” activity
date by asking the least at Takalkarwadi School, boards of education, each counting and reasoning
number of questions. The Maharashtra. of the 29 states in India has (systematically arriving at
students’ strategy is to ana- its own education board! all possible answers and
lyze each question in terms puting into mainstream While English is the com- reasoning on complete-
of the number of dates it curricula, and to train mon language of instruc- ness), iterative patterns
eliminates. teachers so that every child tion in the urban areas, and processes (looking for
Some 300,000 students in India learns computing 70% of the population patterns to generalize and
from 750 schools in 11 as a science by 2030. residing in the rural areas apply to given problem),
states throughout India The National Policy on is educated in the state’s organizing and processing
are learning computing ICT for School Education regional language. information (data collec-
through “unplugged” in India6 advocates the ˲˲ Infrastructure: 63% of tion, representation, and
activities as part of CSpath- development of a model the schools have electric- analysis), discrete math-
shala,1 ACM India’s educa- Curriculum for ICT that ity and only 27% of schools ematical modeling (abstrac-
tion initiative. The name would include conceptual have computers. In rural ar- tions like graphs and trees),
CSpathshala is derived knowledge enhancement eas, electricity may be avail- following and devising in-
from computer science and and enable the develop- able for a few hours a day structions (initially follow-
Pathshala, which means ment of generic skills with and the school may have ing, then devising a precise
place of learning or a focus on digital literacy. only 1–2 computers. Urban set of instructions and later
school. Launched in 2016, Although teaching com- schools are better equipped evaluating multiple solu-
CSpathshala’s primary puter science has already with computer labs that tions) and programming.
goals are to promote com- been introduced in urban allow a computer to be Strategies to address the
puter science education in India, it focuses primarily shared by 2–3 students. challenges mentioned here
K–12, to influence policy- on digital literacy and a bit ˲˲ Teacher skills: A survey include:
makers to introduce com- of programming. we conducted corrobo- ˲˲ Efforts have been
rated findings in Raman directed toward carrying
et al.12 Teachers from rural out a pilot program with
Some 300,000 students from areas had no computing 500+ rural government
background. Moreover, schools and working with
750 schools in 11 states only 59% of the teachers 2–3 education boards. The
throughout India are learning working in urban areas had Tamil Nadu state educa-
exposure to some form of tion board has adopted
computing through “unplugged” computing education, with computational thinking as
activities as part of CSpathshala, only 10% having a comput- part of its math curriculum
er science degree. for 10,000 schools. Another
ACM India’s education initiative. A national curriculum state educational board will
committee explored the begin a pilot shortly with
CSTA K–12 curriculum
framework and recom- a www.code.org
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 55
india region hot topics
Finance | DOI:10.1145/3343458
Creative Disruption in
Fintech from Sri Lanka
BY AJIT SAMARANAYAKE, SAMPATH TILAKUMARA,
THAYAPARAN SRIPAVAN, AND RASIKA WITHANAWASAM
D
U RI N G T H E
1990S,the Sri
Lankan IT
sector was
sandwiched
between
the forces of free mar-
ket competition and the
internal turbulence due to
civil unrest. The relatively
small internal marketplace
made it difficult to attract
foreign investments and
expand businesses beyond
IT offshoring. However,
stock trading was a brick-
and-mortar business that
presented promising
growth potential with the
advent of financial technol- Figure 1. The Colombo Stock Exchange (CSE), the first customer. (Photo courtesy of CSE)
ogy (fintech).
Sensing this opportunity, MillenniumIT intro- Stock Exchange Group lenges served to help con-
a set of seasoned managers at duced novel concepts in acquired MillenniumIT tain the costs of scalability,
an existing IT services busi- designing complex elec- and secured the status while preserving high re-
ness set off with a broader tronic trading systems with of the “fastest trading siliency. LSEG Technology
vision, marking the birth of predictable performance system” through a subse- quit using costlier hard-
MillenniumIT (now known that met the regulated quent technology upgrade. ware-dependent resiliency
as LSEG Technology), a ultra-high resiliency re- Today, LSEG Technology6 by introducing software
fintech product company. quirements. Being an early (as it is now known) is fault-tolerance models into
The farsighted entry into mover allowed for the slow, a key contributor to the a freshly built common
fintech, and the experience steady penetration of the overall group’s technology technology framework for
in mobilizing local talent, company’s technology into strength, and powers over fintech applications.
contributed to the early capital markets around the 40 capital market institu- Such software fault toler-
success of MillenniumIT. world. In 2009, the London tions around the globe. ance introduced patented
Since its inception, cre- models of replication,
ative disruption along four synchronization, and recov-
The early application of ‘creative major themes has served ery that ran on commod-
to keep LSEG Technology ity hardware at a fraction
disruption’ to a niche market competitive. of the cost, establishing
with immense growth potential new industry benchmarks
1. Software resilience on system availability for
has proven to be a very effective for scalable distributed mission-critical fintech ap-
tool and strategy. systems. plications.
Scalable distributed sys- A first-of-its-kind deploy-
tems are at the core of the ment in Sri Lanka’s national
architecture of the organi- stock exchange (CSE) in
zation’s electronic trading 1995 was followed by imple-
systems. Competitive chal- mentations in global trad-
ing hubs in London, Milan, jump in end-to-end latency 4. Description-driven shown to generate ap-
Oslo, and Johannesburg. was possible with a full-stack systems. proximately 80%–90% of the
re-architecture, and being The description-driven code of a system. An initial
2. and 3. the first to infuse emerging approach to software prototype of a fully func-
High-performance and transport technologies (such generation was again a tional post-trade clearing
heterogeneous computing. as Infiniband) helped the disruptive response to meet system demonstrated that a
Execution of complex func- London Stock Exchange gain demands for higher quality 5k-line system description
tionality at ultra-low latency market share and stay ahead and quicker delivery, with will generate up to 950k
is imperative for electronic of other leading exchanges. lower costs. LSEG Technol- lines of C++, JavaScript, or
trading systems. In 2014, an award-win- ogy introduced a patented SQL code.
That pattern of con- ning8,9 low-latency market business rule engine
founding expectations led data distribution platform in 1998, which allowed Summary
to a number of inflection was introduced with the use flexibility in specifying The success of LSEG Tech-
points when LSEG Technol- of field programmable gate business features without nology bears testimony that
ogy offered an ultra-low- arrays (FPGA)2 that yielded a requiring redeployment or despite contextual barriers,
latency trading system to 95% performance improve- upgrades. The core of this organizations in this region
address the London Stock ment (sub 5µs end to end) approach was extensible to can indeed become glob-
Exchange’s requirements in compared to homogeneous the description of an entire ally competitive technology
2011. The low-latency exter- software.3–5,7 The latest gen- system (that is, data model, leaders in specialized niche
nal interfaces developed as eration of the heterogeneous business functionality, markets. The early applica-
part of this solution allowed (FPGA, GPU) application work flows, user interface tion of ‘creative disruption’
co-located high-frequency suite enables new business (UI), and deployment). to a niche market with im-
traders to take advantage of models by realizing financial Using a combination of mense growth potential has
the ultra-low latency of the risk simulations and deep theorem provers and code proven to be a very effective
platform. A sub 100µs step- learning in real time. generators, it has been tool and strategy.
The increasing pace of
technological advance-
1 1 3 Smart systems (AI/ML) ments warrant a balanced
Description driven systems outlook toward agility, far-
Distributed trading Process pair FPGA accelerated Domain specific hardware
system fault tolerance market data Elastic Scalability sighted bets on technology,
and investments in intellec-
1997 2002 2014 tual capital, to exploit the
Future
unfolding opportunities of
the future.1
1996 1998 2009 2016
References
Business rule Trading system GPU accelerated 1. Bloomberg LP U.S., 2019; https://
engine latency < 100µs realtime risk bloom.bg/2KG8NGl
2. Businesswire.com. 2013; http://bit.
ly/2R5Brlm
4 2 3
3. Finextra.com. 2016; http://bit.
ly/2wMtF6D
4. Fnlondon.com. 2016; http://bit.
Figure 2. Themed milestones of creative disruption. ly/2K79sRC
5. Ibsintelligence.com. 2016; http://bit.
ly/2WyHC7w
6. London Stock Exchange Group PLC,
Millennium Exchange Trading System U.K. 2019; https://www.lseg.com/
lseg-technology
Firewalls & Client Network
Client
Algorithms GW SEQ DS MDS
Figure 3. An abstract view of the Millennium Exchange trading system. Copyright held by authors/owners.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 57
india region hot topics
AI Applications | DOI:10.1145/3343447
W
H AT H ITS
A visitor
to India
first,
quite
literally,
is the traffic. The combina-
tion of inadequate road
infrastructure, increasing
vehicle population, and
poor driver training and dis-
cipline makes for a chaotic
and often deadly mix. The
result is a high rate of road
accidents, with the estimate
of fatalities ranging from
one every four minutesa to
over 238,000 each year.b
There is much ongoing
work in academia, indus-
try, and startups on using
artificial intelligence (AI) ditions in India, which are toring of the infrastructure (cars), and larger vehicles
and Internet of Things (IoT) quite different from those to identify such hazards is (trucks, buses); in fact, it is
technologies to improve the in the developed world, quite important. not uncommon for even
situation. The general goal is present interesting chal- Vehicles. For cost pedestrians to share the
to have affordable technolo- lenges (see Figure 1). reasons, vehicles often lack road space with vehicles.
gies that work with humans Road infrastructure. advanced features such as The heterogeneity in vehicle
through effective monitoring The road infrastructure Advanced Driver Assistance sizes and speeds often leads
and feedback, rather than has largely grown organi- Systems (ADAS). Also, the to a chaotic flow of traffic,
replacing humans through cally, without the benefit mix of vehicles tends to be far removed from adher-
full autonomy. of long-term planning. It very heterogeneous, ence to lane discipline.
The road and traffic con- is of uneven quality, with spanning two-wheelers (for Drivers. Driving disci-
safety hazards such as pot- example, scooters and pline is generally lacking,
a http://bit.ly/31y18zX
b https://en.wikipedia.org/wiki/
holes, poor lighting, and motorcycles), three-wheel- with drivers often cutting
List_of_countries_by_traffic-re- inadequate signals and ers (for example, auto corners to “get ahead.” Part
lated_death_rate signage. Therefore, moni- rickshaws), four-wheelers of the reason for this is that
TOP PH OTO BY SNEH IT/SH UTTERSTOCK.COM
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 59
india region hot topics
Skill Evaluation
BY SHASHANK SRIKANT, ROHIT TAKHAR,
VISHAL VENUGOPAL, AND VARUN AGGARWAL
U
P WA RD O F FOU R
million gradu-
ates enter the
labor market
every year in
India alone.
India boasts of a large
services economy, wherein
a single company hires
thousands of new employ-
ees every year. Meanwhile,
product companies and
small and medium enter-
prises (SMEs) look for a
few skilled people each.
This requires cost-effec-
tive and scalable methods
of hiring. Interviewing
every applicant is not a
feasible solution.
On the other hand,
graduates from 30,000+
institutes of higher educa-
tion spread across 20+ In- students from various de- labor market. Aspiring Minds spoken skills using such
dian states face a constant mographics being ignored. was formed 10 years ago to formats over MCQs. Evalu-
challenge in signaling their Further, these students address this challenging ating such responses is an
competence to potential have no mechanism to get problem. We have devel- expensive, time-consuming
employers. Companies, feedback on how their skills oped a scalable platform to process involving human
most of which are located compare to those required deliver standardized assess- graders, and suffers from
in the top 20 biggest cities by the industry. ments to test job skills. The standardization concerns.
in the country, bias their Having systems that can platform tests more than Automated grading has the
search by relying on proxies intelligently and scalably two million students every potential to address these
like university name and assess a wide variety of year and is used by 5,000+ issues and impact millions
the city a college is located skills is essential to address- companies including 100+ of job seekers, trainers,
in. Applying such crude ing this broader problem Fortune 500 companies. and corporations.
filters results in meritorious affecting every modern-day A particular challenge At Aspiring Minds, we
in designing scalable as- have, over the last decade,
sessment technologies distilled a framework to cast
A particular challenge in is evaluating subjective, the question of subjective
open-ended responses assessments as problems in
designing scalable assessment to questions. Such ques- computer science, and spe-
technologies is evaluating tions directly simulate a cifically, in machine learn-
skill or a job task within ing (ML).1 In it, candidate
subjective, open-ended the constraints of a test- responses are data points in
IMAGE F RO M SH UTT ERSTOCK.CO M
languages. assess these specific skills. that are causal and address- Rohit Takhar is a research engineer at
Aspiring Minds, Gurugram, India.
˲˲ Customer service. The ˲˲ Domain knowledge. In ing issues of fairness and
Vishal Venugopal is a senior software
IT-enabled services (ITeS) consultation with subject- bias in grading. These form engineer at Aspiring Minds, Gurugram,
market in India employs matter and industry experts, areas of active research. India.
four million people and is we have designed 300+ Varun Aggarwal is co-founder and Chief
Technology Officer of Aspiring Minds,
a US$181-billion industry. tests for domain knowledge References Gurugram, India.
1. Aggarwal, V., Srikant, S., and
Spoken English skills are across various industry ver- Shashidhar, V. Principles for using
central to this industry. ticals such as IT, ITeS, retail, machine learning in the assessment © 2019 ACM 0001-0782/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 61
india region hot topics
Computing Research at
Tata Consultancy Services
BY GAUTAM SHROFF AND K. ANANTH KRISHNAN
S
OMETIME ture, and playing the role of
IN THE a growth and transforma-
early 1960s tion partner to large enter-
a young gen- prises worldwide.8
eral manager In this article we focus
of the Tata on the unique role that
Electric Co. in Mumbai research and innovation
(then Bombay) visited the has played in TCS’ journey
nearby Tata Institute of Fun- from being the Tata group’s
damental Research (TIFR),a computing division to its
where India’s first electron- current place in global tech-
ic-stored program comput- nology consulting. We shall
er resided. That manager, highlight both some chal-
F.C. Kohli, began using the lenges faced along the way
computer to optimize load as well as lessons learned;
dispatch operations for the accompanying figure is
the city. Within a few years a snapshot of this journey.
J.R.D. Tata, then chair- In 1981, TCS established
man and doyen of the Tata a dedicated corporate
group of companies, called research facility in Pune,
upon Kohli to look after the the Tata Research, Develop-
new computing division— ment, and Design Centre
Tata Consultancy Services (TRDDC).b It was headed by
(TCS). That division soon E.C. Subbarao, a prominent
began catering to clients materials scientist from IIT
both outside the group and Kanpur, who began apply-
outside India—in Europe ing computational materi- successfully automating a product for the Swiss
and North America—pio- als engineering for Tata many of the conversion proj- private banking industry,
neering the ‘offshore devel- Steel, which then domi- ects that TCS won through and charged research with
opment’ model. nated the group’s business. the 1980s, for example, from creating a new generation of
The seeds of the Indian Indeed, research in TCS was one language/database to a CASE tools—MasterCraft.12
software industry had been joined at the hip with busi- more modern platform. By employing code-gener-
sown. ness from its inception. In the late 1980s, TCS ation, the banking product
In 2019, TCS revenues Soon Kesav V. Nori joined was awarded a large, com- remained insulated from
crossed $20 billion, ranking TRDDC from TIFR and plex development project multiple generations of
third after IBM and Accen- began adapting compiler by SEGA, the Swiss financial technology change. Today,
technology to TCS’ fledgling depository, clearing and .NET, Java, and Web ver-
software services business, settlement organization. sions of the product (TCS
a Tata Institute of Fundamental
Research is a research laboratory
Such a project demanded BαNCS9 ) have also been
funded and run by the Govern- b TRDDC is now part of TCS computer-aided software instantiated even while
PHOTO BY BIL A L A LIYAR M/SH UTTERSTOCK.CO M
embedded systems in Ban- Today, TCS Research journey; we mention some research was far from widely
galore and later Kolkata, and covers a wide variety of areas critical ones here: First, shared. Only hindsight
architecture in Delhi. These (as listed in the figure) and research initiatives have al- shows indulgence paying off
new groups quickly began is increasingly multidisci- ways preceded their appli- as the world itself is changing
impacting TCS business: plinary. Many long-term cability, and so continuing so rapidly, with every business
For example, when TCS investments are now actively to invest in research areas increasingly becoming a
launched iON,11 a hosted applied to TCS’ business; for seemingly unrelated to the technology business, across
application platform for the example, marrying materi- current business pays off in industry verticals.
SMB segment in India, a als science research with initially unforeseen ways. To conclude, we submit
runtime configurable multi- machine learning to develop For example, research in the role of research for
tenant architecture6 devel- digital twins to optimize genome-based early predic- players in the technology
oped by the architecture industrial operations,5 and tion of rare diseases4 later services industry, such as
research group served as its with knowledge ontologies to enabled TCS’ business TCS, is to act as the bridge
initial technology base. complete the manufacturing to build genome analysis between fundamental scien-
The 2000s saw a rapid ‘digital thread.’2 Or combin- pipelines for pharma cus- tific advances (in computing
growth in infrastructure ing metadata abstractions tomers. Deep expertise in and beyond) and transfor-
management services, from our earlier software computational chemistry3 mative business ideas and
where TCS would man- engineering tools with deep is now allowing TCS to de- product innovation for large
age entire datacenters and reinforcement learning to sign new chemical formu- enterprises that form their
networks for large custom- optimize supply chains and lations and molecules for customer base.
ers. This was traditionally a drive personalization in cus- customers, a very different
people-intensive exercise, tomer interactions.1 Further, kind of service that could References
1. Barat, S. et al. Actor-based simulation
but also turned out to be TCS eats its own dog food—it potentially expand the very for closed loop control of supply chain
an excellent target for re- has deployed a home-grown scope of its core business using reinforcement learning. In
Proceedings of the 18th Intern. Conf.
searchers to apply a variety enterprise social media in the future. Autonomous Agents and MultiAgent
Systems. Intern. Foundation for
of AI/ML techniques for au- platform7 and more recently Second, translating Autonomous Agents and Multiagent
tomation. After many years a deep-learning-based con- research results into busi- Systems, 2019, 1802–1804.
2. Gautham, B.P., Reddy, S., Das, P. and
of field deployment ‘under versational system across all ness outcomes not only Malhotra, C. Facilitating ICME through
the radar,’ IGNIO,10 an AI- its 400K+ employees.c requires careful shepherd- platformization. In Proceedings of
the 4th World Congress on Integrated
driven enterprise automa- There have been many ing, but also for research to Computational Materials Engineering.
tion product developed in lessons learned along this evolve along with the busi- Springer, Cham, 2017, 93–102.
3. Gupta, R., Dwadasi, B.S., Rai, B. and
TCS Research, was formally ness of the company: Just Mitragotri, S. Effect of chemical
launched in 2015. c https://on.wsj.com/2R4xflP as IBM Research had to permeation enhancers on skin
permeability: In silico screening using
develop ‘services science’d molecular dynamics simulations.
to support its emerging Scientific Reports 9, 1 (2019), 1456.
TCS Research Areas Timeline 4. Punwani, D. et al. Multisystem
global services business, anomalies in severe combined
1981 so too has TCS Research immunodeficiency with mutant
˲˲ Physical Sciences
Establishment of TRDD, Pune; BCL11B. New England J. Medicine 375,
research focus on computational evolved, from initially 22 (2016), 2165–2176.
˲˲ Software Systems and Services engineering to support Tata group. using computer science 5. Runkana, Venkataramana. Model-
based optimization of industrial
1983 → 1990s to accelerate software gas-solid reactors. KONA Powder and
˲˲ Life Sciences Automation of migration to code Particle J. 32 (2015): 115–130.
generation tools; creation of
development and IT tasks, 6. Shroff, G., Agarwal, P. and Devanbu,
MasterCraft.™ to now applying a com- P. Instant multi-tier Web applications
˲˲ Embedded Systems and Robotics without tears. In Proceedings of the
2000s bination of both AI and 2nd India Software Engineering Conf.
˲˲ Cybersecurity and Privacy
Expansion of research beyond Pune domain research to trans- ACM, 2009, 3–12.
to several new centers and focus 7. Singh, M. et al. KNADIA: Enterprise
areas with TCS business expansion form product engineering, KNowledge assisted DIAlogue
˲˲ Computing Systems beyond software development into operations, and business systems using deep learning. In
engineering and infrastructure Proceedings of the IEEE 34th Intern.
services, cloud computing, and
models for its customers. Conf. Data Engineering. IEEE, 2018,
˲˲ Behavioral, Social, and Business process outsourcing … Last but not least, nurturing 1423-1434.
8. TCS Annual Report 2018–2019;
Sciences 2008 a vibrant research environ- https://on.tcs.com/2Ivq8Py
The growing importance of data, ment within a large and 9. TCS BαNCS; https://www.tcs.com/bancs
emergence of big data, maturing 10. TCS Ignio; https://www.digitate.com
˲˲ Data and Decision Sciences often independently 11. TCS iON; https://www.tcsion.com
of AI and deep learning for
enterprise applications … successful core business 12. TCS MasterCraft; https://mastercraft.
tcs.com
˲˲ Deep Learning and AI becomes a challenge in
2015
Launch of Ignio™ itself. For many years, Gautam Shroff (gautam.shroff@tcs.com)
˲˲ Computing Foundations heads TCS Research and is based in Delhi,
2018 research was viewed as a India.
Creation of a theory group, focus
˲˲ Media and Advertising on ad-tech, media, and so on. luxury indulged in because
K. Ananth Krishnan (ananth.krishnan@
one could afford it, that is, the tcs.com) is Chief Technology Officer for
… the evolution continues …
foresight of investing in TCS and is based in Chennai, India.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 63
big trends
DOI:10.1145/ 3347863
ahead in global companies. Increas-
BY PANKAJ JALOTE AND PARI NATARAJAN ingly, leaders of more than a thousand
global enterprises across the U.S., Eu-
The Growth
rope, and other locations have realized
India’s potential and have set up their
own IT or R&D centers to take advan-
tage of the vibrant Indian software
and Evolution
ecosystem.
The current wave of Indian software
entrepreneurs is focusing on build-
ing platforms and products for Indian
of India’s
and global markets. This has led to the
creation of more than 7,000 tech start-
ups in India. India is already home to
18 unicorns (start-ups valued in excess
of US$1 billion), and another 10 are ex-
Software
pected to be added by the end of 2020.
The Indian software industry has
accelerated the adoption of digital
technologies in the country. The
Industry
industry has played a crucial role in
providing digital identities to over
one billion people in the country,
which is further enabling the provi-
sion of services across industries such
as banking, healthcare, and educa-
tion in an efficient manner. The next
generation of Indian software com-
panies is helping millions of small
and medium businesses (SMBs) and
individual workers such as cab drivers
T H E DEVELOPM E NT O F the Indian software industry and delivery personnel move into the
is an archetype of how economic liberalization formal economy.
This article is not just a story of
combined with an entrepreneurial spirit can build the Indian software industry but also
an industry that today contributes as much as 8% to of the entrepreneurial capability of
the GDP of a fast-growing country like India. On the India’s vast talent pool.
help companies manage their legacy large-scale integration)/system de- structure, running the services, pro-
portfolio of applications and infra- sign—also took root, diversifying their viding IT strategy, and other related
structure. The first wave of the global services portfolio. services.
Internet and dot-com era created ˲˲ Circa 2000–2010: The rise of Indi- Global multinational companies
intercontinental Internet infrastruc- an software multinationals and R&D also realized India’s potential in
ture. Indian companies were able to centers. software services and started increas-
leverage this infrastructure to deliver With experience in dealing with ing their direct presence in India
software development-related services complex IT systems and confidence in by setting up IT, business process
to global enterprises remotely. working with international customers, management (BPM), and R&D centers.
Realizing the potential and the several companies became multina- To date, 1,250 companies from around
IMAGE F RO M SH UTT ERSTOCK.CO M
availability of talent, some multina- tionals with offices and centers across the world have set up their own
tional corporations established their countries. They offered a wider range centers in India across almost all key
own offshore development centers of services like executing large and industry verticals. Software/Internet,
in India. Companies involved in the complex projects involving integra- telecom, semiconductor, automotive,
software aspects of hardware—for tion, complete end-to-end solutions and industrial are the top industries
example, design of tools or VLSI (very including management of IT infra- present, with R&D being a strong focal
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 65
india region big trends
Figure 1. Growth of IT services, GICs, and tech start-ups in India. start-ups were established in just the
last year.
There are largely two types of
technology start-ups. The first are
consumer-led and largely focused on
the India market. Initially these were
replicas of U.S. companies, but soon
morphed with unique innovations for
the India market. For example, the
cash on delivery model in e-commerce
was pioneered in India and is now
used globally. The second set of start-
ups are focused on serving the U.S. and
European markets.
In the last few years, 18 start-ups
touched US$1 billion in market capital-
ization. Walmart bought India’s largest
e-commerce company, Flipkart, which
is only about 11 years old, at a valua-
tion of US$21 billion. OYO Rooms, a
Figure 2. The rise of unicorns in India. technology-enabled franchise model
hotel chain, was started by a 20-year-
old, and now has the largest number
of rooms under management in India,
overtaking both traditional Indian and
global hotel chains.
Start-ups are driving innovation at
an accelerated pace. To maintain the
warp speed of innovation, large com-
panies are building partnerships with
the start-ups and are actively looking
at acquisitions, both for talent and
intellectual property.
Companies have also set up pro- tions, thereby improving SMBs’ ability (whose chief conceptualizer and first
cesses to hire, train, and engage thou- to invest and grow their businesses. CEO, Nandan Nilekani, is a product
sands of employees. In fact, Indian IT Digital infrastructure. Within a of the IT industry), and others.
services companies spend over US$1.6 span of about a decade, Indian IT India’s global perception. Finally,
billion a year on employee training. companies have taken several ser- it should be noted the software sector
Large technology companies have vices being provided to citizens and has perhaps played the most crucial
set up campuses exclusively focused corporations and moved them online. role in changing the global percep-
on training their employees on skills Most of these systems have been tion of India. Until the 1980s, India
relevant to their global customers. developed by indigenous IT compa- was perceived as a poor country that
Over 500,000 engineers in India are nies, and many are also maintained needed support from more developed
already equipped with relevant digital and managed by them. Examples nations. Today, this view has changed,
skills to drive digital transforma- include the Ministry of Corporate and India now has a seat at the global
tion. FutureSkills, an initiative of the Affairs system for corporate tax filing, table. The world is aware of India’s
National Association of Software and the income tax management system, technology prowess and is actively
Services Companies (NASSCOM), has including e-filing of tax returns, the looking to make investments, form
an ambitious goal of training another entire India Stack digital infrastruc- partnerships, and tap India’s bustling
two million people in digital technolo- ture, the Goods and Services Tax sys- technology ecosystem. Frequent for-
gies over the next few years. tem, the passport system, the Indian eign travelers can attest to the fact that
The extensive engineering educa- rail reservation system (that books the quality of interaction with local
tion system and the deployed talent over 200 million tickets annually), the people has evolved dramatically over
pool in the IT industry are also helping Aadhaar unique identification infra- the last quarter century due to the IT
improve the digital capabilities of structure—the largest in the world industry’s widespread impact.
Indian enterprises. The technical and
managerial talent from IT companies Figure 3. Local impact of Indian software sector.
have moved to Indian enterprises
to help them accelerate their digital
transformation initiatives.
Female empowerment. The indus-
try has been supportive of women in
the workforce, an aspect where India
has traditionally lagged. Some 30% of
the IT sector workforce is comprised
of women employees and this has
been a trend since the early stages of
its development. The sector has not
only helped empower women but has
also provided them with highly aspira-
tional career options.
Start-up ecosystem. The start-up
ecosystem in India attracted over
US$10 billion in investments from
venture capitalists from across the
world between 2016 and 2018. Figure 4. Global impact of Indian software sector.
US$6 billion has already been invested
in Indian start-ups by SoftBank out of
its US$100 billion Vision Fund.
Start-ups such as Flipkart, Ola, and
Swiggy have helped create or digitally
enable millions of jobs such as cab
drivers and e-commerce/food delivery
professionals. These companies are
also empowering the country’s 60
million small and medium businesses
by digitally enabling their operations.
Start-ups such as Power2SME and
CapitalFloat are offering innovative
financial services for SMBs, including
“flow-based lending;” a lending model
that provides credit to SMBs based on
an analysis of their financial transac-
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 67
india region big trends
regions like the United States, Mexico, A strong process and continuous im- creation of many companies with
China, Europe, and Australia, employ- provement focus is a catalyst for both valuation in excess of US$1 billion
ing thousands of software engineers quality and productivity. across India.
and managers. The Indian IT services Industry collaboration. One key The government is expected to
industry is estimated to employ a total factor in the success of the Indian accelerate the creation of public
of 40,000+ locals in the United States IT sector has been its ability to bring digital infrastructure to streamline
alone. Additionally, the industry is companies together to develop an existing citizen services and create
also exporting its massive talent train- industry. The sector has created a huge new services. The India Stack model
ing infrastructure to global locations. collaboration ecosystem in the form will be expanded to create industry-
Tata Consultancy Services has set up a of an industry body—NASSCOM. The specific initiatives in areas such as
training hub in Cincinnati and is cre- association has helped develop best healthcare, supply chain, and educa-
ating a pipeline of graduates coming practices that get disseminated to tion. This will result in technology
out of U.S. universities. companies across the sector. getting weaved into the fabric of the
The worldwide impact of the Indian Scale and entrepreneurship. Wide- Indian workforce across agriculture,
software industry is widely evident: spread industry effort to promote, healthcare, education, and other
˲˲ Code written by Indians is present cultivate, and celebrate entrepreneur- industries. Millions of digitally en-
in almost all systems with software, in- ship has created an ecosystem for abled jobs and job categories will be
cluding cars, consumer electronics, en- entrepreneurs to conceptualize, fund, created in the process.
terprise software solutions, industrial and scale IT companies. The first Further, second- and third-tier
products, banking systems, and more; generation of entrepreneurs focused locations will join India’s software
˲˲ Indian designers are involved in on building and scaling IT services ecosystem due to the strong mobility
most chip and system designs by major companies, while the second is focus- network, education, and digital infra-
multinationals; ing on building IT products and IP-led structure built over the last decade.
˲˲ Indian IT firms have development services companies. Global companies, Indian IT compa-
centers in over 80 countries around the nies, and start-ups will leverage these
world; and The Way Forward cities to drive innovation.
˲˲ Over 1,000 companies develop The Indian IT sector is in a unique Over the last three decades, India
global products from their centers in position to lead global technology in- has risen as a technology and software
India. novations over the next decade. trailblazer, and with concerted efforts
The Indian education infrastruc- by the entire ecosystem including
Key Lessons ture is being rapidly overhauled by Indian IT companies, multinationals,
Observing the growth and impact of dedicated government initiatives. start-ups, and the government, India
the Indian IT industry provides a set of The government has announced the has the potential to further establish
valuable lessons that can be replicated creation of 17 new Indian Institutes of its standing as a world leader in the
for the development of other indus- Technology (IIT) across India to fur- software sector.
tries in India and other countries. ther improve the quality of engineer-
Government involvement. Minimal ing education. IITs have some of the Pankaj Jalote (jalot@iiitd.ac.in) is Distinguished Professor
(and founding director 2008–2018) at Indraprastha Institute
government interference coupled country’s best engineering faculty and of Information Technology (IIIT), New Delhi, India.
with supportive incentive policies education infrastructure. The millions Pari Natarajan (pari@zinnov.com) is the Chief Operating
was a key success factor. The Indian of engineers who will graduate in the Officer of Zinnov, Bangalore, India.
government did not regulate the in- next few years will be adept at machine
dustry and created tax incentives for learning, cloud computing, and other
Suggested Reading
both importing technology and for new-age digital technologies. As a re- NASSCOM, Strategic Review: IT-BPM
revenue from exports. sult India will continue to be a source Sector in India 2019: Decoding Digital
Skilling and development. Focus of skilled digital talent and intellectual
NASSCOM, Future Skills—A NASSCOM
on skills and talent development has property for more than 2,000 global Initiative
been instrumental in the growth of the enterprises.
NASSCOM, Women ‘in’Equality—Not
IT industry. Even in their early stages, India’s per capita income is Anymore!
IT companies spent significant time expected to cross US$3,500 by 2025
NASSCOM-Zinnov, GCC 3.0: Spotlight on
and money developing the skills of from the current US$2,000. This will Digital, Partnerships, New Delivery Models
their employees ahead of time. This increase discretionary spending by & Future Skills, 2019.
helped companies rapidly address the the population, creating a huge con-
NASSCOM-Zinnov, Indian Tech Start-Up
changing technology needs of their sumer market, potentially triggering Ecosystem 2018: Approaching Escape
global customers. the next wave of digital entrepreneurs Velocity
Process orientation. The heightened building India-focused technology. Press Information Bureau, Government of
focus on process orientation in the Venture capital activities will increase India, Ministry of Electronics & I (May 3,
Indian software industry has un- due to massive domestic opportuni- 2017), Employment Prospects in India’s IT
doubtedly contributed to its meteoric ties and the ability of Indian start-ups Sector: Robust Outlook
growth and has also enhanced the to build global products. The avail- Copyright held by authors/owners.
perception of the industry globally. ability of capital will catalyze the Publications rights licensed to ACM.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 69
big trends
DOI:10.1145/ 3343456
These are social and commercial needs,
BY PUSHPAK BHATTACHARYYA, HEMA MURTHY, whose servicing requires user interac-
SURANGIKA RANATHUNGA, AND RANJIVA MUNASINGH tion and information dissemination
in languages other than English. Only
Indic
around 10% of India’s population, or
about 125 million people, can speak
English; only about half that number
is comfortable reading and writing in
Language
that language. The social media activity
of the youth of the Indian subcontinent
(where 65% of the population is below
the age of 35) generates a huge amount
Computing
of e-content, much of which is in text
form, is multilingual, and even code-
mixed (text in multiple languages at the
same time, often in Roman script). The
numbers are mind-boggling:c
˲˲ 462.1 million Internet users (34% of
the population; the global average is 53%).
˲˲ 430.3 million users access the In-
ternet via mobile devices (79% of total
Web traffic).
˲˲ 250 million social media users
(19% of the population; the global aver-
I N APRIL 2019,following the Easter Sunday bomb age is 42%).
˲˲ 260 million WhatsApp users, and
attacks, the Government of Sri Lanka had to shut 53 million Instagram users.
down Facebook and YouTube for nine days to stop Sri Lanka alone has seven million
the spreading of hate speech and false news, posted Internet users (2018 data), which
equates to a penetration of 32%.
mainly in the local languages Sinhala and Tamil. There is no doubt that speech and
This came about simply because these social media natural language processing (NLP) of
Indic languages is hugely important
platforms did not have the capability to detect and and relevant, and has the potential to
warn about the provocative content. influence the lives and activity of at
India’s Ministry of Human Resource Development least 20% of the world’s population.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 71
india region big trends
Diversity is the name of the game for Indic-language computing; shown here are scripts in Other languages offer very little
Devanagari, Brahmi, Odia, Tamil, Telugu, Malayalam, and Sinhala, among other languages. language data. For example, available
parallel corpora for Sinhala-Tamil are
well below 50,000 sentences. Even raw,
clean corpora are of great value for
language computing. Modern-day deep
learning techniques start with word
embeddings (WEs). WEs are learned
from huge amounts of corpora (mil-
lions of words) that capture the context
distribution for words and phrases.
Such distribution captures semantics,
which is an elusive entity, computation-
ally speaking. Many Indic languages
do not have a processable clean corpus
from word lists, WEs, and a rich lexi-
con can be built. Another application
area that is affected by paucity of data
is ASR-TTS. Spoken signals must be
correct, with proper text units. Then
Austro-Asiatic (Khasi in Meghalaya, language boundary could be an impor- there are transcriptions of spoken
and Munda in Chhotonagpur). These tant cue for semantics (assuming the utterances that need to be accurate.
language families each have their own lexicon accounts for the vocabulary of Although there are subtitled You-
linguistic characteristics, whose rich- both languages). Also, Indian language Tube videos and lectures, they require
ness and complexity have been delved words are included in an English sen- curation, as time alignments are quite
into in multiple scholarly treatises.11 tence, where gerundification (such as poor. However, the number of available
These complexities, along with techno- “I’m chalaaoing a car,” meaning “I am hours of training data is small, leading
human constraints, give rise to the driving a car”) of Indian-language nouns to poor alignments.
challenges of Indic language comput- is common. In TTS, producing code- Absence of basic speech and NLP
ing, some of which are described here. switched systems requires the prosodic tools. The NLP pipeline starts with
Scale and diversity. For Indic lan- characteristics of the language and the word-level processing, and goes all the
guages, solutions must be simultane- speaker are preserved, especially when way up to discourse computation (con-
ously proposed for multiple languages. code switching involves stress-timed necting many sentences together with
There are 22 major languages in India, and syllable-timed languages. The attention to coherence and cohesion).2
written in 13 different scripts, with interplay between languages in terms The tools used at each stage of this
over 720 dialects. There is a need to de- of prosody needs to be understood to pipeline are affected by the accuracy
velop approaches that are generic, and make the sentences sound natural. of tools in the preceding stages. For
scaling to multiple languages should Resource scarcity. Indic-language English, since many groups across
be only a task of adaptation. As the lan- computing is bogged down by paucity the world have worked on the compu-
guages are quite different, there is a lot of data. Language computing these days tational processing of the language,
of effort required to arrive at common is primarily data-driven, with sophis- a staged development of NLP tools
solutions. Although E2E (end-to-end) ticated machine learning techniques of English occurred. NLTK,d a GATE-
is the buzzword today, use of multiple employed on the data. The success of likee NLP framework came into being,
scripts for Indian languages makes these approaches depends crucially on paving the way for large application
systems complex (as illustrated in the the availability of large amounts of high- development in English. In contrast,
accompanying figure). quality data. We take the example from even basic morphology analyzers that
Long utterances. Indian-language automatic machine translation (MT), split words into their roots and suffixes
utterances are much longer in duration which is highly data-driven these days: do not exist for most Indic languages,
compared to English, and hardly con- the Hansard corpus for English-French and even if they exist, their accuracy
tain punctuation. A typical English sen- contains 1.6 billion words; the Europarl level is low.
tence has about 70 characters, while a Parallel Corpus for 21 European lan- Absence of linguistics knowledge.
sentence in an Indian language typically guages contains about 30 million words; Though speech processing and NLP
averages 130 characters. E2E systems WMT 15 data for English-Czeck contains are data-driven, linguistics insight and
perform poorly with long sentences. about 16 million parallel sentences; understanding of language phenom-
Code mixing. Code mixing is the use and WMT 14 data for English-German ena often help solve the problem of ac-
of more than one language in text/utter- contains about 4.5 million parallel curacy saturation. Deep understanding
ance. Handling code switching from one sentences. An Indic-language example of language phenomena helps design
language to another in both automatic with comparable size is the CFILT-IITB
speech recognition (ASR) and text to English-Hindi corpus, which includes d https://www.nltk.org/
speech (TTS) is a challenge. In ASR, the 800,000 parallel sentences. e https://gate.ac.uk/
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 73
india region big trends
and language technologies. This will Evaluation (FIRE) initiativet has taken up
help pool low resources across various the cause of evaluation in information
languages to build robust ASR systems retrieval and allied tasks. A FIRE-like
for Indian languages. initiative is needed for all areas of ILT.
˲˲ In the context of TTS, the major
issue to be addressed is the input Conclusion
method. Text is available in multiple Indic Language Computing (ILC) is
Indian scripts, but digital resources in too important a problem to be lying in
terms of high-quality parallel corpora oblivion. Given spectacular advance-
are few and far between. In the context ments to date in computing science
of both ASR and TTS, generic acoustic and technology, Internet, AI, machine Code mixing
models across various languages, gener-
ic language models in the former, and a
learning, and NLP, the time is ripe for
a concerted thrust for realization and
must find ways
generic Indic voice in the latter need to social penetration of ILC. The energy of to preserve
be designed. This will also address the
issue of code switching.
the start-up echo system has to be har-
nessed with government support, and the speaker’s voice
˲˲ In TTS, code mixing must find ways
to preserve the speaker’s voice across
guidance from academia. Language re-
source creation is a precondition for ILC
across languages.
languages. Further, the influence of the revolution, and as in all cases of large Further,
native tongue on a non-native tongue
must be preserved. For instance, there
infrastructure building (roads, internet,
gas lines, waterways), government spon-
the influence
are as many varieties of English as there sorship is needed for resource building. of the native tongue
are native tongues. Replacing non-
native English (which is syllable-timed)
t http://fire.irsi.res.in/fire/2019/home on a non-native
with stress-timed English can make it References
1. Bahdanau, D., Cho, K. and Bengio, Y. Neural machine
tongue must
difficult for the listener to understand.
˲˲ Text in social media generally
translation by jointly learning to align and translate.
ICLR, 2015. be preserved.
2. Bhattacharyya, P. Natural language processing:
includes code switching/mixing. Fur- A perspective from computation in presence of
ther, there are many words that have ambiguity, resource constraint and multilinguality. CSI
J. Computer Science and Engineering 1, 2 (2012).
a local cultural connotation. Building 3. Chakrabarti, D., Mandalia, H., Priya, R., Sarma, V., and
language resources to address these re- Bhattacharyya, P. Hindi compound verbs and their
automatic extraction. In Proceedings of Computational
quires the expertise of linguists, speech Linguistics, Manchester, U.K., Aug. 2008.
4. Jha, G.N. The TDIL program and the Indian language
scientists, natural language processing corpora initiative. In Proceedings of LREC, 2010.
engineers, and ethnographers. 5. Koehn, P. Europarl: A parallel corpus for statistical
˲˲ Data is the new oil, and NLP and machine translation. In Proceedings of the Machine
Translation Summit, 2005.
ILT is no exception. There is no doubt 6. Kunchukuttan, A., Mishra, A., Chatterjee, R., Shah, R. and
Bhattacharyya, P. Shata-Anuvadak: Tackling multiway
that resources with quality and cover- translation of Indian languages. In Proceedings of the
age need to be created, and created fast. Language Resources Evaluation Conference, 2014.
7. Kunchukuttan, A., Mehta, P., and Bhattacharyya, P.
Thinking creatively on how to engage The IIT Bombay English-Hindi parallel corpus. In
even a small portion of 1 billion hands Proceedings of LREC, (Miyazaki, Japan, May 7–12, 2018).
8. Liu. B. Sentiment Analysis and Opinion Mining. Morgan
for resource creation is a must. Crowd- and Claypool Publishers, 2012.
sourcing, in spite of its criticism with 9. Murthy, R., Kunchukuttan, A., and Bhattacharyya, P.
Addressing word-order divergence in multilingual
respect to quality, seems to be the way neural machine translation for extremely low resource
forward. Providing attractive, helpful languages. In Proceedings of LREC, 2019.
10. Ranathunga, S., Farhath, F., Thayasivam, U., Jayasena,
interfaces and remuneration can go a S., and Dias, G. Si-Ta: Machine translation of Sinhala
long way toward resource creation. In and Tamil official documents. In Proceedings of the
National Information Technology Conference, 2019.
this context, the Language Data Consor- 11. Subbarao K.V. South Asian Languages—A Syntactic
Typology. Cambridge, 2012.
tium for Indian Languages (LDC-IL)p
initiative of Central Institute of Indian
Pushpak Bhattacharyya (pb@cse.iitb.ac.in) is a professor
Languages (CIIL) is noteworthy. in the computer science and engineering department of IIT
˲˲ Evaluation is the key to actual use of Bombay, and director of IIT Patna.
language resources and should be taken Hema Murthy (hema@cse.iitm.ac.in) is a professor in
the computer science and engineering department of
very seriously. Like TRECq (USA), CLEFr IIT Madras.
(Europe), and NTCIRs (CJK countries), Surangika Ranathunga (surangika@cse.mrt.ac.lk) is a
India’s Forum for Information Retrieval senior lecturer in the department of computer science and
engineering and a member of the faculty of engineering at
the University of Moratuwa.
Ranjiva Munasinghe (ranjiva@mindlanka.org) is chief
p http://www.ldcil.org/ executive officer of MIND Analytics and Management in
q https://trec.nist.gov/ Colombo, Sri Lanka.
r http://www.clef-initiative.eu/
s http://research.nii.ac.jp/ntcir/index-en.html © 2019 ACM 0001-0782/19/11 $15.00
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 75
big trends
DOI:10.1145/ 3355625
incredibly young—about 50% of its
BY VIVEK RAGHAVAN, SANJAY JAIN, population is below the age of 25, with
AND PRAMOD VARMA approximately 65% of the population
below the age of 35.e India expects to
India Stack—
have 100 million people entering the
workforce over the next 10 years. In
short, the country is young, ambitious,
and connected.
Digital
Social welfare is delivered through a
complex network of over 950 schemes
and funds by the Union government
alone. The Union government spent
close to $45 billion on subsidies last
Infrastructure
year. The states would cumulatively
spend another $10 billion. A migrant
worker population of over 453 mil-
lion people,f moves from their homes
as Public Good
either seasonally or permanently, add-
ing to the complexity of welfare service
delivery.
In 1985, the Prime Minister of In-
dia said that out of every rupee spent
by the central government, only 15
paise (15%) reaches the beneficiary.
This is because, distribution of
welfare has typically taken place in
kind, through a multi-layered supply
I N DI A IS HOM Eto almost one-fifth of the world’s chain. Realistically, it is estimated
that leakages in welfare programs
population. Its scale and diversity rival those of spanned from 10% to 60%, depend-
continents, not countries. India has “official” 22 ing on the program.
Moreover, price subsidies tend to
languages but unofficially 19,500 languages have been be regressive because they are untar-
recognized as having 10,000 speakers or more.a There geted. As the economic survey of 2015
is incredible diversity, but also incredible disparity. defined it, “a rich household benefits
more from the subsidy than a poor
About 45 million people still live in extreme poverty,b household.”g The report found the bot-
and less than 4% of its 1.3 billion people paid any tom 50 of the country consumed less
than 25% of the subsidized LPG (cook-
income tax at all.c ing gas). Similarly, 41% of the kerosene
At the same time, digital inclusion in India has taken supplied through the public distribu-
off in a significant way in the last few years. It has tion system was lost to “leakages,”
and only 46% of the remainder went to
1.2 billion mobile connections and over 500 million poor households.
Internet users.d India is now the world’s second- It is important to remember that
largest market for smartphones, with an estimated 400 some of these problems and numbers
are as recent as 2015. Clearly, the state
million smartphones in India having access to one of needed to move away from the price
the cheapest mobile data plans in the world. India is subsidy model to a more targeted
a Census of India, 2011; http://bit.ly/2Sysodk e Age structure and marital status, Census of
b World Poverty Clock Statistics on India https://worldpoverty.io/ India, 2011.
c Two crore Indians file returns but pay zero income tax. f Census 2011 Data; http://bit.ly/2Y9nrdj
Economic Times, Oct. 23, 2018; http://bit.ly/2wWziiU g Economic Survey of India, 2014–15, Chapter 3:
d IAMAI I-CUBE 2019 Report, http://bit.ly/2MQELCF Wiping every tear from every eye.
A young woman applies for an Aarhaar card, the world’s largest biometrics ID system.
and efficient service delivery model. of infrastructure. While the technolo- multiple efforts by multiple teams.
Starting in 2009, India began to create gies themselves are commendable, Each API or standard may have an
digital infrastructure to move from their real “disruptive” power has been owner and their own licensing nu-
people and paper-intensive inefficient what applications they enable. For ances. It is a set of loosely coupled
service delivery, to an efficient, direct, example, the Internet may have been technologies and protocols, and
digital service delivery. born of a specific need, but its success there is no master directive. Each
This was not just the need of the is because of its design. It was a mass- technology tries to do one thing and
State, the Indian markets felt the scale, open, and interoperable pro- do it well. The innovation comes
same way. Despite its large size, and tocol. The use cases for the Internet from the combinatorial use of these
consistently high growth rates, the were not restricted by the imagination technologies by entrepreneurs and
Indian markets have not turned out of its founders. governments alike.
to be stellar for many players. The The India Stack is a name given to What they do have in common is
high cost of customer acquisition, a family of APIs, open standards, and that each lowers the cost of doing
KYC (Know-Your-Customer) pro- infrastructure components that allow transactions. The reason for cost
cess, various claims verification, and a user in India to demand services savings is multifold—it eliminates
overall cost of business meant market digitally. As of 2019, the services the paper, but also eliminates the need
players could not provide affordable India Stack offers are proving identity, for physical presence during a trans-
and accessible products or services. completing KYC, making digital pay- action. Digital payments eliminate
PHOTO BY MELTING SP OT/SHU TTERSTOCK.COM
A large population was not in the ments, signing documents digitally cash and the cost of cash handling.
formal economy. This is the context and sharing of data. While the list It can also simplify compliance, such
that—beginning in 2009—over the of APIs is growing, the APIs listed in as in the case of KYC compliance for
next 10 years led to the creation of the Table 1 are now mature, well under- financial or telecom institutions. It
India Stack. stood, and enable efficient delivery of could reduce “leakages” through the
services in India. verification of identity and elimina-
India Stack Why India Stack? Just like the tion of duplicates.
Leapfrogs. There have been various modern Web, the India Stack did not The breadth of India Stack and its
technologies that have played the role come out of one place, but through potential use cases are too wide to
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 77
india region big trends
open architecture, the Aadhaar project points for India’s 1,300 million Interface is a protocol that simplifies
was able to develop a vibrant and open people,k indicating many features the sending and receiving of value
vendor base for critical components of card-based payment systems (for from any stored-value account to any
of the hardware and software running example, high cost of payments and other stored-value account. That is,
Aadhaar. The project is deployed on cumbersome user experiences) were the UPI specifications allowed sending
commodity computing resources to not effective at reaching most of the money from bank accounts to bank
prevent costly maintenance bills. Fur- Indian market. The National Payments accounts, but also from bank accounts
ther, scaling to hundreds of millions of Corporation of India (NPCI) realized to mobile wallets and loyalty accounts,
transactions per month and billions of that for digital payments to be suc- among others.
records has not been a problem. This cessful in India, it needed a low-cost UPI provides a set of interoper-
has led to massive cost savings, with payments system that worked for high able APIs that innovators use to build
each enrollment ultimately costing volumes of low-value transactions. payment apps or make payments as a
less than $1 per successful enrollment The outcome of NPCI’s delib- feature into their current workflows.
and authentications to approximately erations was the Unified Payments Normally, this would have required
one cent per authentication.j Interface (UPI). The Unified Payments bilateral agreements with all banks,
Using identity. Aadhaar is a digital but since almost all the banks in India
identity, and its value is derived from k Bankwise CARD Statistics, RBI; https://rbi. use the UPI specifications for transfer-
the fact that to confirm the user who org.in/Scripts/ATMView.aspx ring money between bank accounts
furnished the ID is indeed the true
owner of that identity Aadhaar provides Figure 1. India Stack’s design layers.
multiple channels for authentication.
This allows governments and business-
es to trust the person they are transact-
ing with is truly who they claim to be.
Aadhaar serves as foundational
identity and does not collect informa-
tion on purpose of authentication.
It has been envisioned that many
domain-specific federated identities
will be derived from Aadhaar. For
example, India’s tax ID—the Perma-
nent Account Number (PAN)—uses
Aadhaar to deduplicate its registers.
Since these two databases remain
separate, the CIDR has no informa-
tion on the tax IDs of its users. This
principle is also reflected in the insti-
tutional design of the program—the
Unique Identity Authority of India
(UIDAI)—which is a separate agency
that does not fall under an existing Figure 2. India Stack’s open architecture.
function-specific ministry.
Aadhaar was aware of the growing
privacy risks if identity and transaction
data is collected in one central place.
Hence, Aadhaar envisioned a feder-
ated model during use of Aadhaar.
Payments
Despite having credit cards for more
than 40 years, their penetration in
India has been very low. In 2015, there
were only approximately 20 million
credit cards in the country and two
million digital payment acceptance
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 79
india region big trends
party model that is so important to vent replay attacks. Residents can lock declared Aadhaar did not intrinsically
UPI’s success. In UPI, the payment-ad- and unlock (for short periods of time) violate an individual’s fundamental
dress-issuing entity is not necessarily their biometrics using the multiple right to privacy, but its mandated use
the same as the one providing the un- channels such as the Aadhaar mobile ought to be restricted only to govern-
derlying bank account. This means a application or the Web portal. ment-provided subsidies and benefits,
user can use any app to send or receive Aadhaar introduced temporary vir- tax collection, and other proportional
money directly from their bank ac- tual IDs that allowed users to mask their use cases where permitted by law.n
count. They are no longer restricted to Aadhaar numbers during an authenti- While it may seem contentious
just the app provided by their banking cation request. The means the Aadhaar and politically charged, such con-
service provider. This has increased number does not need to be shared versations are a feature, not a bug,
competitiveness to acquire users, and with an authenticating agency. In the of democracy. The executive, judi-
as a result the responsiveness and per- digitally signed response, Aadhaar re- ciary, and UIDAI were responsive to
formance of bank apps has improved turns agency-specific UID tokens, which the public’s needs and evolved the
dramatically since the launch of UPI. are unique and cannot be correlated system based on what the people
With over 800 million transactions across agencies. In addition, residents wanted. Our experience underscores
worth more than US$1.9 billion being can lock their Aadhaar number and the importance of stakeholder
transacted monthly after approximate- authenticate using only the virtual ID. conversations during the design and
ly two years,l the Unified Payments Aadhaar has introduced the con- implementation of the program.
Interface (UPI) is the fastest-growing cept of offline KYC verification, which
open-loop digital payments platform allows residents to directly share their Conclusion
in the world. digitally signed KYC information India’s experience with creating
with a verification agency XML/QR digital infrastructure platforms as
Criticism and Evolution code formats. This allows residents public goods offers multiple lessons
As Aadhaar gained coverage, traction, to share non-tamperable credentials learned in technology, system, and
and the trust of service providers as a without direct involvement of the regulatory architecture. It demon-
unique and robust proof of identity, Aadhaar system. Local validation of strates how multiple such systems
it began to be requested (and some- the photograph through face match- can be leveraged in concert—such
times mandated) as a foundational ing and mobile number are possible. as the India Stack—for develop-
document across a variety of public Sensitive data such mobile number is ment objectives. Governments and
and private services, in particular, for stored using a one-way hash; the data businesses alike are building for
government subsidies, banking, and is revealed only if residents share the diverse use cases on top of the stack.
telecommunications. data with the verification agency. By lowering the transaction costs of
As a result, there was pushback from Problems with authentication using serving the poor, we are achieving
media, civil society, and academics fingerprints by manual laborers or better inclusion.
around issues of privacy and security of senior citizens were addressed through Such digital infrastructure is
individual data, and the possibility of the introduction of multiple biometric not a unique requirement in India.
exclusion from access to services due to modalities such as face and iris match- It is estimated that approximately
lack of an Aadhaar or due to authenti- ing. In addition, multiple modalities 161 countries currently have or are
cation errors. Meaningful engagement can be combined through fusion to building their own digital ID systems.
on all criticisms is not possible in this further reduce rejections in the field. Many countries have local interbank
article, the issues are wide ranging and Finally, exception processes are put in payment systems and are now look-
need detailed, nuanced discussions on place to ensure 100% of residents can ing to upgrade them for a mobile-
design trade-offs. authenticate using the Aadhaar system. first world. As various countries build
What we would like to highlight is Aadhaar’s open architecture meant their own systems, the Indian experi-
some of the outcomes from the cri- such a solution could be rolled out ence with Aadhar serves as a real-
tique of Aadhaar. The UIDAI was able quickly in response to public demand. world example to learn from. Even
to see the increasingly vocal demand The criticism and civil society if the systems may look different,
for better privacy controls, resulting movement also bought into the public we believe the principles adopted in
in design changes to the program as discourse India’s lack of a Data Privacy their development would serve well
it evolved. Aadhaar has rolled out a Law, which is necessary whether or not globally.
number of features to further enhance there is an Aadhaar. While trying the
the security, privacy, and inclusion of Aadhaar case, the judges were forced n Justice K.S. Puttaswamy (Retd) vs Union of India,
Aug 26, 2018.
the Aadhaar system. to ask if the constitution guarantees a
Biometric capture devices are regis- fundamental right to privacy. A nine-
Vivek Raghavan (vivek.raghavan68@gmail.com) is Chief
tered with the Aadhaar ecosystem and judge bench found the answer was Product Officer of UIDAI, Bangalore, India.
all biometrics captured are signed and affirmative.m Sanjay Jain (snjyjn@gmail.com) is Chief Innovation
encrypted at the capture device to pre- A second Supreme Court judgment Officer of CIIE, IIMA, Bangalore, India.
Pramod Varma (pramodkvarma@gmail.com) is Chief
Architect at UIDAI, Bangalore, India.
l UPI Product Statistics; https://www.npci.org. m Justice K.S. Puttaswamy (Retd) vs Union of India,
in/product-statistics/upi-product-statistics Aug. 24, 2017. © 2019 ACM 0001-0792/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 81
big trends
DOI:10.1145/ 3344432
eyeing Australia); an insurance
BY CHARLES ASSISI, AVINASH RAGHAVA, aggregator called PolicyBazaar; the
AND NS RAMNATH e-commerce site Paytm Mall; an
eyewear retailer called Lenskart;
The Rise
food technology aggregators such
as Swiggy and Zomato, and hotel-
room aggregators like OYO and
FabHotels.
of the Indian
Thousands of entrepreneurs
start up every year and aspire to
become one of the new unicorns.
Venture capitalists invested over
Start-Up
$20 billion on start-ups last year,
and evidence suggests they are
likely to invest much more by the
end of this year.
The rise of the Indian start-up
Ecosystem
ecosystem can be characterized by
three major changes over the last
decade:
1. A shift from models copy-past-
ed from elsewhere to the creation of
models built for India.
2. A move from the IT services
model to technology products.
3. A statement of intent from
entrepreneurs that the time for
Jugaad is over, and cutting-edge
innovations are where the future
lies. The Hindi word Jugaad roughly
translates as “to work around.”
WA L K INTO ANY one of the many start-up events The notion was a result of resource
organized across India, and inevitably the image of an constraints faced by a number of
Indian bazaar comes to mind: people rushing around, enterprising Indians, especially
those living in rural areas. Jugaad
shouting, bargaining, answering phones with great is a well-researched theme and has
excitement, laughing loudly, boasting, blushing, been extensively documented by
Navi Radjou,b a French-American
and generally being optimistic, as if they are at the scholar based in Silicon Valley, in
beginning of a rising trend of well-being. his book Frugal Innovation.
Such optimism might seem justified. According to These changes offer interesting
insights for the start-up ecosystem
data compiled by Fortune magazine,a from just eight in India and across the world.
‘unicorns’ in 2015, the number of start-ups in India
From ‘Copy-Paste’ Models
valued at more than $1 billion has grown to 26. What to Local Innovation
is interesting is that in 2018 alone, India added The current crop of Indian start-ups
eight unicorns to the club. trace their origins to the mid-1990s
and late 2000s. They were driven by
These include diverse entities such as Ola, started entrepreneurs and venture capital-
in India as a competitor to Uber and has since ists (VCs) from the U.S. (Silicon Val-
ley, in particular), or were heavily
expanded its footprint into the U.K. (and is
a http://bit.ly/2IHMrl3 b https://thinkers50.com/biographies/navi-radjou/
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 83
india region big trends
and make interesting innovations the country with $2,000 per capita cated, and if you take 25 years to do
based on domestic markets. to $20,000 per capita, it is really a it, then they will be adults and you
This realization led a range of major challenge. You have to fix can’t do it. The only way to square
start-ups that targeted specific seg- basic things like health, education, the circle is by using AI (artificial in-
ments across Indias 1, 2, and 3. and access to financial services. The telligence) and the cloud to deliver
Banking and finance have always classical way of doing that would personalized health, education, and
been early adopters of technology, have been to say ‘let’s have more finance to a billion people. That
and financial technology start-ups doctors, let’s have more teachers,’ will drive the economy. For a coun-
began trying to solve some of the and so on. That’s certainly not pos- try that has low per-capita income,
problems that banks and financial sible in the timeframe that we have. to use this as a strategic tool ... is
institutions could not solve with the If you want to get everyone edu- very important.”
burden of brick-and-mortar infra-
structure. Sachin (left) and Binny Bansal. How India earns.
One of the early validations that ˲˲ 1 Indian earns 30% of the total and makes
using digital technologies could over Rs 1.5 lakh a month
help in financial inclusion came ˲˲ 14 Indians earn 30% of the total and make
from an experiment that IFMR around Rs 20,000 a month each
Trust (now Dvara Trust) did with its ˲˲ The next 30 Indians earn 30% of the total
KGFS model. For a long time, finan- and make Rs 8,000 per month each
cial inclusion meant micro credit. ˲˲ The poorest 55 Indians earn 10% of the total
However, the designers of the and make only Rs 1,500 per month each
Kshetriya Gramin Financial Ser- ˲˲ 1 Indian owns 53% of the wealth
vices (KGFS) model, which included ˲˲ 9 Indians own 23% of the wealth
Nachiket Mor, then heading up the
˲˲ 40 Indians own 20% of the wealth
ICICI Foundation, and Bindu An-
˲˲ 50 Indians at the bottom own only
ant, who was with venture capital
4.1% of the total
firm IFMR Trust, wanted to create
a system that didn’t just give credit,
but also provided a range of finan-
cial services, including savings India’s unicorns.
and insurance products. Savings
products were not yet being offered Name Value* Incorporated Industry Investors
to the poor, because the money BigBasket $1 6/5/19 E-commerce/ Alibaba Group, Bessemer Venture
marketplace Partners, Helion Venture Partners
they put in did not even cover the
Dream11 $1 9/4/19 Sports/gaming Kaalari Capital, Tencent Holdings,
paperwork needed to accept it. As Steadview Capital
an experiment, the KGFS design- Udaan $1 3/9/18 E-commerce DST Global, Lightspeed Venture Partners,
ers digitized the entire process in Microsoft ScaleUp
money market mutual funds, and PolicyBazaar $1 6/25/2018 Fintech Info Edge, Softbank Capital
found it worked. InMobi $1 2/12/14 Adtech Kleiner Perkins Caufield & Byers, Softbank
While KGFS could not scale up Corp., Sherpalo Ventures
some of its products, it showed Shopclues $1.1 12/1/16 eCommerce/ Nexus Venture Partners, GIC Special
fintech how going digital can help Marketplace Investments, Tiger Global Management
financial inclusion by bringing Swiggy $3.3 6/21/2018 On-demand Accel India, SAIF Partners, Norwest
Venture Partners
down transaction costs.
Hike $1.4 8/16/2016 Social Foxconn, Tiger Global management,
Fintech is one example of how Tencent
Indian start-ups, even as they Delhivery $1.6 2/27/2019 Supply Chain Times Internet, Nexus Venture Partners,
pursue growth and profits, also fill and Logistics SoftBank Group
the gaps that government, busi- ReNew Power $2 2/14/2017 Energy and Goldman Sachs, JERA, Asian Development
nesses, and the social sector either Utilities Bank
would not or could not in the past. Zomato $2.18 10/4/15 Social Sequoia Capital, VY Capital
By making technology work, many BYJU'S $5 7/25/2017 Ed Tech Tencent Holdings, Lightspeed India
start-ups today are aligned with Partners, Sequoia Capital India
broader societal goals. Oyo Rooms $4.3 9/25/2018 Travel Tech SoftBank Group, Sequoia Capital India,
Lightspeed India Partners
As Nandan Nilekani, chairman
Ola Cabs $6.2 10/27/2014 On-demand Accel Partners, SoftBank Group, Sequoia
of Infosys and former chairman of Capital
the Unique Identification Author- Snapdeal $7 5/21/2014 E-commerce SoftBankGroup, Blackrock, Alibaba Group
ity of India (UIDAI), said during a One97 $10 12/5/15 Fintech Intel Capital, Sapphire Ventures,
conversation with Microsoft CEO Communications Alibaba Group
Satya Nadella in Bangalore: “When * In billions as of press time.
you think of the challenge of taking
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 85
india region big trends
down upon initially, it is only now of asking deeper questions about oncology treatment and supporting
that investors are beginning to ap- design. more effective and efficient cancer
preciate why products matter in the Yet Jugaad has evolved in its own drug development. GreyOrange is a
long run. As the cost of labor start- way, and has come to mean frugal focused robotics warehouse man-
ed to rise over the years, margins innovation; coming up with solu- agement company. Julia Computing
from IT services started to come tions using minimal resources, for has developed a unique, high-per-
down, and competition emerged a market that could not afford ex- formance programming language
from other geographies where labor pensive products or solutions. Con- with rich applications in AI and
is cheaper. sumer products company Godrej machine learning capabilities.
Between entities such as Infosys, developed ChotuKool, a portable
Cognizant, TCS, Wipro, and some refrigerator that consumes minimal Lessons
mavericks who pushed the pedal to power, for villages facing continu- These three shifts offer two broad
the metal on products, start-ups in ing power outages. lessons to start-up ecosystems
India are looking at a contemporary Underlying the change from across the world:
narrative. One example of such a Jugaad to frugal innovation is the The starting point does not mat-
maverick is Zoho, which was found- belief that it is possible to build ter; the direction does.
ed in 1996. Zoho builds a range of world-class products with limited It does not matter where the
Web-based technology tools aimed resources. One of the insights of story starts, or where the motiva-
at improving the productivity of C.K. Prahalad is that innovations tion comes from. It could have been
businesses. Its founder, Sridhar made for the bottom of the pyramid in copy-pasting Western business
Vembu, on encountering the unfair often work for segments that oc- models, bidding for software cod-
conditions of venture capitalists, cupy the higher levels. ing services based on labor arbi-
swore to build a company without These three shifts—from the trage, or even Jugaad. What matters
going to VCs. Zoho’s revenue is esti- copy-paste model to local innova- is the evolution.
mated to be around $500 million. tion, from software outsourcing to Context matters; start-ups come
One interesting innovation of Zo- product development, from Jugaad to life in a society.
ho’s is its training program, which to frugal innovation—have given However, evolution seldom hap-
selects students (typically from way to a new breed of start-ups that pens on its own. Evolution happens
poor backgrounds) from schools are hugely aspirational. within a context; when entrepre-
and teaches them to code. Some Sharad Sharma, co-founder of neurs start solving a problem for
students from early instances of the the Indian Software Products In- the society in which they live, they
program are now product managers dustry Round Table (iSPIRT), makes experiment, scale up, and reach out
at Zoho. the distinction between mercenary to new customers.
Zoho is not the only enterprise start-ups whose primary goal is to
tech company that has made its make money, and missionary start- Charles Assisi is co-founder and director of Founding Fuel,
Mumbai, India. He is co-author (with NS Ramnath) of The
mark in the product space. Fresh- ups whose primary goal is to solve Aadhaar Effect: Why the World’s Largest Identity Project
works, started by a former employee impossible problems. Matters.
of Zoho, also focuses on the SMB An example of such a mission- Avinash Raghava is Community Platform Evangelist for
Accel, Bangalore, India.
market and has innovated on the ary start-up is TeamIndus, India’s
NS Ramnath is part of the founding team at
Inside Sales model. only entrant in the competition for Founding Fuel, Bengaluru, India, where he now serves
India today has more than 10 the Lunar X Prize, in which teams as a senior writer.
B2B companies with Unicorn are challenged to “land a robot on The views expressed here are the authors’ and do not
status; some reached this mile- the surface of the Moon, travel 500 necessarily reflect those of their employers.
stone in less than three years. meters over the lunar surface, and
Companies like BlackBuck, Udaan, send images and data back to the
Power2SME, Delhivery, and Capil- Earth.” TeamIndus did not win the
lary Technologies are trying to solve Lunar X Prize; no one did. But the
some of India’s problems. Deep point about start-ups such as Team-
technology security companies like Indus is their goal is not just to win
Druva, Qubole, and CloudCherry a prize, but to show it is possible to
are leveraging India as a base for aim high.
their development. TeamIndus is just one example
of a start-up that many would not
From Jugaad to Cutting Edge instinctively associate with India.
For years, Indian innovation was There are many start-ups in India
mostly associated with the word that work on cutting-edge technolo-
Jugaad. The flip side to this is that gies. Medical diagnostics start-up
it indicates short-term vision that SigTuple Labs uses AI to analyze
moves from one kind of ‘duct tape’ visual medical data, while Mitra
to fix a problem to another, instead Biotech is advancing personalized © 2019 ACM 0001-0792/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 87
big trends
DOI:10.1145/ 3343454
Highlights
of Software
R&D in India
and IEEE TSE. In fact, ICSE 2014 was software was filled with stories of across the world. Indian software
held in Hyderabad and POPL 2015 was delayed and poor-quality software companies took a lead in deploying
held in Mumbai under the General projects. For improving this sce- this model for improving software
Chairship of Pankaj Jalote and Sriram nario, a five-level Capability Maturity quality and productivity. In the early
PHOTO BY NOPPASIN WONGCHUM /SH UT T ERSTOCK. CO M
Rajamani, respectively. Model (CMM)b was developed by the stages of model deployment, a large
India also has its flagship annual Software Engineering Institute (SEI). percentage of the companies at high
conference called Innovations in CMM is a framework and model maturity levels (CMM level 5) were
Software Engineering (ISEC), which for evaluating and improving the from India—a situation that contin-
provides a platform for sharing experi- software development process in an ues even today. Companies across
ences of various research groups. organization in a staged manner, the world, and countries desirous
and has been adopted by companies of developing their software sector,
Indian Industry’s Leadership wanted to learn from the Indian expe-
in the Software Process b https://en.wikipedia.org/wiki/ rience in employing rigorous software
Up until the 1990s, the world of Capability_Maturity_Model processes using quantitative tech-
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 89
india region big trends
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 91
big trends
DOI:10.1145/ 3343449
Research Highlights
BY MEENA MAHAJAN, MADHAVAN MUKUND, Algorithms. Maximizing the flow that
AND NITIN SAXENA can be routed in a network is one of
the most well-studied algorithmic
Research in
problems, with immense practi-
cal applicability. In the 1970s, when
computer science research in India
was taking root, Sachin Maheshwari
Theoretical
and his co-authors V.M. Malhotra and
M. Pramodh Kumar devised a max-
flow algorithm that matched the best
bounds at that time, but was concep-
tually much simpler and hence ideal
Computer
for exposition.
Scheduling and facility location
problems are often cast as multi-
commodity flow problems and are
Science
NP-hard. Using ideas from flows and
linear programming, efficient ap-
proximation problems can be devised
in many settings. The Indian Institute
of Technology (IIT) Delhi is at the
forefront of international research in
this area.
Parameterized algorithms and
complexity is a relatively recent field
that focuses on multivariate analysis
of algorithm performance and the
development of algorithms for hard
problems where combinatorial explo-
sion is confined to specified param-
eters. This burgeoning field has a very
has been a vibrant
T H EOR ETICAL CO MPUT E R S C I E NC E close connection with India—the first
international event wholly devoted
part of computing research in India for the past 30 to this theme took place in Chennai
years. India has always had a strong mathematical in 1999—and has seen cutting-edge
tradition. One could also argue that in the 1980s and contributions from India, notably
from the Institute of Mathematical
1990s, theory offered a unique opportunity to keep Sciences, Chennai (IMSc) and Chennai
up with international research in computing despite Mathematical Institute (CMI).
Matchings in graphs come in many
limited access to state-of-the-art hardware. different flavors—perfect, maximum,
The Annual International Conference Foundations stable, popular. Indian researchers
of Software Technology and Theoretical Computer have made significant contributions
toward obtaining combinatorial
Science (FSTTCS) was launched in 1981. FSTTCS2 characterizations, devising new algo-
allowed Indian researchers a natural opportunity to rithms, and understanding the paral-
lel complexity of these problems.
interact with leading academics worldwide. Data structures are crucial to the
Another early impetus was funding for efficiency of many state-of-the-art algo-
international collaboration through agencies rithms. Indian researchers have been
part of the community designing data
such as the Indo-French Centre for Promotion of structures for static succinct represen-
Advanced Research (CEFIPRA). tations and for maintaining dynamic
The annual Foundations of Software Technology and Theoretical Computer Science (FSTTCS) Conference, organized by the Indian
Association for Research in Computer Science, is a premier forum for presenting original results in initial aspects of CS and software
technology. The images here show participants from FSTTCS 2018, held last December at India’s Ahmedabad University.
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 93
india region big trends
computing date back to the early Nitin Saxena is a professor at the Indian Institute of
Technology Kanpur, Kanpur, India.
1980s—a time that also saw the first
generation of graduate students from © 2019 ACM 0001-0792/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 95
big trends
DOI:10.1145/ 3345671
The Positive
and Negative
Effects of
Social Media
in India
THERE HAS BEEN a phenomenal increase in the use of
online social media (OSM) services in India, including
Facebook, Twitter, Instagram, LinkedIn, and YouTube.
In addition to these services, one-to-one messaging
services like WhatsApp have 200 million users, the
highest in the world. India has 462 million users
accessing the Internet, among these: Facebook has
in spammers and phishers, users are
250+ million users, LinkedIn 42+ million, and Twitter losing social skills, and more. Newness
23+ million users, and the majority of users access of technology/mobile phones, low-
these services through their mobile phones. literacy rates, and cheaper mobile data
rates are cited as negative impacts of
These services have had a profound impact in social media services on society.
India—overall digital literacy has increased, people Research has been mainly directed
toward regulation of content gener-
are more connected, dissemination of local language ated on OSM. It can be classified in the
content has increased, information exchanged during following categories:
crises is substantial, and more. The deep penetration ˲˲ Identifying topical interests and
expertise of the users in online be-
of social media services also has negative effects—the havior1,11 and efficiently matching the
propagation of false information and hate, an increase consumers and producers of content;
˲˲ Mining useful content from social ˲˲ Identifying bias in content recom- that clicking selfies in many cases can
media, for example, finding actionable mendation of news to users of social lead to accidents. Hence, another line
information from the OSM to help law- media;2 of research has been to accurately
enforcement agencies9 and relief and ˲˲ Impact of content on determining communicate to users risks involved
rescue teams during disaster;7,8 the dynamics of opinion over social with a location chosen for taking
PHOTO BY F ILIP JED RASZAK/ SH UTT ERSTO CK .COM
˲˲ Identifying harmful content, name- networks.3,4 selfies, as with the Saftie and Saftie
ly analyzing hate and spam content on Note that with the rising usage of Camera apps.b
YouTube and Twitter,5 and analyzing local and code-mixed (that is, local Research enumerated earlier
the spread of misinformation/fake language + English) languages in con- provides an overview of some of the
content on social media (TweetCred, tent generation, a lot of research is ongoing work in the area of social
Facebook Inspector, WhatsFarziaa); also directed toward mining in pres- media conducted by Indian scientists,
ence of such content.6 Selfies form but is by no means exhaustive. Here,
a http://precog.iiitd.edu.in/research/whatsapp- a substantial part of social media
misinformation/ image content and it has been found b http://labs.precog.iiitd.edu.in/killfie/
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 97
india region big trends
is a widespread practice of writing to have been killed while taking selfies,i unprecedented rise in use of local or
Indian languages using Roman script with India dominating these statistics code-mix languages; hence the need for
as well as mixing it with English dur- with 141 deaths. Given the increasing special attention from Indian research-
ing writing/speaking,g a phenomenon penetration of mobile technology, ers. Another diagonally opposite area
referred to as linguistic code-mixing high usage statistics, and the distur- of research would be to leverage social
or code-switching. For any analysis bances caused by such behavior, India media for social good; work on post-
of social media content from India, is one of the prime regions where disaster management as reported here;
correct processing of code-mixed text this problem is particularly relevant. and future scopes including utilizing
is an absolute necessity; however, Research conducted by Precog@IIIT social media content to devise better
traditional natural language process- Delhij identifies dangerous selfies. governance mechanisms, supporting
ing (NLP) modules such as language The researchers have created datasets, individuals/groups with health-related
identifiers, POS taggers, translators, classifiers, apps, and location-marker issues, and making quality education
and word aligners treat linguistic code- tools in this context. A convolutional accessible to the huge population by
switching data either as noise or as a neural network-based classifier to connecting teachers with students
new language (for example, Hinglish identify dangerous selfies posted on located in different places.
for Hindi-English code mixing). Both social media using only the image (no Acknowledgments. The authors
views are limited because the former metadata) gives an accuracy of 98%. thank Sunita Sarawagi, Abir De, and
does not recognize the complexity and The Saftie Camerak app based on the the anonymous reviewers for provid-
socio-pragmatics of the phenomenon, developed classifier works in real world ing constructive feedback.
whereas the latter does not utilize settings and detects and warns a user if
the fact that code mixing is a gram- the location is potentially dangerous. References
1. Bhattacharya, P. et al. Deep Twitter diving: Exploring
matically informed combination of two Important funding initiatives. topical groups in microblogs at scale. In Proceedings
languages. Further, bilingual speakers There has been a lot of funding ini- of the 17th ACM Conf. Computer Supported Cooperative
Work and Social Computing, 2014, 197–210.
show different language references tiatives both from government and 2. Chakraborty, A., Messias, J., Benevenuto, F., Ghosh,
S., Ganguly, N.and Gummadi, K.P. Who makes trends?
depending on the topic of discussion non-government agencies to popu- Understanding demographic biases in crowdsourced
and sentiment expressed. This implies larize social media research. Among recommendations. In Proceedings of the 11th Intern.
AAAI Conf. Web and Social Media, 2017.
that ignoring code-mixed patterns or those initiatives is the Indo-German 3. De, A., Bhattacharya, S.and Ganguly, N. Demarcating
conducting content-analysis only for Max Planck Center for Computer Endogenous and Exogenous Opinion Diffusion Process
on Social Networks. In Proceedings of the 2018 World
the predominant language over social Science—a five-year project on Un- Wide Web Conf., 2018, 549–558.
media (usually English) can lead to mis- derstanding, leveraging and deploying 4. De, A., Valera, I., Ganguly, N., Bhattacharya, S. and
Gomez-Rodriguez, M. Learning and forecasting opinion
leading conclusions, and are bound to online social networks, jointly funded dynamics in social networks. In Proceedings of the
miss out on social and discourse-level by the Indian Department of Science 30th Inter. Conf. Neural Information Processing
Systems, 2016, 397–405.
nuances in the data. Several research- and Technology and Max Planck Soci- 5. Maity, S.K., Chakraborty, A., Goyal, P. and Mukherjee,
ers from India have worked to address ety. Another initiative is the Media A. Opinion conflicts: An effective route to detect
incivility in Twitter. In Proc. ACM Hum.-Comput.
different aspects of code-switching; Lab Asia and Information Technol- Interact. Article 117 (2018), 117:1–117:27.
Microsoft Research India, under ogy Research Academy (ITRA)-funded 6. Pratapa, A., Bhat, G., Choudhury, M., Sitaram, S.,
Dandapat, S. and Bali, K. Language modeling for code-
project Melange,h has largely led the five-year project on Post disaster situ- mixing: The role of linguistic theory based synthetic
initiative. Several semi-supervised10 ation analysis and resource manage- data. In Proceedings of the 56th Annual Meeting of the
Assoc. Computational Linguistics, Vol.1. (Melbourne,
techniques to automatically produce ment, which patronized the research Australia, 2018), 1543–1553; https://www.aclweb.org/
anthology/P18- 1143
a large, annotated code-mixed dataset on investigating the role of social 7. Rudra, K., Ganguly, N., Goyal, P. and Ghosh, S.
are being developed to help the com- media for disaster management. Extracting and summarizing situational information
from Twitter social media during disasters. ACM
munity efficiently perform downstream Challenges. Presently, the world is Trans. Web 12, 3 (July 2018), 17:1–17:35.
supervised NLP tasks. witnessing several negative impacts of 8. Rudra, K., Goyal, P., Ganguly, N., Mitra, P. and Imran, M.
Identifying sub-events and summarizing disaster-
Killfies for social media. In recent OSMs. Hence, it is important for the related information from microblogs. In Proceedings
years, the posting of selfies (or digital computing world, with intense research of the 41st Intern. ACM SIGIR Conf. Research and
Development in Info. Retrieval, 2018, 265–274.
self-portraits) on social media websites input from scientists all over the world, 9. Sachdeva, N. and Kumaraguru, P. Call for service:
such as Facebook, Instagram, and to mitigate these impacts. The specific Characterizing and modeling police response to
serviceable requests on Facebook. In Proceedings of
Snapchat has become a part of main- problems are many—fake news, hate the ACM Conf. Computer-Supported Cooperative Work
stream culture. Often people portray speech, the shaming of individuals or and Social Computing, 2017.
10. Samanta, N., Nangi, S.R., Jagirdar, H., Ganguly, N.,
their adventurousness by posting groups. It is now clear that in the garb Charabarti, S. A deep generative model for code
dangerous selfies (aka killfies). Since of spontaneity, companies, political switched text. In Proceedings of IJCAI, 2019.
11. Zafar, M.B., Bhattacharya, P., Ganguly, N., Ghosh,
March 2014, 238 people are reported parties, and individuals are constantly S. and Gummadi, K.P. On the wisdom of experts
manipulating the systems to produce vs. crowds: Discovering trustworthy topical news
in microblogs. In Proceedings of the ACM Conf.
trending topics and thus control discus- Computer-Supported Cooperative Work and Social
Computing, 2016, 438–451.
g for example, a bilingual Hindi/English speak- sions on social media. The problems
er posts on Twitter: “aj patakhe to india me hi are compounded in India with the Niloy Ganguly (niloy@cse.iitkgp.ac.in) is a professor at IIT
phutenge, sure it would be,” where the itali- Kharagpur, India.
cized segment (“today fireworks will occur in In-
Ponnurangam Kumaraguru (pk@iiitd.ac.in) is an
dia only”) is in Hindi written in Roman script. i http://bit.ly/saftie- bot associate professor at IIIT Delhi, India.
h https://www.microsoft.com/en-us/research/ j http://precog.iiitd.edu.in/
project/melange/ k http://bit.ly/saftie-cam © 2019 ACM 0001-0792/19/11
N OV E MB E R 2 0 1 9 | VO L. 6 2 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 99
big trends
DOI:10.1145/ 3343452
literacy barriers prevent 26% of adults
BY ADITYA VASHISTHA, UMAR SAIF, in India and 42% of adults in Pakistan
AND AGHA ALI RAZA from using text-based interfaces. Most
South Asian languages and dialects
The
are still unsupported by the advance-
ments in natural language processing
ruling out the use of voice interfaces
like Siri and Alexa.
In light of these constraints,
Human-Computer Interaction for
Internet of
Development (HCI4D) researchers and
practitioners have used interactive
voice response (IVR) technology to cre-
ate voice-based services that overcome
connectivity barriers by using ordinary
phone calls, literacy barriers by using
the Orals
local language speaking and listening
skills, and socioeconomic barriers by
using toll-free (1-800) lines. These ser-
vices let users call a phone number to
record and listen to voice messages in
their local languages. Because of their
accessible and usable design, these
services have found applications in
diverse domains and have profoundly
impacted marginalized communities
in low-resource environments. This
article follows the evolution of these
services over the last two decades
social media, online
I N TER NET SER VI C E S LI K E (see the accompanying figure), and
their big challenges and new frontiers.
discussion forums, and crowdsourcing marketplaces
have transformed how people participate in the First Wave: Access and Inclusion
The first wave of voice-based services
information ecology and digital economy. These focused on improving information
services empower mostly urban, affluent, and literate access for people in low-resource
people, and improve their reach to information and communities. For example, Health-
Line enabled low-literate frontline
instrumental needs. However, these services currently health workers in Pakistan to retrieve
exclude billions of people worldwide who are too relevant information by speaking
out predefined commands.6 While
poor to afford Internet-enabled devices, too remote to initial efforts like HealthLine allowed
access the Internet, or too low literate to navigate the users to only consume information,
mostly text-driven Internet. subsequent services took the form of
voice forums and enabled marginal-
In India and Pakistan alone, there are nearly ized communities to also produce
1.1 billion people offline. Although 70% of their and share information. This included
Avaaj Otalo (an agriculture discus-
populations have access to mobile phones, most sion forum in India),3 CGNet Swara (a
people still use basic or feature phones, making it citizen journalism service in India),2
difficult to extend existing Internet services on these MobileVaani (a social media service
in India), Ila Dhageyso (a civic engage-
devices running custom operating systems. Even when ment portal in Somaliland),1 and
people can afford smartphones and the Internet, IBM’s Spoken Web (a user-generated
information directory in India). The by conducting lab-trainings as well development. In an initial test, 34,000
success of these initial services dem- as door-to-door field campaigns, Polly users listened to 728 job adver-
onstrated their great potential to en- but it was quickly realized that these tisements nearly 386,000 times within
able information access and connec- approaches were not scalable. Raza a year.
tivity among underserved populations et al. used a ludic design approach Over the last seven years, Polly has
in diverse HCI4D contexts. However, to train users and promote usability been successfully used in multiple
the vast majority of these services ran and spread. They built Polly, a voice- countries to rapidly spread useful
into the hurdles of user training and based entertainment service that lets information to underserved popu-
technology adoption. users make a short audio recording, lations. In 2014, at the peak of the
apply funny voice modifications to Ebola crisis in West Africa, Polly-San-
Second Wave: Training and Spread it, and share it with their friends via té (Polly-Health) was deployed as an
Nearly a decade ago, the biggest road- automated voice calls.5 They deployed emergency disaster-response service
blocks to designing voice forums were Polly to five low-income people in in Guinea to spread reliable informa-
usability, motivation, and spread; Pakistan in early 2012. Within a year, tion about prevention, symptoms,
target populations faced difficulties Polly spread virally to over 165,000 and cure of Ebola.12 The informa-
in using even the simplest of speech- users via 636,000 calls without any tion originated from the Centers for
based telephone interfaces, they did outreach efforts. Polly’s ludic inter- Disease Control and the service was
not exhibit interest or trust in using fact design trained users to navigate funded by the U.S. Embassy in Cona-
such services, and it was difficult to IVR interfaces, and also led to its viral kry. One of the hurdles to informa-
advertise and spread such services to adoption. Raza et al. then used Polly tion dissemination in the Guinean
underconnected people. Research- to share instrumental information context is great linguistic diversity
ers tried to overcome these barriers with users to aid their socioeconomic and the lack of a widely understood
Three waves of voice forums in low-resource environments. platform like Facebook might be inef-
fective for voice forums, and vice versa.
This presents interesting research
2 Training and Spread challenges of identifying indecorous
• Polly
• Polly-Santé
content in local language audio, filter-
ing out spreaders of disinformation,
and addressing situations where the
collective ignorance of community
2007 2011 2015 2019
members eclipse their collective intel-
ligence. The HCI4D community must
tackle these grand challenges to make
the Internet of the orals more diverse,
1 Access and Inclusion 3 Managing Content and Costs inclusive, and impactful.
• HealthLine • MobileVaani • Sangeet Swara
• Avaaj Otalo • Spoken Web • Baang References
• CGNet Swara • Ila Dhageyso • Respeak and similar systems 1. Gulaid, M. and Vashistha, A. Ila Dhageyso: An
interactive voice forum to foster transparent
governance in Somaliland. In Proceedings of the
6th Intern. Conf. Information and Communications
Technologies and Development: Notes, Vol. 2 (Cape
Town, South Africa, 2013), 41–44.
2. Mudliar, P. et al. Emergent practices around CGNet
tion and get community support. to reduce random speech recognition Swara, voice forum for citizen journalism in rural
Moreover, they demonstrated that a errors. It then pays users in mobile India. In Proceedings of the 5th Intern. Conf.
Information and Communication Technologies and
community of low-income, low-liter- airtime based on the accuracy of Development (Atlanta, GA, USA, 2012), 159–168.
3. Patel, N. et al. Avaaj Otalo: A field study of an
ate people can moderate themselves transcripts submitted in them. In the interactive voice forum for small farmers in rural
without any outside support, thereby last three years, Respeak has been India. In Proceedings of the SIGCHI Conf. Human
Factors in Computing Systems (Atlanta, GA, USA,
addressing the content management used by low-income students, blind 2010), 733–742.
challenge of these voice forums. people, and rural residents in India to 4. Raza, A.A. et al. Baang: A viral speech-based social
platform for under-connected populations. In
The second key challenge in scal- produce speech transcriptions with Proceedings of the 2018 CHI Conf. Human Factors
ing voice forums is the airtime cost. over 90% accuracy at one-fourth of in Computing Systems (Montreal, QC, Canada, 2018),
643:1–643:12.
Often, these services use expensive the market rate, generating sufficient 5. Raza, A.A. et al. Job opportunities through
toll-free lines to remain accessible profit to subsidize their participation entertainment: Virally spread speech-based services
for low-literate users. In Proceedings of the SIGCHI
to low-income users. The resultant costs. One minute of crowd work on Conf. Human Factors in Computing Systems (Paris,
cost poses a huge burden to sustain- Respeak enable users to earn eight France, 2013), 2803–2812.
6. Sherwani, J. et al. Healthline: Speech-based access to
ability, often putting these services minutes of airtime.8 health information by low-literate users. Inter. Conf.
at risk of being shut down as the Information and Communication Technologies and
Development (Bangalore, India, 2007), 1–9.
usage grows. While a few services Grand Challenges: Harassment, 7. Vashistha, A. et al. BSpeak: An accessible voice-based
sustain themselves through adver- Misinformation, and Disinformation crowdsourcing marketplace for low-income blind
people. In Proceedings of the 2018 CHI Conf. Human
tisements, grants, and partnerships Voice forums, like any other social Factors in Computing Systems (Montreal, QC, Canada,
2018), 57:1–57:13.
with telecoms or governments, these platform, come with their own pitfalls. 8. Vashistha, A. et al. ReCall: Crowdsourcing on basic
options are often beyond the reach of They end up reflecting the existing phones to financially sustain voice forums. In
Proceedings of the 2019 CHI Conf. Human Factors in
most voice forum providers. To make sociocultural norms and values of the Computing Systems (Glasgow, Scotland, U.K., 2019).
these services financially sustainable, society, including its shortcomings 9. Vashistha, A. et al. Respeak: A voice-based, crowd-
powered speech transcription system. In Proceedings
Vashistha et al. examined whether and biases. For example, while Swara of the 2017 CHI Conf. Human Factors in Computing
low-income users of voice forums and Baang served as instruments of Systems (Denver, CO, USA, 2017), 1855–1866.
10. Vashistha, A. et al. Sangeet Swara: A community-
could complete useful work on their inclusion for low literate, rural, indig- moderated voice forum in rural India. In Proceedings
mobile phones to offset their partici- enous, and visually impaired commu- of the 33rd Annual ACM Conf. Human Factors in
Computing Systems (Seoul, South Korea, 2015),
pation costs. In 2016, they created nities, they failed to create a welcom- 417–426.
Respeak, the first voice-based crowd- ing environment for female users.11 11. Vashistha, A. et al. Threats, abuses, flirting, and
blackmail: Gender inequity in social media voice
sourcing marketplace that pays users Women faced systemic discrimination forums. In Proceedings of the 2019 CHI Conf. Human
to transcribe audio files vocally.7–9 and harassment in the form of mes- Factors in Computing Systems (Glasgow, Scotland,
U.K., 2019).
Respeak sends short audio segments sages that contained abuses, threats, 12. Wolfe, N. et al. Rapid development of public health
to multiple voice forum users and and flirtatious behavior. education systems in low-literacy multilingual
environments: Combating Ebola through voice
pays them via mobile airtime for Both mainstream social media messaging. In Proceedings of the ISCA Special
Interest Group on Speech and Language Technology in
each submitted transcript. Instead of platforms and voice forums face grand Education (Leipzig, Germany, 2015).
typing the transcript, users respeak challenges when tackling misinforma-
audio content into an off-the-shelf tion, disinformation, harassment, and Aditya Vashistha is an assistant professor at Cornell
speech recognition engine and abuse. These platforms and forums University, Ithaca, NY, USA.
submit the autogenerated transcript. differ greatly in terms of scale, fea- Umar Saif is UNESCO Chair, ICTD, Lahore, Pakistan.
Respeak combines the transcripts for tures, interfaces, supported languages, Agha Ali Raza is an assistant professor at Information
Technology University, Lahore, Pakistan.
each segment from multiple users us- and target users. Consequently, solu-
ing sequence-alignment algorithms tions to tackle these challenges on a © 2019 ACM 0001-0792/19/11
The Effects
differences, combining human and
machine intelligence produces better
outcomes than when each works sepa-
rately. People also view this form of col-
of Mixing
laboration between humans and ma-
chines as a possible way to mitigate the
problems of bias in machine learning,
a problem that has taken center stage
Machine
in recent months.12
We decided to investigate this type
of collaboration between humans and
machines using risk-assessment algo-
Learning
rithms as a case study. In particular,
we looked at the Correctional Offender
Management Profiling for Alternative
Sanctions (COMPAS) algorithm, a well-
known (perhaps infamous) risk-predic-
and Human
tion system, and its effect on human
decisions about risk. Many state courts
use algorithms such as COMPAS to pre-
dict defendants’ risk of recidivism, and
Judgment
these results inform bail, sentencing,
and parole decisions.
Prior work on risk-assessment algo-
rithms has focused on their accuracy
and fairness, but it has not addressed
their interactions with human deci-
sion makers who serve as the final ar-
bitrators. In one study from 2018, Julia
Dressel and Hany Farid compared risk
assessments from the COMPAS soft-
ware and Amazon Mechanical Turk
Deep Blue software beat the World
I N 1 997, IB M ’S
workers, and found that the algorithm
Chess Champion Garry Kasparov in a series of six and the humans achieved similar
matches. Since then, other programs have beaten levels of accuracy and fairness.6 This
study signals an important shift in the
human players in games ranging from “Jeopardy!” to literature on risk-assessment instru-
Go. Inspired by his loss, Kasparov decided in 2005 to ments by incorporating human sub-
jects to contextualize the accuracy and
test the success of Human+AI pairs in an online chess fairness of the algorithms. Dressel and
tournament.2 He found the Human+AI team bested Farid’s study, however, divorces the
the solo human. More surprisingly, he also found human decision makers and the algo-
rithm when, in fact, the current model
the Human+AI team bested the solo computer, even indicates that humans and algorithms
though the machine outperformed humans. would work in tandem.
demonstrates that algorithmic risk pact of the COMPAS algorithm on hu- and criminal history. These profiles
scores act as anchors that induce a man judgments concerning the risk described real people arrested in Bro-
cognitive bias: If we change the risk of recidivism. COMPAS risk scores ward County, FL, based on information
prediction made by the algorithm, par- were used because of the data avail- from the dataset that ProPublica used
ticipants assimilate their predictions able on that system, its widespread in its analysis of risk-assessment algo-
to the algorithm’s score. usage in prior work about algorithmic rithms.1 While this dataset originally
The results highlight potential fairness, and the use of the system in contained 7,214 entries, this study ap-
shortcomings with the existing hu- numerous states. plied the following filters before sam-
man-in-the-loop frameworks. On the Methods. The experiment entailed pling for 40 profiles that were present-
one hand, when algorithms and hu- a 1 x 3 between-subjects design with ed to participants:
mans make sufficiently similar deci- the following treatments: control, in ˲˲ Limit to black and white defendants.
sions their collaboration does not which participants see only the defen- Prior work on the accuracy and fair-
achieve improved outcomes. On the dant profiles; score, in which partici- ness of the COMPAS algorithm lim-
other hand, when algorithms fail, hu- pants see the defendant profiles and its their analyses to white and black
defendants.3,4,6 To compare the results From this filtered dataset 40 de- [JUVENILE-FELONY COUNT] juvenile
from this experiment with those in pri- fendants were randomly sampled. A felony charges and [JUVENILE-MISDE-
or studies, this study considers only the profile was generated containing in- MEANOR COUNT] juvenile misdemeanor
subset of defendants who identify as formation about the demographics, charges on their record.
either African-American (black) or Cau- alleged crime, criminal history, and The descriptive paragraph in the
casian (white). algorithmic risk assessment for each score treatment added the following
˲˲ Exclude cannabis crimes. Interest- of the defendants in the sample. The information:
ingly, the pilot study showed partici- descriptive paragraph in the control COMPAS is risk-assessment software
pant confusion about cannabis-related treatment assumed the following for- that uses machine learning to predict
crimes such as possession, purchase, mat, which built upon that used in whether a defendant will commit a crime
and delivery. In the free-response section Dressel and Farid’s study:6 within the next two years. The COMPAS
of the survey, participants made com- The defendant is a [RACE] [SEX] aged risk score for this defendant is [SCORE
ments such as “Cannabis is fully legal [AGE]. They have been charged with: NUMBER]: [SCORE LEVEL].
here.” To avoid confusion about the le- [CRIME CHARGE]. This crime is clas- Finally, the descriptive paragraph in
gality of cannabis in various states, this sified as a [CRIMINAL DEGREE]. They the disclaimer treatment provided the
study excludes defendants charged with have been convicted of [NON-JUVENILE following information below the COM-
crimes containing the term cannabis. PRIOR COUNT] prior crimes. They have PAS score, which mirrored the lan-
guage the Wisconsin Supreme Court
Figure 1. Defendant profile from score treatment. recommended in State v Loomis:18
Some studies of COMPAS risk-as-
sessment scores have raised questions
about whether they disproportionately
classify minority offenders as having a
higher risk of recidivism.
Upon seeing each profile, partici-
pants were asked to provide their own
risk-assessment scores for the defen-
dant and indicate if they believed the
defendant would commit another
crime within two years. Using drop-
down menus, they answered the ques-
tions shown in Figure 1.
We deployed the task remotely
through the Qualtrics platform and re-
cruited 225 respondents through Ama-
zon Mechanical Turk, 75 for each treat-
ment group. All workers could view the
task title, “Predicting Crime;” task de-
scription, “Answer a survey about pre-
Figure 2. Accuracy rate in treatment groups. dicting crime;” and the key words asso-
ciated with the task, “survey, research,
and criminal justice.” Only workers
60%
54%
living in the U.S. could complete the
54%
51% task, and they could do so only once.
50% During the pilot study among an initial
test group of five individuals, the sur-
vey required an average of 15 minutes
40%
Accuracy (Overall)
The error bars represent the 95% con- ent ways: some ignored them, some re- Farid’s study demonstrates the strik-
fidence intervals. The results suggest lied heavily on them, some used them ing similarity between recidivism pre-
the provision of COMPAS scores did as starting points, and others used dictions by Mechanical Turk workers
not significantly affect the overall ac- them as sources of validation. and the COMPAS algorithm.6 This sim-
curacy of human predictions of recidi- Figure 3 has excerpts of participant ilarity may preclude the possibility of
vism. In this experiment, the overall responses with a summary of answers complementarity. Our study reinforces
accuracy of predictions in the control to the free-response question: How this similarity, indicating the combina-
treatment (54.2%) did not significantly did you incorporate the COMPAS risk tion of human and algorithm is slightly
vary from those in the score treatment scores into your decisions? (although not statistically significantly)
(51.0%) (p = 0.1460). Discussion. When assessing the worse than the algorithm alone and
The inclusion of a written advise- risk that a defendant will recidivate, similar to the human alone.
ment about the limitations of the COM- the COMPAS algorithm achieves a sig- Moreover, this study shows that the
PAS algorithm did not significantly af- nificantly higher accuracy rate than accuracy of participant predictions
fect the accuracy of human predictions participants who assess defendant pro- of recidivism does not significantly
of recidivism, either. Participants in files (65.0% vs. 54.2%). The results from change when a written advisement
the disclaimer treatment achieved an this experiment, however, suggest that about the appropriate usages of the
average overall accuracy rate of 53.5%, merely providing humans with algo- COMPAS algorithm is included. The
whereas those in the score condition rithms that outperform them in terms Wisconsin Supreme Court mandated
achieved 51.0%; a two-sided t-test in- of accuracy does not necessarily lead the inclusion of an advisement without
dicated this difference was not statisti- to better outcomes. When participants indicating that its effect on officials’
cally significant (p = 0.1492). incorporated the algorithm’s risk score decision-making was tested.11 Psychol-
Upon the conclusion of the task into their decision-making process, ogy research and survey-design litera-
block in the exit survey, 99% of par- the accuracy rate of their predictions ture indicate that people often skim
ticipants responded that they found did not significantly change. The inclu- over such disclaimers, so they do not
the instructions for the task clear, sion of a written advisement providing perform their intended purpose.10 In
and 99% found the task satisfying. In information about potential biases in concurrence with such theories, the
their feedback, participants indicated the algorithm did not affect participant results here suggest that written ad-
they had positive experiences with the accuracy, either. visements accompanying algorithmic
study, leaving comments such as: “I Given research in complementary outputs may not affect the accuracy of
thoroughly enjoyed this task;” “It was computing that shows coupling human decisions in a significant way.
a good length and good payment;” and and machine intelligence improves
“Very good task.” their performance,2,9,11 this finding Experiment Two: Algorithms
Participants did not mention the seems counterintuitive. Yet successful as Anchors
advisement when asked how they took instances of human and machine col- The first experiment suggested that
the COMPAS scores into account. Rath- laboration occur under circumstances COMPAS risk scores do not impact hu-
er, their responses demonstrated that in which humans and machines dis- man risk assessments, but research in
they used the COMPAS scores in differ- play different strengths. Dressel and psychology implies that algorithmic
COMPAS Disclaimer
Ignore “I tried not to look at them after awhile, because I felt some were off “I thought it was fairly random, so I didn’t invest
(lol) but I still took them into account somewhat. I mostly went with much faith in it.”
my gut and opinions, though.”
“Generally I just ignored it and made my own guess.”
Rely Heavily “I kept my scores within 2 points of the COMPAS score.” “I relied on it, it eliminates bias.”
Starting Point “I used that as a baseline.“ “It was only a starting point. I paid more attention
to the criminal charge, prior charges and guessing
on whether the defendant would be convicted.”
susceptible to forms of cognitive bias When predicting the risk that a de- Accountability in Algorithmic
such as anchoring.7,15 fendant will recidivate, the COMPAS Decision-Making
These findings also, importantly, algorithm achieved a significantly Nicholas Diakopoulos
highlight problems with existing higher accuracy rate than the partici- https://queue.acm.org/detail.cfm?id=2886105
frameworks to address machine pants who assessed defendant pro-
References
bias. For example, many research- files (65.0% vs. 54.2%). Yet when par- 1. Angwin, J., Larson, J. Machine bias. ProPublica
ers advocate for putting a “human in ticipants incorporated the algorithm’s (May 23, 2016).
2. Case, N. How to become a centaur. J. Design and
the loop” to act in a supervisory ca- risk assessments into their decisions, Science (Jan. 2018).
pacity, and they claim this measure their accuracy did not improve. The 3. Chouldechova, A. Fair prediction with disparate
impact: A study of bias in recidivism prediction
will improve accuracy and, in the experiment also evaluated the effect of instruments. Big Data 5, 2 (2017), 153–163.
context of risk assessments, “ensure presenting an advisement designed to 4. Corbett-Davies, S., Pierson, E., Feller, A., Goel, S. and
Huq, A. Algorithmic decision making and the cost of
a sentence is just and reasonable.”12 warn of the potential for disparate im- fairness. In Proceedings of the 23rd ACM SIGKDD
Intern. Conf. Knowledge Discovery and Data Mining.
Even when humans make the final pact on minorities. The findings sug- ACM Press, 2017, 797–806.
decisions, however, the machine- gest, however, that the advisement did 5. Critcher, C.R. and Gilovich, T. Incidental environmental
anchors. J. Behavioral Decision Making 21, 3 (2008),
learning models exert influence by not significantly impact the accuracy of 241–251.
anchoring these decisions. An algo- recidivism predictions. 6. Dressel, J. and Farid, H. The accuracy, fairness, and
limits of predicting recidivism. Science Advances 4, 1
rithm’s output still shapes the ulti- Moreover, researchers have in- (2018), eaao5580.
mate treatment for defendants. creasingly devoted attention to the 7. Englich, B., Mussweiler, T. and Strack, F. Playing dice
with criminal sentences: the influence of irrelevant
The subtle influence of algorithms fairness of risk-assessment software. anchors on experts’ judicial decision making.
via this type of cognitive bias may While many people acknowledge Personality and Social Psychology Bulletin 32, 2
(2006), 188–200.
extend to other domains such as fi- the potential for algorithmic bias in 8. Furnham, A. and Boo, H.C. A literature review of the
nance, hiring, and medicine. Future these tools, they contend that leav- anchoring effect. The J. Socio-Economics 40, 1 (2011),
35–42.
work should, no doubt, focus on the ing a human in the loop can ensure 9. Goldstein, I.M., Lawrence, J. and Miner, A.S. Human-
collaborative potential of humans fair treatment for defendants. The machine collaboration in cancer and beyond: The
Centaur Care Model. JAMA Oncology 3, 10 (2017), 1303.
and machines, as well as steps to pro- results from the second experiment, 10. Green, K.C. and Armstrong, J.S. Evidence on the
mote algorithmic fairness. But this however, indicate that the algorith- effects of mandatory disclaimers in advertising. J.
Public Policy & Marketing 31, 2 (2012), 293–304.
work must consider the susceptibil- mic risk scores acted as anchors that 11. Horvitz, E. and Paek, T. Complementary computing:
policies for transferring callers from dialog systems to
ity of humans when developing mea- induced a cognitive bias: Partici- human receptionists. User Modeling and User-Adapted
sures to address the shortcomings of pants assimilated their predictions Interaction 17, 1-2 (2007), 159–182.
12. Johnson, R.C. Overcoming AI bias with AI fairness.
machine learning models. to the algorithm’s score. Participants Commun. ACM (Dec. 6, 2018).
who viewed the set of low-risk scores 13. Jukier, R. Inside the judicial mind: exploring judicial
methodology in the mixed legal system of Quebec.
Conclusion provided risk scores, on average, European J. Comparative Law and Governance
The COMPAS algorithm was used here 42.3% lower than participants who (Feb. 2014).
14. Kahneman, D. Thinking, Fast and Slow. Farrar, Straus
as a case study to investigate the role viewed the high-risk scores when as- and Giroux, 2011.
of algorithmic risk assessments in hu- sessing the same set of defendants. 15. Mussweiler, T. and Strack, F. Numeric judgments
under uncertainty: the role of knowledge in anchoring.
man decision-making. Prior work on Given this human susceptibility, an J. Experimental Social Psychology 36, 5 (2000),
the COMPAS algorithm and similar inaccurate algorithm may still result 495–518.
16. Northcraft, G.B. and Neale, M.A. Experts, amateurs, and
risk-assessment instruments focused in erroneous decisions. real estate: an anchoring-and-adjustment perspective
on the technical aspects of the tools by Considered in tandem, these find- on property pricing decisions. Organizational Behavior
and Human Decision Processes 39, 1 (1987), 84–97.
presenting methods to improve their ings indicate that collaboration be- 17. Shaw, A.D., Horton, J.J. and Chen, D.L. Designing
accuracy and theorizing frameworks tween humans and machines does incentives for inexpert human raters. In Proceedings
of the ACM Conf. Computer-supported Cooperative
to evaluate the fairness of their predic- not necessarily lead to better out- Work. ACM Press, 2011, 275–284.
18. State v Loomis, 2016.
tions. The research has not considered comes, and human supervision does 19. Tversky, A. and Kahneman, D. Judgment under
the practical function of the algorithm not sufficiently address problems uncertainty: Heuristics and biases. Science 185, 4157
(1974), 1124–1131.
as a decision-making aid rather than as when algorithms err or demonstrate 20. Wansink, B., Kent, R.J. and Hoch, S.J. An anchoring
a decision maker. concerning biases. If machines are and adjustment model of purchase quantity decisions.
J. Marketing Research 35, 1 (1998), 71.
Based on the theoretical findings to improve outcomes in the criminal
from the existing literature, some justice system and beyond, future re-
policymakers and software engineers search must further investigate their Michelle Vaccaro received a bachelor’s degree in
computer science in 2019 from Harvard College,
contend that algorithmic risk assess- practical role: an input to human de- Cambridge, MA, USA.
ments such as the COMPAS software cision makers. Jim Waldo is a Gordon McKay Professor of the practice of
can alleviate the incarceration epi- computer science at Harvard University, Cambridge, MA,
USA, where he is also a professor of technology policy at
demic and the occurrence of violent the Harvard Kennedy School. Prior to joining Harvard, he
Related articles
crimes by informing and improving on queue.acm.org
spent more than 30 years in the industry, much of that at
Sun Microsystems.
decisions about policing, treatment,
The Mythos of Model Interpretability
and sentencing.
Zachary C. Lipton
The first experiment described here https://queue.acm.org/detail.cfm?id=3241340
thus explored how the COMPAS algo- The API Performance Contract
rithm affects accuracy in a controlled Robert F. Sproull and Jim Waldo Copyright held by authors/owners.
environment with human subjects. https://queue.acm.org/detail.cfm?id=2576968 Publications rights licensed to ACM.
Article development led by
queue.acm.org
Write
Amplification
vs. Read
Perspiration
indexing we did, the more our ability to dom place. Linear searches for your
update became slower than molasses. wallet might be tractable in a small
I learned this is a common trade-off. apartment but not so much when the
Reading fast frequently means writing search space gets bigger in a larger
slow.
Row-store vs. column-store. I have Search home in the suburbs. To reduce the
read perspiration, LSM trees invest en-
focused most of my misspent career on
distributed systems and online trans-
makes reading ergy to organize the data by rewriting
it as you go.
action processing (OLTP)-style data- the documents When a new file is freshly writ-
bases. It’s natural for me to associate
high-performance updates with what
a lot easier. ten from the storage engine, it has a
bunch of key-value pairs. To make it
today is called a row-store. It dramatically easy to find keys, these are merged
Another approach is to organize
data by columns: Take a bunch of rows
lowers the read with files that were written earlier.
Each LSM tree has some form of fan-
and organize the data by its column perspiration. out where lower levels of the tree (with
values. Every row containing the state keys written earlier) are kept across
of California, for example, keeps just more files. For example, you may have
the single column’s data together. Co- 10 times as many files at level 1 as
lumnar databases are super fast for do- at the brand-new level 0. Each file at
ing queries because many logical rows level 1 has approximately one-tenth
with the same value are physically close as large a key range represented but
to each other. approximately 10 times the amount
However, updating a column-store of update time represented. Similarly,
is not as easy. Typically, updates are moving down to level 2 results in 100
kept separately in an integrated row- files, each with a narrower key range
store. Queries check the small row- and longer time range.
store in a fashion that’s relatively fast The depth of an LSM tree depends
because it’s small. These queries are on the fan-out, the size of each file, and
combined with the results of the fast- the number of key-value pairs in the
er column-store to give a unified ac- tree. In general, most of the storage is
curate answer. Periodically, the new in the lowest level of the tree.
row-store updates are merged with the So, within this basic LSM structure
column-store to make a new column- that is gaining so much popularity,
store. This may be done in a cascading there are varieties of implementation
fashion somewhat like the merges in choices. Consider:
an log-structured merge (LSM) tree, ˲˲ Leveling merges. When a new file
described in the next section. is added to a level, pick the next file in
When inserting into a column-store the round-robin traversal and merge
(or really its attached row-store), you it with the files in the next level be-
are incurring a debt to be paid later. low. Suppose you pick a fan-out of 10;
This debt to rewrite and integrate the you will find the key range in the file
new data is a form of write amplifica- dropping down typically covers the
tion where a single write turns into key range in about 10 files in the level
more writes later. below. You merge 11 files together as
LSM trees were first proposed in one drops down onto 10 and you get
1996.6 The idea is to track changes to 11 files out. Now, the next level has
a key-value store as transactions, with gotten fatter by one file, so you repeat
new values kept in memory. As trans- and merge down again.
actions commit, the sorted collection ˲˲ Tiering merges. In this different
of recent key-value pairs can be writ- but related approach, you let a bunch
ten to disk in a uniquely named file. of files stack up on each level before
This file contains the sorted key-value doing the merge. Say you stack up 10
pairs along with an index into the keys files before you merge down at each
in the file. Once written to disk, the level. That dramatically reduces the
newly committed changes do not need amount of merging required.
to be kept in memory. Leveling merges have a large write
Now, if you keep doing this, look- amplification. Each write of a new
ing up values by key starts looking like key-value pair to level 0 will be re-
what happens to me when I try to find written 10 or 11 times at each level it
something I set down in some ran- moves through. On the other hand,
they have a small read perspiration, Internet-scale search systems clearly consistent read. This, in turn, can be
as a reader typically checks only one offer excellent and low read perspiration. seen as a tension between write ampli-
place per level. Large-scale caches. Lots of big fication and read perspiration.
Tiering merges have a much lower Internet systems have ginormous
write amplification but a larger read caches. Consider a product catalog at Conclusion
perspiration. Because new files stack a big ecommerce retailer. Whenever I have looked at just a few of the exam-
up at each level before merging, there anything changes, lots of servers are ples where there are trade-offs in our
is less merging and hence less writing. updated with the new product descrip- systems between write and read.1 It is
On the other hand, reads must check tion. This makes for a very easy and endemic in so many environments. We
a lot more places, leading to the larger fast read in exchange for a lot of writes. see emerging systems that adapt and
read perspiration. Normalization and denormaliza- optimize for these trade-offs as they
There’s a bunch of fun work lately tion. Growing up in the relational da- watch their usage patterns. Fun stuff!
on the trade-offs of these schemes.2,5 tabase world, I was imbued with the
Indexing and searching. Search is determination to have normalized
Related articles
in many ways a variation of database data contained in the database. Work- on queue.acm.org
indexing. In database indices, the no- ing to avoid update anomalies was
Immutability Changes Everything
tion of identity exists hidden within deemed to be extremely important.
Pat Helland
the database as a row-id or a primary Performing a large number of joins https://queue.acm.org/detail.cfm?id=2884038
key. Within a relational system, up- to get an answer was a small penalty
Disambiguating Databases
dates to indices are transactionally to pay to ensure the database wasn’t Rick Richardson
integrated, and the user sees only a damaged by an errant update. https://queue.acm.org/detail.cfm?id=2696453
performance difference. Increasingly, I view this as the The Pathologies of Big Data
Search systems are a bit different equivalent of throwing salt over your Adam Jacobs
in that they deal with documents. shoulder if you spill some. Yeah… https://queue.acm.org/detail.cfm?id=1563874
Most search systems asynchronous- I’ve seen others do it, but I’m not
ly update the search index after the sure I should. References
1. Athanassoulis, M., Kester, M.S., Maas, L. M., Stoica, R.,
change to the document occurs. This Most systems are getting more dis- Idreos, S., Ailamaki, A. and Callaghan, M. Designing
is knit together with some form of tributed. Most of these have key-value access methods: The RUM conjecture. In Proceedings
of the 19th International Conference on Extending
document identity.3 pairs containing their data, which is Database Technology (2016).
Search makes reading the docu- sharded for scale. By grouping related 2. Dayan, N. and Idreos, S. Dostoevsky: better space-
time tradeoffs for LSM-tree-based key-value stores
ments a lot easier. It dramatically low- data into the value of a pair—typically via adaptive removal of superfluous merging. In
ers the read perspiration. Updates to in a JSON (JavaScript Object Notation) Proceedings of the Intern. Conf. Management of Data
(2018), 505–520.
the documents asynchronously im- representation or something simi- 3. Helland, P. Identity by any other name. Commun.
pose a debt onto the system to get them lar—it’s easy to grab the value, per- ACM 62, 4 (Apr. 2019), 80.
4. Helland, P. Normalization is for sissies (July 23, 2007);
indexed. Creating and merging search haps as a string, and squirt it over to http://bit.ly/30iL7g3
indices is a complex job that I think of the distant system issuing the request. 5. Luo, C., and Carey, M.J. Forthcoming. LSM-
based storage techniques. Computing Surveys;
as a form of write amplification. If you were to normalize the data in arXiv:1812.07527.
To index, you must scour the cor- this big and sharded system, the nor- 6. O’Neil, P., Cheng, E., Gawlick, D. and O’Neil, E. The log-
structured merge-tree (LSM-tree). Acta Informatica
pus to find recently written or updated malized values would not be on the 33, 4 (1996).
documents. Each of these needs to same shard together. Doing a distrib-
have an identifier and then must be uted join is more annoying than doing Pat Helland has been implementing transaction systems,
databases, application platforms, distributed systems,
processed to locate the search terms a centralized join. fault-tolerant systems, and messaging systems since
(sometimes called n-grams; https:// To cope with this, people superim- 1978. He currently works at Salesforce.
en.wikipedia.org/wiki/n-gram). Each pose versioning on their data. It’s not
of these many n-grams found in a typi- perfect but it’s less challenging than
cal document then needs to be sent distributed joins or trying to do mas-
to an indexer that covers one of many sive updates across the denormalized
shards. So, the document identifier data. The classic example for the value
now is associated with each term (or of normalization in databases is a de-
n-gram) located in the searchable doc- normalized table with employees, their
ument—all of this because the user manager, and their manager’s phone
did a write or created a document! number.4 Because the manager’s phone
I worked for a few years on an In- number is copied in many tables for
ternet-scale search engine and know many employees, it’s hard to change
how they work. I’m still in awe that all it. Increasingly, I see systems store “as-
this machinery can keep up with the of” data in their denormalized struc-
work involved in all that write amplifi- tures—for example, the manager’s
cation. It’s a lot of work for each docu- phone is captured “as-of” June 1.
ment written—and there are lots and Large-scale distributed systems put Copyright held by author/owner.
lots of documents. a lot of pressure on the semantics of a Publication rights licensed to ACM.
The Five-
Today, enterprise database engines
use a three-tier storage hierarchy as
depicted in Figure 1. DRAM or NAND
flash solid state device (SSD)-based per-
Minute Rule
formance tier is used for hosting data
accessed by latency-critical transaction
processing and real-time analytics ap-
plications. The HDD-based capacity
30 Years Later
tier hosts data accessed by latency-in-
sensitive batch analytics applications.
The archival tier is not used for online
query processing, but for storing data
on the Storage
tape and is extremely crucial as a long-
term data repository for several appli-
cation domains like physics, banking,
security, and law enforcement.
Hierarchy
In this article, we revisit the five-
minute rule three decades after its in-
ception. We recomputed break-even
intervals for each tier of the modern,
multi-tiered storage hierarchy and use
guidelines provided by the five-minute
rule to identify impending changes in
the design of data management en-
gines for emerging storage hardware.
We summarize our findings here:
˲˲ HDD is tape. The gap between
THE DESIGN OF data management systems has always DRAM and HDD is increasing as the
five-minute rule valid for the DRAM–
been influenced by the storage hardware landscape. HDD case in 1987 is now a four-hour
In the 1980s, database engines used a two-tier storage rule. This implies the HDD-based ca-
hierarchy consisting of dynamic random access memory pacity tier is losing relevance for not
just performance sensitive applica-
(DRAM) and hard disk drives (HDD). Given the disparity tions, but for all applications with a
in cost between HDD and DRAM, it was important to non-sequential data access pattern.
˲˲ Non-volatile memory is DRAM.
determine when it made economic sense to cache data The gap between DRAM and SSD is
in DRAM as opposed to leaving it on the HDD. shrinking. The original five-minute
In 1987, Jim Gray and Gianfranco Putzolu rule is now valid for the DRAM–SSD
case, and the break-even interval is
established the five-minute rule that gave a precise less than a minute for newer non-vol-
answer to this question: “1KB records referenced every atile memory (NVM) devices like 3D-
The five-minute rule explores the trade- Rearranging the formulation by swap- the break-even interval for 4KB pages
off between the cost of DRAM and the ping the denominators provides the and the page sizes for which the five-
cost of disk I/O by providing a formula intuition behind the five-minute rule, minute rule is applicable across four
to predict the break-even interval—the as it reduces the equation to price-per- decades. In 1987, the break-even inter-
val was 400 seconds for 1KB pages. This even rare accesses to HDD should be owed by the decrease in technology ra-
was rounded down to five minutes, performed in large granularities. tio with SSDs, resulting in the interval
thus, lending the name for the rule. For DRAM–SSD. SSDs are being increas- shrinking.
4KB pages, the break-even interval was ingly used as the storage medium of SSD–HDD. As SSDs can also be used
100 seconds. When the study was re- choice in the latency-critical perfor- as a cache for HDD, the same formula
peated in 1997, the break-even interval mance tier due to their superior random can also be used to estimate the break-
had increased to nine minutes for 4KB access capability compared to HDDs. even interval for the SSD–HDD case.
pages, and the five-minute rule was de- Thus, the five-minute rule can be used From Table 3, we see the break-even
termined to hold only for 8KB pages. to compute a break-even interval for the interval for this case has increased by
Between 1997 and 2007, DRAM and case where DRAM is used to cache data a factor of 10× from 2.25 hours in 2007
HDD prices dropped further result- stored in SSDs. Table 3 shows the inter- to 1.5 days in 2018. The SSD–HDD in-
ing in the economic ratio increasing val in 2007, when SSDs were in the ini- terval is nine times longer than the
from 133 ($2k/$15) to 1700 ($80/$0.05). tial stages of adoption, and today, based DRAM–HDD interval of four hours.
However, the technology ratio did not on metrics listed in Table 1. Implications. There are two impor-
drop proportionately due to a lack of We see the interval has dropped tant consequences of these results.
improvement in HDD random access from 15 minutes to five minutes for First, in 2007, the turnover time in the
latency. As a result, the break-even in- 4KB pages. Thus, the five-minute rule DRAM–HDD case was six times higher
terval for 4KB pages increased 10×, is valid for SSDs today. This is in stark than the DRAM–SSD case (1.5h/15m).
from nine minutes to 1.5 hours. The contrast with the DRAM–HDD case, In 2018, it is nearly 50× higher (4h/5m).
five-minute rule was applicable only where the interval increased 2.7× from Thus, in systems tuned using econom-
for 64KB pages in 2007. 1.5 hours to four hours. In both DRAM– ic considerations, one should replace
Continuing this trend, the break- HDD and DRAM–SSD cases, the drop HDD with SSD, as it would not only im-
even interval for DRAM–HDD case to- in DRAM cost/MB dominated the eco- prove performance, but also reduce the
day is four hours for 4KB pages. The nomic ratio. However, unlike the 2.5× amount of DRAM required for cach-
five-minute rule is valid today for 512KB improvement in random I/Os-per- ing data. Second, given the four-hour
pages. The break-even interval trend in- second (IOPS) with HDDs, SSDs have DRAM–HDD and one day SSD–HDD in-
dicates it is more economical to store managed to achieve an impressive 11× tervals, it is important to keep all active
most data in DRAM instead of the HDD, improvement (67k/6.2k). Thus, the in- data in the DRAM or SSD-based per-
and the page size trend indicates that crease in economic ratio was overshad- formance tier and relegate the HDD-
based capacity tier to storing only in-
Figure 1. Storage tiering for enterprise databases. frequently accessed data. The growing
gap between performance and capac-
ity tiers also implies that SSD vendors
DRAM should optimize for $/IOPS, and HDD
SSD
Performance vendors, in contrast, should optimize
for $/GB. Next, we highlight recent
Storage Cost
DRAM-based SSD products. By the mid have found that 3D XPoint provides the cost of NAND flash is likely to drop
2000s, improvements in performance predictable access latencies that are faster than DRAM. This, in turn, will
and reliability of NAND flash resulted much lower than several state-of-the- result in the economic ratio dropping
in flash-based serial AT attachment art NAND flash devices even under se- further leading to a reduction in the
(SATA) SSDs gaining popularity in vere load.23 break-even interval.
niche application domains. The late Break-even interval and implica- Second, modern PCIe SSD is a highly
2000s witnessed the emergence of a tions. When we apply the five-minute parallel device that can provide very
new breed of peripheral component rule formula using metrics given in high random I/O throughput by ser-
interconnect express (PCIe) flash SSDs Table 4, we get a break-even interval vicing multiple outstanding I/Os con-
that could deliver two orders of mag- of one minute for 4KB pages in both currently. New non-volatile memory
nitude higher throughput than their the DRAM–NAND Flash PCIe SSD and technologies like 3D XPoint promise
SATA counterparts. Since then, a rapid DRAM–3D XPoint cases. Comparing further improvements in both through-
increase in capacity, drop in pricing, these results with Table 2, we see that put and access latencies over NAND
and new low-overhead interfaces like the breakeven interval is 10× shorter flash. With interfaces like NVMe, the
non-volatile memory express (NVMe), when PCIe SSDs or new PM technolo- end-to-end latency of accessing data
have all resulted in PCIe flash SSDs gies are used as the second tier instead from PCIe 3D XPoint SSDs is just tens
displacing their SATA counterparts as of SATA SSDs. This can be attributed to of µs. Thus, further improvements in
server accelerators of choice. the drop in technology ratio caused by non-volatile solid-state storage media
Table 4 (first row) shows the price/ the improvement in random IOPS. will result in a drop in technology ratio,
performance characteristics of a rep- Implications. Today, in the era of thereby reducing the break-even inter-
resentative, state-of-the-art PCIe SSD. in-memory data management, several val further.
In comparison to Table 1, we find the database engines are designed based Third, SSDs consume substantially
PCIe SSD offers five times higher read on the assumption that all data is resi- lower power than DRAM. The Intel 750
IOPS and sequential access bandwidth dent in DRAM. However, the dramatic SSD consumes 4W of power when idle
than its SATA counterpart. drop in breakeven interval computed and 22W when active. In contrast, 1TB
NVDIMM. As SSD vendors continue by the five-minute rule challenges this of DRAM in a server would consume
to improve throughput and capacity, trend of DRAM-based in-memory data 50W when idle and 100W when active.1
the bottleneck in the storage subsys- management due to three reasons. It is also well known that DRAM power
tem has shifted from the device itself First, recent projections indicate that consumption increases non-linearly
to the PCIe bus that is used to inter- flash density is expected to increase with capacity, as high-density DRAM
face with the SSD. Thus, over the past 40% annually over the next five years.5 consumes substantially more power
few years, NAND flash has started DRAM, in contrast, is doubling in ca- than their low-density counterparts.
transitioning once again from stor- pacity every three years.17 As a result, A recent study that focuses on power
age devices that are interfaced via the
high-latency, bandwidth-limited PCIe Table 2. The evolution of the page size for which the five-minute rule holds across four
decades based on appropriate price, performance, and page size values.
bus into non-volatile memory (NVM)
devices that are interfaced via the low-
latency, high-bandwidth memory bus. 1987 1997 2007 2018
These devices, also referred to as non- Break-even (4KB page) 100s 9m 1.5h 4h
volatile DIMMs (NVDIMM), use a com- Page size (5-minute interval) 1KB 8KB 64KB 512KB
bination of DRAM and flash storage
media packaged together as a dual in-
line memory module (DIMM).
NVM. Today, NVDIMMs are niche Table 3. The evolution of the break-even interval across four decades based on appropriate
price, performance, and page size values.
accelerators compared to PCIe SSDs
due to a high cost/GB. Unlike these
NVDIMM technologies that rely on Tier 1987 1997 2007 2018
NAND flash, new NVM technologies DRAM–SSD — — 15m 5m
that are touted to have better endur- SSD–HDD — — 2.25h 1.5d
ance, higher throughput, and lower
latency than NAND flash are being ac-
tively developed.
Table 4 (second row) shows the Table 4. Price/performance metrics for the NAND-based Intel 750 PCIe SSD and 3D-XPoint-
based Intel Optane P4800X PCIe SSD.
characteristics of Intel Optane DC
P4800X—a PCIe SSD based on 3D
XPoint, a new phase-changed-media- Device Capacity Price($) IOPS(k) B/w(GB/s)
based NVM technology. The cost/GB Intel 750 800GB 589 460 2.5
of 3D XPoint is higher than NAND Intel P4800X 480GB 617 550 2.5
flash today as the technology is yet to
mature. However, preliminary studies
consumption in main memory data- enabled HDDs to increase capacity at and the Information Storage Industry
bases showed that in a server equipped Kryder’s rate (40% per year), outstrip- Consortium (INSIC)13 project a contin-
with 6TB of memory, the idle power ping Moore’s Law. However, over the ued increase in density for the foresee-
of DRAM would match that of four ac- past few years, HDD vendors have hit able future.
tive CPUs.1 Such a difference in power walls in scaling areal density with con- Table 5 shows the price/perfor-
consumption between SSD and DRAM ventional Perpendicular Magnetic Re- mance metrics of tape storage both in
directly translates into higher Opera- cording (PMR) techniques resulting in 1997 and today. The 1997 values are
tional Expenses (OPEX), and hence, annual areal density improvement of based on the corresponding five-min-
higher Total Cost of Ownership (TCO), only around 16% instead of 40%.19 ute rule paper.8 The 2018 values are
for DRAM-based database engines. HDDs also present another prob- based on a SpectraLogic T50e tape li-
Given these three factors, the break- lem when used as the storage medium brary22 using LTO-7 tape cartridges.
even interval from the five-minute rule of choice for building a capacity tier, With individual tape capacity in-
seems to suggest an inevitable shift namely, high idle power consump- creasing 200× since 1997, the total ca-
from DRAM-based data management tion. Although enterprises gather vast pacity stored in tape libraries has ex-
engines to NVM-based persistent-mem- amounts of data, as one might expect, panded from hundreds of gigabytes to
ory engines. In fact, this change is al- not all data is accessed frequently. Re- hundreds of petabytes today. Further,
ready well under way, as state-of- the-art cent studies estimate that as much as a single LTO-7 cartridge is capable of
database engines are being updated to 80% of enterprise data is “cold,” mean- matching, or even outperforming a
fully exploit the performance benefits ing infrequently accessed, and that HDD, with respect to sequential data
of PCIe NVMe SSDs.26 Researchers have cold data is the largest growing seg- access bandwidth as shown in Table
recently highlighted the fact that data ment with a 60% Cumulative Annual 6. As modern tape libraries use multi-
caching systems that trade-off perfor- Growth Rate (CAGR).10–12 Unlike tape, ple drives, the cumulative bandwidth
mance for price by reducing the amount which consumes no power once un- achievable using even low-end tape li-
of DRAM are gaining market share over mounted, HDDs consume a substan- braries is 1–2GB/s. High-end libraries
in-memory database engines.18 tial amount of power even while idle. can deliver well over 40GB/s. These
Such power consumption translates to benefits have made tape the prefer-
The Capacity Tier a proportional increase in TCO. able media of choice in the archival
HDD. Traditionally, HDDs have been Tape. The areal density of tape has tier both on-premise and in the cloud,
the primary storage media used for been increasing steadily at a rate of for several applications ranging from
provisioning the capacity tier. For sev- 33% per year and roadmaps from the natural sciences, like particle physics
eral years, areal density improvements Linear Tape Open consortium (LTO)25 and astronomy, to movies archives in
the entertainment industry.15,20 How-
Table 5. Price/performance characteristics of tape. ever, random access latency of tape is
still 1000× higher than HDD (minutes
1997 2018 vs. ms) due to the fact that tape librar-
Tape library cost ($) 10,000 11,000 ies need to mechanically load and
Number of drives 1 4 wind tape cartridges before data can
Number of slots 14 10 be accessed.
Max capacity per tape 35GB 15TB Break-even interval and implica-
Transfer rate per drive (MB/s) 5 750 tions. Using metrics from Tables 1, 5
Access latency 30s 65s to compute the break-even interval for
the DRAM–tape case results in an in-
terval of over 300 years for a page size
of 4KB! Jim Gray referred to tape drives
Table 6. Price/performance metrics of DRAM, HDD, and tape. as the “data motel” where data checks
in and never checks out,7 and this is
Metric DRAM HDD Tape certainly true today. Figure 2 shows
Unit capacity 16GB 2TB 10 × 15TB the variation in break-even interval
Unit cost ($) 80 50 11,000 for both HDD and tape for various
Latency 100ns 5ms 65s page sizes. We see that the interval
Bandwidth 100 GB/s 200 MB/s 4 × 750MB/s asymptotically approaches one min-
Kaps 9,000,000 200 0.02 ute in the DRAM–HDD case and 10
Maps 10,000 100 0.02 minutes in the DRAM–tape case. The
Scan time 0.16s 3hours 14hours HDD asymptote is reached at a page
$/Kaps 9e-14 5e-09 8e-03 size of 100MB and the tape asymp-
$/Maps 9e-12 8e-09 8e-03 tote is reached at a size of 100GB. This
$/Tbscan 8e-06 0.003 0.03 clearly shows that randomly access-
$/TBscan (97) 0.32 4.23 296 ing data on these devices is extremely
expensive, and data transfer sizes
with these devices should be large to
amortize the cost of random accesses. 70× cheaper than tape. However, today to cooling and power restrictions en-
However, the primary use of the it is only 10× cheaper. Unlike HDD, se- forced by hardware. Access to data in
capacity tier today is not sup-porting quential data transfer bandwidth of any of the spun-up disks can be done
applications that require high-perfor- tape is predicted to double for the fore- with latency and bandwidth compa-
mance random accesses. Rather, it is seeable future. Hence, this difference rable to that of the traditional capacity
to reduce the cost/GB of storing data is likely to shrink further. Thus, in the tier. For instance, Pelican, OpenVault
over which latency-insensitive batch near future, it might not make much of Knox, and ArticBlue provide between
analytics can be performed. Indeed, a difference whether data is stored in a 1–2GB/s of throughput for reading
Gray and Graefe noted that metrics tape or HDD with respect to the price data from spun-up disks.2,21,27 How-
like KB-accesses-per-second (Kaps) are paid per TB scan. ever, accessing data on a spun-down
less relevant for HDD and tape as they Implications. Today, all data gener- disk takes several seconds, as the disk
grow into infinite-capacity resourc- ated by an enterprise has to be stored has to be spun up before data can be
es.8 Instead, MB-accesses-per-second twice, once in the traditional HDD- retrieved. Thus, CSDs form a perfect
(Maps) and time to scan the whole de- based capacity tier for enabling batch middle ground between HDD and tape
vices are more pertinent to these high- analytics, and a second time in the with respect to both cost/GB and ac-
density storage devices. Table 6 shows tape-based archival tier for meeting cess latency.
these new metrics and their values for regulatory compliance requirements. On the application front, there is a
DRAM, HDD, and tape. In addition to The shrinking difference in $/TBscan clear bifurcation in demand between
Kaps, Maps, and scan time, the table between HDD and tape suggests that latency-sensitive interactive applica-
also shows $/Kaps, $/Maps, and $/TB- it might be economically beneficial to tions and latency insensitive batch ap-
scan, where costs are amortized over a merge the capacity and archival tiers plications. As interactive applications
three-year time frame as proposed by into a single cold storage tier.3 However, are isolated to the performance tier,
Gray and Graefe.8 with such a merger, the cold storage tier the cold storage tier only has to cater
Looking at $/Kaps, we see that DRAM would no longer be a near-line tier that to the bandwidth demands of latency-
is five orders of magnitude cheaper is used rarely during disaster recovery, insensitive batch analytics applica-
than HDD, which, in turn, is six orders but an online tier that is used for run- tions. Nearline storage devices like
of magnitude cheaper than tape. This ning batch analytics applications. Re- tape libraries and CSD are capable of
is expected given the huge disparity cent hardware and application trends providing high-throughput access for
in random access latencies and is in indicate that it might be feasible to sequentially accessed data. Thus, re-
accordance with the five-minute rule build such a cold storage tier. searchers have recently started investi-
that favors using DRAM for randomly On the hardware front, storage ven- gating extensions to batch processing
accessed data. Looking at $/Maps, we dors have recently started building frameworks for enabling analytics di-
see that the difference between DRAM new cold storage devices (CSD) for stor- rectly over data stored in tape archives
and HDD shrinks to roughly 1,000×. ing cold data. Each CSD is an ensemble and CSD. For instance, Nakshatra im-
This is due to the fact that HDDs can of HDDs grouped in a massive array of plements prefetching and I/O schedul-
provide much higher throughput for idle disks (MAID) setup where only a ing extensions to Hadoop so that ma-
sequential data accesses over random small subset of disks are active at any preduce jobs can be scheduled to run
ones. However, HDD continue to be six given time.2,4,27 For instance, Pelican directly on tape archives.14 Skipper is a
orders of magnitude cheaper than tape CSD pro vides 5PB of storage using query-processing framework that uses
even for MB-sized random data access- 1,152 SMR disks packed as a 52U rack adaptive query processing techniques
es. This, also, is in accordance with the appliance.2 However, only 8% of disks in combination with customized cach-
HDD/tape asymptote shown in Figure can be spun up simultaneously due ing and I/O scheduling to enable que-
2. Finally, $/TBscan paints a very dif-
ferent picture. While DRAM remains Figure 2. Break-even interval asymptotes for DRAM–HDD and DRAM–tape cases.
300× cheaper than HDD, the difference
between HDD and tape shrinks to 10×. 1E+09
DRAM-HDD DRAM-Tape
Comparing the $/TBscan values 100,000,000
with those reported in 1997, we can see
Break-even interval (min)
10,000,000
two interesting trends. First, the dispar-
ity between DRAM and HDD is growing 1,000,000
over time. In 1997, it was 13× cheaper 10,000
to use DRAM for a TBscan than HDD.
1,000
Today, it is 300× cheaper. This implies
that even for scan-intensive applica- 100
tions, unsurprisingly, optimizing for 10
performance requires avoiding using
1
HDD as the storage medium. Second,
1 1,000 1,000,000 1E+09 1E+12
the difference between HDD and tape
Page size (KB)
is following the opposite trend and
shrinking over time. In 1997, HDD was
ry execution over CSD.3 Skipper even answered in order for the cold storage 4. Colarelli, D. and Grunwald, D. Massive arrays of idle
disks for storage archives. In Proceedings of 2002
shows that for long-running batch que- tier to be feasible in practice. Conference on Supercomputing.
ries, using CSD results in query execu- Over the past few years, several other 5. Coughlin, T. Flash memory areal densities exceed
those of hard drives; http://bit.ly/2NbDh5T.
tion time increasing by only 35% com- systems have been built to reduce the 6. Graefe, G. The five-minute rule 20 years later (and
pared to a traditional HDD despite the cost of storing cold data using alterna- how flash memory changes the rules). Commun. ACM
52, 7 (July 2009).
long disk spin-up latency. With such tive storage media. For instance, DT- 7. Gray, J. The five-minute rule; research.microsoft.com/
frameworks, it should be possible for Store16 uses LTFS tape archive for re- en-us/um/people/gray/talks/fiveminuterule.ppt.
8. Gray, J. and Graefe, G. The five-minute rule ten years
installations to switch from the tradi- ducing the TCO of online multimedia later, and other computer storage rules of thumb.
tional three-tier hierarchy to a two-tier streaming services by storing cold data SIGMOD Rec. 26, 4 (1997).
9. Gray, J. and Putzolu, F. The 5-minute rule for trading
hierarchy consisting of just a perfor- in tape drives. ROS28 is a PB-sized, rack- memory for disc accesses and the 10-byte rule for
mance tier with DRAM and SSDs, and scale cold storage library built using trading memory for CPU time. In Proceedings of
SIGMOD, 1987.
a cold storage tier with CSDs. thousands of optical discs packed in 10. Horison Information Strategies Report. Tiered storage
takes center stage, IDC. Technology assessment:
a single 42U Rack. Today, it is unclear Cold storage is hot again — finding the frost point;
Conclusion and Future Work as to how these alternative storage op- http://www.storiant.com/resources/Cold-Storage-Is-
Hot-Again.pdf.
Modern database engines use a three- tions fare with respect to HDD-based 11. Intel. Cold Storage in the Cloud: Trends, Challenges,
tier storage hierarchy across four CSD as the storage media of choice for and Solutions, 2013; https://intel.ly/2ZG74F6.
12. I.S.I. Consortium. International magnetic
primary storage media (DRAM, SSD, storing cold data. Furthermore, in or- tape storage roadmap; http://www.insic.org/
HDD, and tape) with widely varying der for the Cold Storage Tier to be re- news/2015roadmap/15index.html
13. Kathpal, A. and Yasa, G.A.N. Nakshatra: Towards
price-performance characteristics. In alized in practice, an ideal cold storage running batch analytics on an archive. In Proceedings
this article, we revisited the five-minute media needs to support batch analytics of MASCOTS, 2014.
14. Lantz, M. Why the future of data storage is (still)
rule in the context of this modern stor- workloads. CSD, tape, and optical me- magnetic tape; http://bit.ly/2XChrMO
age hierarchy and used it to highlight dia are all primarily used today for ar- 15. Lee, J., Ahn, J., Park, C., and Kim, J. Dtstorage:
Dynamic tape-based storage for cost-effective and
impending changes based on recent chival storage where data is rarely read. highly-available streaming service. In Proceedings of
trends in the hardware landscape. Further research is required to under- CCGRID, 2016.
16. Lim, K., Chang, J., Mudge, T., Ranganathan, P.,
In the performance tier, NAND flash stand the reliability implications of us- Reinhardt, S.K., and Wenisch, T.F. Disaggregated
is inching its way closer to the CPU re- ing these storage devices under batch memory for expansion and sharing in blade servers. In
Proceedings of ISCA, 2009.
sulting in dramatic improvements in analytics workloads. 17. Lomet, D. Cost/performance in modern data stores:
How data caching systems succeed. In Proceedings of
both access latency and bandwidth. For Finally, with widespread adoption DaMoN, 2018.
state-of-the-art PCIe SSDs, the break- of cloud computing, the modern enter- 18. Moore, F. Storage outlook 2016; http://bit.ly/2KBLgao.
19. Perlmutter, M. The lost picture show: Hollywood
even interval predicted by the five-min- prise storage hierarchy not only spans archivists cannot outpace obsolescence, 2017; http://
ute rule is one minute for 4KB pages. several storage devices, but also differ- bit.ly/2KDaqWd.
20. Spectra. Arcticblue deep storage disk. Product, https://
Going forward, further improvements ent geographic locations from direct- www.spectralogic.com/products/arcticblue/.
in NAND flash and the introduction of attached low-latency devices, through 21. SpectraLogic. Spectralogic t50e; http://bit.ly/2Ych8pl.
22. StorageReview. Intel optane memory review. http://
new NVM technologies will likely result network-attached storage servers, to www.storagereview.com/intel_optane_memory_review.
in this interval dropping further. As the cloud-hosted storage services. The 23. TPC-C. Dell-microsoft sql server tpc-c executive
summary, 2014; http://www.tpc.org/tpcc/results/
data reuse window shrinks, it will soon price-performance characteristics of tpcc_result_ detail.asp?id=114112501.
be economically more valuable to store these storage configurations vary dra- 24. Ultrium. LTP ultrium roadmap;
25. http://www.ltoultrium.com/lto-ultrium-roadmap/.
most, if not all, data on solid-state stor- matically depending not only on the 26. Umamageswaran, K. and Goindi, G. Exadata:
age devices instead of DRAM. This will storage media used, but also on other Delivering memory performance with shared flash;
http://bit.ly/2LhBVEa.
invariably necessitate revisiting several factors like the total capacity of data 27. Yan, M. Open compute project: Cold storage hardware
techniques pioneered by traditional stored, the frequency and granular- v0.5, 2013; http://bit.ly/2X6H2Ot.
28. Yan, W., Yao, J., Cao, Q., Xie, C., and Jiang, H. Ros:
HDD-based database engines, but es- ity of I/O operations used to access the A rack-based optical storage system with inline
accessibility for long-term data preservation. In
chewed by in-memory engines, like data, the read–write ratio, the duration Proceedings of EUROSYS, 2017.
buffer caching, on-disk storage layout, of data storage, and the cloud service
and index persistence, to name a few, provider used, to name a few. Given the Raja Appuswamy (raja.appuswamy@eurecom.fr) is an
for these new low-latency, high-band- multitude of factors, determining the assistant professor in the Data Science Department at
EURECOM, Biot, Provence-Alpes-Côte d’Azur, France.
width storage devices. break-even interval for cloud storage is
Goetz Graefe (goetzg@google.com), Google, Inc.,
Traditionally, HDDs have been a complicated problem that we did not Madison, WI, USA.
used for implementing the capac- consider in this work. Thus, another
Renata Borovica-Gajic (renata.borovica@unimelb.edu.
ity tier. However, our analysis showed interesting avenue of future work is ex- au) is an assistant professor in the School of Computing
that the difference between HDD and tending the five-minute rule to such a and Information Systems at the University of Melbourne,
Australia.
tape is shrinking when $/TBScan is distributed cloud storage setting.
Anastasia Ailamaki (anastasia.ailamaki@epfl.ch) is a
used as the metric. Given the latency- professor at EPFL, Lausanne, Switzerland, and director of
insensitive nature of batch analytics References its Data-Intensive Applications and Systems (DIAS) lab.
workloads, it is economically benefi- 1. Appuswamy, R., Olma, M., and Ailamaki, A. Scaling
the memory power wall with dram-aware data
cial to merge the HDD-based capacity management. In Proceedings of DaMoN, 2015.
2. Balakrishnan, S. et al. Pelican: A building block
tier and the tape-based archival tier for exascale cold data storage. In Proceedings of
into a single cold storage tier as dem- OSDI, 2014.
3. Borovica-Gajic, R., Appuswamy, R., and Ailamaki, A.
onstrated by recent research.3 Howev- Cheap data analytics using cold storage devices. In
er, several open questions still need to Proceedings of VLDB 9, 12 (2016). © 2019 ACM 0001-0782/19/11
M
&C
Association for Computing Machinery
1601 Broadway, 10th Floor, New York, NY 10019-7434, USA
Phone: +1-212-626-0658 Email: acmbooks-info@acm.org
review articles
DOI:10.1145/ 3363294
An
Elementary
Introduction
to Kalman
Filtering
of Bayesian inference, assuming that
KAL M AN F ILTER I NG I S a state estimation technique noise is Gaussian. This leads to the
used in many application areas such as spacecraft common misconception that Kalman
filtering can be applied only if noise
navigation, motion planning in robotics, signal is Gaussian.15
processing, and wireless sensor networks because Abstractly, Kalman filtering can be
seen as a particular approach to combin-
of its ability to extract useful information from ing approximations of an unknown val-
noisy data and its small computational and memory ue to produce a better approximation.
requirements.12,20,27–29 Recent work has used Kalman Suppose we use two devices of different
filtering in controllers for computer systems.5,13,14,23 key insights
Although many introductions to Kalman filtering are ˽˽ This article presents an elementary
available in the literature,1–4,6–11,17,21,25,29 they are usually derivation of Kalman filtering, a classic
state estimation technique.
focused on particular applications such as robot motion ˽˽ Understanding Kalman filtering is
or state estimation in linear systems, making it difficult to useful for more principled control
of computer systems.
see how to apply Kalman filtering to other problems. Other ˽˽ Kalman filtering is used as a black box
presentations derive Kalman filtering as an application by many computer scientists.
One ad hoc solution is to use the for- estimator 0.0*x1 + 1.0*x2. Kalman filter-
mula 0.5*x1+0.5*x2 to take the average ing tells us that in general, this intui- a An extended version of this article that in-
of the two measurements, giving them tively reasonable linear estimator is not cludes additional background material and
equal weight. Formulas of this sort are “optimal;” paradoxically, there is useful proofs is available.30
application of these general concepts. random sample from the distribution the pdf for the possible values of x2. If
First, the informal ideas discussed here for that device. We write to the random variables are only uncor-
are formalized using the notions of distri- denote that xi is a random variable with related, knowing x1 might give us new
butions and random samples from distri- pdf pi whose mean and variance are µi information about x2 such as restricting
butions. Confidence in estimates is and , respectively; following conven- its possible values but the mean of x2|x1
quantified using the variances and covari- tion, we use xi to represent a random will still be µ2. Using expectations, this
ances of these distributions.b Two algo- sample from this distribution as well. can be written as E[x2|x1] = E[x2], which is
rithms are described next. The first one Means and variances of distribu- equivalent to requiring that E[(x1−µ1)(x2−
shows how to fuse estimates (such as core tions model different kinds of inaccura- µ2)], the covariance between the two vari-
temperature measurements) optimally, cies in measurements. Device i is said to ables, be equal to zero. This is obviously
given a reasonable definition of optimal- have a systematic error or bias in its a weaker condition than independence.
ity. The second algorithm addresses a measurements if the mean µi of its dis- Although the discussion in this sec-
problem that arises frequently in practice: tribution is not equal to the actual tem- tion has focused on measurements,
estimates are vectors (for example, the perature xc (in general, to the value being the same formalization can be used for
position and velocity of a robot), but only a estimated, which is known as ground estimates produced by an estimator.
part of the vector can be measured truth); otherwise, the instrument is unbi- Lemma 1(i) shows how the mean and
directly; in such a situation, how can an ased. Figure 1 shows pdfs for two devices variance of a linear combination of pair-
estimate of the entire vector be obtained that have different amounts of systematic wise uncorrelated random variables can
from an estimate of just a part of that error. The variance on the other hand be computed from the means and vari-
vector? The best linear unbiased esti- is a measure of the random error in the ances of the random variables.18 The
mator (BLUE) is used to solve this prob- measurements. The impact of random mean and variance can be used to quan-
lem.16,19,26 It is shown that the Kalman errors can be mitigated by taking many tify bias and random errors for the esti-
filter can be derived in a straightfor- measurements with a given device and mator as in the case of measurements.
ward way by using these two algorithms averaging their values, but this approach An unbiased estimator is one whose
to solve the problem of state estimation will not reduce systematic error. mean is equal to the unknown value
in linear systems. The extended Kalman In the formulation of Kalman fil- being estimated and it is preferable to a
filter and unscented Kalman filter, tering, it is assumed that measuring biased estimator with the same variance.
which extended Kalman filtering to non- devices do not have systematic errors. Only unbiased estimators are considered
linear systems, are described briefly at However, we do not have the luxury of in this article. Furthermore, an unbiased
the end of the article. taking many measurements of a given estimator with a smaller variance is pref-
state, so we must take into account the erable to one with a larger variance as we
Formalizing Estimates impact of random error on a single would have more confidence in the esti-
Scalar estimates. To model the behav- measurement. Therefore, confidence mates it produces. As a step toward gener-
ior of devices producing noisy tempera- in a device is modeled formally by the alizing this discussion to estimators that
ture measurements, we associate each variance of the distribution associated produce vector estimates, we refer to the
device i with a random variable that has with that device; the smaller the vari- variance of an unbiased scalar estimator
a probability density function (pdf) pi(x) ance, the higher our confidence in the as the mean square error of that estimator
such as the ones shown in Figure 1 (the measurements made by the device. In or MSE for short.
x-axis in this figure represents tempera- Figure 1, the fact we have less confi- Lemma 1(ii) asserts that if a random
ture). Random variables need not be dence in the first device has been illus- variable is pairwise uncorrelated with
Gaussian.c Obtaining a measurement trated by making p1 more spread out a set of random variables, it is uncor-
from device i corresponds to drawing a than p2, giving it a larger variance. related with any linear combination of
The informal notion that noise should those variables.
affect the two devices in “unrelated
b Basic concepts such as probability density func- ways” is formalized by requiring that Lemma 1. Let
tion, mean, expectation, variance and covari-
ance are introduced in the online appendix.
the corresponding random variables be be a set of pairwise uncorrelated
c The role of Gaussians in Kalman filtering is uncorrelated. This is a weaker condition random variables. Let be a
discussed later in the article. than requiring them to be independent, random variable that is a linear combi-
as explained in our online appendix nation of the xi’s.
Figure 1. Using pdfs to model devices with (http://dl.acm.org/citation.cfm?doid=
systematic and random errors. Ground truth
is 60°C. Dashed lines are means of pdfs.
3363294&picked=formats). Suppose we (i) The mean and variance of y are:
are given the measurement made by
p2(x)
one of the devices (say x1) and we have (1)
to guess what the other measurement
(x2) might be. If knowing x1 does not give
(2)
p(x)
p1(x)
us any new information about what x2
might be, the random variables are inde-
x pendent. This is expressed formally by (ii) If random variable xn+1 is pair-wise
58 60 63
the equation p(x2|x1) = p(x2); intuitively, uncorrelated with x1,..,xn, it is
knowing the value of x1 does not change uncorrelated with y.
(9)
(13)
(14)
The expressions for y and
complicated because they contain the
are
Kalman filtering Equations 13 and 14 generalize
reciprocals of variances. If we let ν1 and
ν2 denote the precisions of the two dis-
can be seen as a Equations 10 and 11.
Incremental fusing is optimal. In
tributions, the expressions for y and νy particular approach many applications, the estimates x1, x2,
can be written more simply as follows: to combining …, xn become available successively over
a period of time. Although it is possible to
(10) approximations of store all the estimates and use Equations
(11)
an unknown value 13 and 14 to fuse all the estimates from
scratch whenever a new estimate
to produce a better becomes available, it is possible to save
These results say the weight we should
give to an estimate is proportional to the approximation. both time and storage if one can do this
fusion incrementally. We show that just
confidence we have in that estimate, as a sequence of numbers can be added
and that we have more confidence in the by keeping a running sum and adding
fused estimate than in the individual esti- the numbers to this running sum one at a
mates, which is intuitively reasonable. To time, a sequence of n>2 estimates can be
use these results, we need only the vari- fused by keeping a “running estimate”
ances of the distributions. In particular, and fusing estimates from the sequence
the pdfs pi, which are usually not avail- one at a time into this running estimate
able in applications, are not needed, and without any loss in the quality of the final
the proof of Theorem 1 does not require estimate. In short, we want to show that
these pdfs to have the same mean. yn(x1, .., xn)=y2(y2(..y2(x1, x2)…), xn). A little
Fusing multiple scalar estimates. bit of algebra shows that if n>2,
These results can be generalized Equations 13 and 14 for the optimal linear
to optimally fuse multiple pairwise estimator and its precision can be
uncorrelated estimates x1, x2, …, xn. expressed as shown in Equations 15 and 16.
Let yn,α(x1, .., xn) denote the linear esti-
mator for fusing the n estimates given
parameters α1, .., αn, which we denote
by α (the notation yα(x1, x2) introduced (15)
previously can be considered to be an
abbreviation of y2,α(x1, x2) ).
(16)
Theorem 2. Let for (1≤i≤n)
be a set of pairwise uncorrelated ran- This shows that yn(x1, .., xn) = y2(yn−1
dom variables. Consider the linear esti- (x1, .., xn−1), xn). Using this argument
mator where recursively gives the required result.d
. The variance of the estimator To make the connection to Kalman
is minimized for filtering, it is useful to derive the
same result using a pictorial argu-
ment. Figure 2 shows the process of
incrementally fusing the n estimates.
In this picture, time progresses from
left to right, the precision of each esti-
The minimal variance is given by the mate is shown in parentheses next to
following expression: it, and the weights on the edges are
(12) the weights from Equation 10. The
contribution of each xi to the final
value y2(y2(…), xn) is given by the prod-
uct of the weights on the path from xi
As before, these expressions to the final value, and this product is
are more intuitive if the variance is obviously equal to the weight of xi in
replaced with precision: the contribu-
tion of xi to the value of yn(x1, .., xn) is d We thank Mani Chandy for showing us this
proportional to xi’s confidence. approach to proving the result.
(ν1 + ν2) (ν1 + ν2 + ν3) (ν1 + ··· + νn−1) (ν1 + ··· + νn)
y2(x1, x2) y2(y2(x1, x2), x3) y2(y2(···), xn−1) y2(y2(···), xn)
ν1 ν1+ν2 ν1+···+νn−1
(ν1) ν1+ν2 ν1+ν2+ν3 ν1+···+νn
...
x1
ν2 ν3 νn−1 νn
ν1+ν2 ν1+ν2+ν3 ν1+···+νn−1 ν1+···+νn
x2 x3 xn−1 xn
(ν2) (ν3) (νn−1) (νn)
Equation 13, showing that incremen- For vectors, the linear estimator random variables. Consider the linear
tal fusion is optimal. is where estimator , where
Summary. The results in this section . Here A stands for the matrix . The value of MSE( yA) is mini-
can be summarized informally as fol- parameters (A1, …, An). All the vectors (xi) mized for
lows. When using a linear estimator to fuse are assumed to be of the same length.
uncertain scalar estimates, the weight given To simplify notation, we omit the sub- (23)
to each estimate should be inversely propor- script n in yn,A in the discussion here
tional to the variance of the random vari- as it is obvious from the context. Therefore the optimal estimator is
able from which that estimate is obtained. Optimality. The parameters A1, …,
Furthermore, when fusing n>2 estimates, An in the linear data fusion model are (24)
estimates can be fused incrementally with- chosen to minimize MSE(yA) which is
out any loss in the quality of the final result. E[(yA−µyA)T(yA−µyA)]. The covariance matrix of y can be
These results are often expressed formally Theorem 3 generalizes Theorem 2 to computed by using Lemma 2.
in terms of the Kalman gain K, as shown the vector case. The proof of this theorem
in Figure 3; the equations can be applied is given in the appendix. Comparing (25)
recursively to fuse multiple estimates. Theorems 2 and 3, we see that the
Note that if ν1ν2, K≈0 and y(x1,x2)≈x1; expressions are similar, the main dif- In the vector case, precision is the
conversely if ν1ν2, K≈1 and y(x1,x2)≈x2. ference being that the role of variance inverse of a covariance matrix, denoted
in the scalar case is played by the covari- by N. Equations 26–27 use precision to
Fusing Vector Estimates ance matrix in the vector case. express the optimal estimator and its
The results for fusing scalar estimates variance and generalize Equations 13–14
can be extended to vectors by replacing Theorem 3. Let xi∼pi(µi, ∑i) for (1≤i≤n) to the vector case.
variances with covariance matrices. be a set of pairwise uncorrelated
(26)
Figure 3. Optimal fusion of scalar estimates.
(27)
2 2
x1∼p1(µ1,σ1), x2∼p2(µ2,σ2)
As in the scalar case, fusion of n>2 vec-
2 tor estimates can be done incrementally
σ1 ν2
K= = (17) without loss of precision. The proof is
σ12 + σ22 ν1 + ν2
similar to the scalar case and is omitted.
y(x1, x2)=x1 + K(x2 − x1) (18)
σy2=(1−K )σ12 or νy = ν1 + ν2 (19) Figure 5. BLUE line corresponding to
Equation (31).
y
Figure 4. Optimal fusion of vector estimates. −1
(y − µy)=Σyx Σxx(x − µx)
y1
µx
µy
x
x1
There are several equivalent expres- Best Linear Unbiased Estimator In our context, however, x and y are
sions for the Kalman gain for the fusion (BLUE) random variables, so such a precise
of two estimates. The following one, In some applications, the state of the functional relationship will not hold.
which is easily derived from Equation system is represented by a vector but Figure 5 shows an example in which x
23, is the vector analog of Equation 17: only part of the state can be measured and y are scalar-valued random vari-
directly, so it is necessary to estimate ables. The gray ellipse in this figure,
(28) the hidden portion of the state corre- called a confidence ellipse, is a pro-
sponding to a measured value of the jection of the joint distribution of x
The covariance matrix of the opti- visible state. This section describes an and y onto the (x, y) plane that shows
mal estimator y(x1, x2) is the following. estimator called the best linear unbiased where some large proportion of the
estimator (BLUE)16,19,26 for doing this. (x, y) values are likely to be. Suppose
(29) Consider the general problem of x takes the value x1. Even within the
(30) determining a value for vector y given confidence ellipse, there are many
a value for a vector x. If there is a func- points (x1, y), so we cannot associate
Summary. The results in this sec- tional relationship between x and y (say a single value of y with x1. One possi-
tion can be summarized in terms of the y=F(x) and F is given), it is easy to com- bility is to compute the mean of the
Kalman gain K as shown in Figure 4. pute y given a value for x (say x1). y values associated with x1 (that is,
(d) Implementation of the dataflow diagram (b) for systems with partial observability.
the expectation E[y|x=x1]) and return line. However, if we know that x has a par- during that interval. This is usually
this as the estimate for y if x=x1. This ticular value x1, we can use the correlation expressed by an equation of the form xt
requires knowing the joint distribu- between y and x to estimate a better value = ft(xt−1, ut) where ut is the control input.
tion of x and y, which may not always for y from the difference (x1−µx). Note The function ft is nonlinear in the gen-
be available. that if ∑yx = 0 (that is, x and y are uncor- eral case, and can be different for differ-
In some problems, we can assume related), the best estimate of y is just µy, ent steps. If the system is linear, the
that there is an unknown linear rela- so knowing the value of x does not give relation for state evolution over time
tionship between x and y and that us any additional information about can be written as xt = Ftxt−1 + Btut, where Ft
uncertainty comes from noise. y as one would expect. In Figure 5, and Bt are time-dependent matrices
Therefore, we can use a technique simi- this corresponds to the case when the that can be determined from the physics
lar to the ordinary least squares (OLS) BLUE line is parallel to the x-axis. At the of the system. Therefore, if the initial
method to estimate this linear relation- other extreme, suppose that y and x are state x0 is known exactly and the system
ship, and use it to return the best esti- functionally related so y = Cx. In that dynamics are modeled perfectly by the Ft
mate of y for any given value of x. In Figure case, it is easy to see that ∑yx = C∑xx, and Bt matrices, the evolution of the state
5, we see that although there are many so as expected. In Figure 5, over time can be computed precisely as
points (x1, y), the y values are clustered this corresponds to the case when the shown in Figure 6a.
around the line as shown in the figure so confidence ellipse shrinks down to the In general, however, we may not
the value is a reasonable estimate for the BLUE line. know the initial state exactly, and the
value of y corresponding to x1. This line, Equation 31 is a generalization of system dynamics and control inputs
called the best linear unbiased estimator ordinary least squares in the sense that if may not be known precisely. These inac-
(BLUE), is the analog of ordinary least we compute the relevant means and vari- curacies may cause the state computed
squares (OLS) for distributions. ances of a set of discrete data (xi, yi) and by the model to diverge unacceptably
Computing BLUE. Consider the substitute into Equation 31, we get the from the actual state over time. To avoid
estimator . We choose A same line that is obtained by using OLS. this, we can make measurements of the
and b so that this is an unbiased esti- state after each time step. If these mea-
mator with minimal MSE. The “∧” over Kalman Filters for Linear Systems surements were exact, there would of
the y is notation that indicates that we We now apply the algorithms for opti- course be no need to model the system
are computing an estimate for y. mal fusion of vector estimates (Figure 4) dynamics. However, in general, the mea-
and the BLUE estimator (Theorem 4) to surements themselves are imprecise.
Theorem 4. Let the particular problem of state estima- Kalman filtering was invented to
tion in linear systems, which is the clas- solve the problem of state estimation in
sical application of Kalman filtering. such systems. Figure 6b shows the data-
Figure 6a shows how the evolution of flow of the computation, and we use it
The estimator for esti- the state of such a system over time can to introduce standard terminology. An
mating the value of y for a given value of be computed if the initial state x0 and estimate of the initial state, denoted by
x is an unbiased estimator with minimal the model of the system dynamics are , is assumed to be available. At each
MSE if known precisely. Time advances in dis- time step t=1, 2, .., the system model is
crete steps. The state of the system at used to provide an estimate of the state
any time step is a function of the state of at time t using information from time
the system at the previous time step and t−1. This step is called prediction and
the control inputs applied to the system the estimate that it provides is called the
(31)
a priori estimate and denoted by . problem when only a portion of the of wt is denoted by Qt, and the
The a priori estimate is then fused with state can be measured directly. noise terms in different time
zt, the state estimate obtained from the State evolution model and p
rediction. steps are assumed to be uncorre-
measurement at time t, and the result is The evolution of the state over time is lated to each other (such as,
the a posteriori state estimate at time t, described by a series of random vari- E[wiwj]=0 if i≠j) and to x0.
denoted by . This a posteriori estimate ables x0, x1, x2,…
is used by the model to produce the a For estimation, we have a random
priori estimate for the next time step • The random variable x0 captures variable x0|0 that captures our belief
and so on. As described here, the a priori the likelihood of different initial about the likelihood of different states
and a posteriori estimates are the means states. at time t=0, and two random variables
of certain random variables; the covari- • The random variables at succes- xt|t−1 and xt|t at each time step t = 1, 2, …
ance matrices of these random variables sive time steps are related by the that capture our beliefs about the likeli-
are shown within parentheses each esti- following linear model: hood of different states at time t before
mate in Figure 6b, and these are used to and after fusion with the measurement,
weight estimates when fusing them. (32) respectively. The mean and covariance
We first present the state evolution matrix of a random variable xi|j are
model and a priori state estimation. Here, ut is the control input, which denoted by and ∑i|j, respectively. We
Then we discuss how state estimates is assumed to be deterministic, assume (no bias).
are fused if an estimate of the entire and wt is a zero-mean noise term Prediction essentially uses xt−1|t−1 as
state can be obtained by measurement. that models all the uncertainty in a proxy for xt−1 in Equation 32 to deter-
Finally, we discuss how to address this the system. The covariance matrix mine xt|t−1 as shown in Equation 33.
(33) with zt, which is used in fusion, are eas- Figure 6d shows the computation
ily proved to be correct by induction on for this case. The fusion phase can be
For state estimation, we need only the t, using Lemma 2(ii). Figure 6b gives the understood intuitively in terms of the
mean and covariance matrix of xt|t−1. intuition: xt|t−1 for example is an affine following steps.
The predictor box in Figure 6 computes function of the random variables x0|0, w1,
these values; the covariance matrix v1, w2, v2, …, wt, and is therefore uncor- i. The observable part of the a pri-
is obtained from Lemma 2 under the related with vt (by assumption about vt ori estimate of the state
assumption that wt is uncorrelated with and Lemma 2(ii) ) and hence with zt. is fused with the measurement
xt−1|t−1, which is justified here. Figure 7 shows the computation picto- (zt), using Equations 20–22.
Fusing complete observations rially using confidence ellipses to illus- The quantity is
of the state. If the entire state can trate uncertainty. The dotted arrows at called the innovation. The result is
be measured at each time step, the the bottom of the figure show the evolu- the a posteriori estimate of the
imprecise measurement at time t is tion of the state, and the solid arrows show observable state .
modeled as follows: the computation of the a priori estimates ii. The BLUE of Theorem 4 is used to
and their fusion with measurements. obtain the a posteriori estimate
(34) Fusing partial observations of the of the hidden state by adding
state. In some problems, only a portion to the a priori estimate of the hid-
where vt is a zero-mean noise term with of the state can be measured directly. den state a value obtained
covariance matrix Rt. The noise terms The observable portion of the state is from the product of the covariance
in different time steps are assumed to specified formally using a full row-rank between Htxt|t–1 and Ctxt|t–1 and the
be uncorrelated with each other (such matrix Ht called the observation matrix: difference between and .
as, E[vivj] is zero if i≠j) as well as with if the state is x, what is observable is Htx. iii. The a posteriori estimates of the
x0|0 and all wk. A subtle point here is that For example, if the state vector has two observable and hidden portions
xt in this equation is the actual state of components and only the first component of the state are composed to pro-
the system at time t (that is, a particular is observable, Ht can be [1 0]. In general, the duce the a posteriori estimate of
realization of the random variable xt), Ht matrix can specify a linear relationship the entire state .
so variability in zt comes only from vt between the state and the observation,
and its covariance matrix Rt. and it can be time-dependent. The The actual implementation pro-
Therefore, we have two imprecise imprecise measurement model intro- duces the final result directly without
estimates for the state at each time step duced in Equation 34 becomes: going through these steps as shown in
t = 1, 2, …, the a priori estimate from the Figure 6d, but these incremental steps
predictor and the one from the (35) are useful for understanding how all
measurement (zt). If zt is uncorrelated to this works, and their implementation
xt|t−1, we can use Equations 20–22 to The hidden portion of the state is shown in more detail in Figure 8.
fuse the estimates as shown in Figure 6c. can be specified using a matrix Ct, Figure 6d puts all this together.
The assumptions that (i) xt−1|t−1 is which is an orthogonal complement of In the literature, this dataflow is
uncorrelated with wt, which is used in Ht. For example, if Ht = [1 0], one choice referred to as Kalman filtering.
prediction, and (ii) xt|t−1 is uncorrelated for Ct is [0 1]. Unlike in Equations 18 and 21, the
Velocity (m/s)
Distance (m)
Distance (m)
1500 1500
Variance
Variance
150 150 75 75
1000 1000
100 100 50 50
500 500
50 50 25 25
0 0 0 0 0 0
0 0 5 5 10 10 15 15 20 20 0 0 5 5 10 10 15 15 20 20
Time (s) Time (s) Time (s) Time (s)
(a) Evolution
(a) of
Evolution
state: Distance
of state: Distance (b) Evolution
(b) of
Evolution
state: Velocity
of state: Velocity
(MVUE) ). This result is proved using a nonlinear transformation (y = g(x) ). Process. Mag. 5, 29 (2012), 128–132.
11. Grewal, M.S., Andrews, A.P. Kalman Filtering: Theory and
the Cramer-Rao lower bound.24 The random variable x is sampled Practice with MATLAB, 4th edn. Wiley-IEEE Press, 2014.
using a carefully chosen set of sigma 12. Hess, A.-K., Rantzer, A. Distributed Kalman filter
algorithms for self-localization of mobile devices.
Extension to Nonlinear Systems points and these sample points are In Proceedings of the 13th ACM International
The extended Kalman filter (EKF) and propagated through the nonlinear Conference on Hybrid Systems: Computation and
Control, HSCC ‘10, 2010, 191–200.
unscented Kalman filter (UKF) are heu- function g. The statistics of y are esti- 13. Imes, C., Kim, D.H.K., Maggio, M., Hoffmann, H. POET:
ristic approaches to using Kalman mated using a weighted sample mean A portable approach to minimizing energy under soft
real-time constraints. In 21st IEEE Real-Time and
filtering for nonlinear systems. The and covariance of the posterior sigma Embedded Technology and Applications Symposium,
state evolution and measurement points. The UKF tends to be more April 2015, 75–86.
14. Imes, C., Hoffmann, H. Bard: A unified framework
equations for nonlinear systems with robust and accurate than the EKF but for managing soft timing and power constraints. In
additive noise can be written as fol- has higher computation overhead International Conference on Embedded Computer
Systems: Architectures, Modeling and Simulation
lows; in these equations, f and h are due to the sampling process. (SAMOS), 2016.
15. Julier, S.J., Uhlmann, J.K. Unscented filtering and non-
nonlinear functions. linear estimation. Proc. IEEE 3, 92 (2004), 401–422.
Conclusion 16. Kitanidis, P.K. Unbiased minimum-variance linear
state estimation. Automatica 6, 23 (1987), 775–778.
(42) In this article, we have shown that 17. Lindquist, A., Picci, G. Linear Stochastic Systems.
two concepts—optimal linear estima- Springer-Verlag, 2017.
(43) 18. Maybeck, P.S. Stochastic Models, Estimation, and
tors for fusing uncorrelated estimates Control, volume 3. Academic Press, 1982.
and best linear unbiased estimators 19. Mendel, J.M. Lessons in Estimation Theory for Signal
Processing, Communications, and Control. Pearson
Intuitively, the EKF constructs for correlated variables—provide the Education, 1995.
linear approximations to the nonlin- underpinnings for Kalman filtering. 20. Nagarajan, K., Gans, N., Jafari, R. Modeling human gait
using a Kalman filter to measure walking distance. In
ear functions f and h and applies the By combining these ideas, standard Proceedings of the 2nd Conference on Wireless Health,
Kalman filter equations, whereas the results on Kalman filtering for linear WH ‘11 (New York, NY, USA, 2011). ACM, 34:1–34:2.
21. Nakamura, E.F., Loureiro, A.A.F., Frery, A.C. Information
UKF constructs approximations to systems can be derived in an intuitive fusion for wireless sensor networks: methods, models,
probability distributions and propa- and straightforward way that is simpler and classifications. ACM Comput. Surv. 3, 39 (2007).
22. Petersen, K.B., Pedersen, M.S. The Matrix Cookbook.
gates these through the nonlinear than other presentations of this mate- http://www2.imm.dtu.dk/pubdb/views/publication_
functions to construct approximations rial in the literature. This approach details.php?id=3274. November 2012. Version 20121115.
23. Pothukuchi, R.P., Ansari, A., Voulgaris, P., Torrellas, J.
to the posterior distributions. makes clear the assumptions that Using multiple input, multiple output formal control
to maximize resource efficiency in architectures. In
EKF. Examining Figure 6d, we see underlie the optimality results associ- Proceedings of the 2016 ACM/IEEE 43rd Annual
that the a priori state estimate in the pre- ated with Kalman filtering and should International Symposium on Computer Architecture
(ISCA) (2016), IEEE, 658–670.
dictor can be computed using the sys- make it easier to apply Kalman filtering 24. Rao, C.R. Information and the accuracy attainable
tem model: . However, to problems in computer systems. in the estimation of statistical parameters. Bull.
Calcutta Math. Soc., 37 (1945), 81–89.
as the system dynamics and measure- 25. Rhudy, M.B., Salguero, R.A., Holappa, K. A Kalman
ment equations are nonlinear, it is not Acknowledgments filtering tutorial for undergraduate students. Int. J.
Comp. Sci. Eng. Surv. (1), 8 (2017).
clear how to compute the co-variance This research was supported by NSF 26. Sengupta, S.K. Fundamentals of statistical signal
matrices for the a priori estimate and grants 1337281, 1406355, and 1618425, processing: Estimation theory. Technometrics (4), 37
(1995), 465–466.
the measurement. In the EKF, these and by DARPA contracts FA8750-16- 27. Souza, É.L., Nakamura, E.F., Pazzi, R.W. Target
matrices are computed by linearizing 2-0004 and FA8650-15-C-7563. We are tracking for sensor networks: A survey. ACM Comput.
Surv. (2), 49 (2016), 30:1–30:31.
Equations 42 and 43 using the Taylor indebted to K. Mani Chandy (Caltech), 28. Thrun, S., Burgard, W., Fox, D. Probabilistic Robotics
series expansions for the nonlinear Ivo Babuska (UT Austin,) and Augusto (Intelligent Robotics and Autonomous Agents). The
MIT Press, 2005.
functions f and h. This requires comput- Ferrante (Padova, Italy) for their valu- 29. Welch, G., Bishop, G. An Introduction to the Kalman
ing the following Jacobians,e which play able feedback. Filter. Technical Report. Chapel Hill, NC, USA, 1995.
30. Pei, Y., Biswas, S., Fussell, D.S., Pingali, K. An
the role of Ft and Ht in Figure 6d. Elementary Introduction to Kalman Filtering. ArXiv
References e-prints. October 2017.
1. Babb, T. How a Kalman filter works, in pictures
| bzarg. 2018. https://www.bzarg.com/p/how-a-
kalman-filter-works-in-pictures/. Accessed: 2018-11-30 The online appendix for this article can be found at http://
2. Balakrishnan, A.V. Kalman Filtering Theory. Optimization dl.acm.org/citation.cfm?doid=3363294&picked=formats
The EKF performs well in some Software, Inc., Los Angeles, CA, USA, 1987.
3. Barker, A.L., Brown, D.E., Martin, W.N. Bayesian
applications such as navigation systems Estimation and the Kalman Filter. Technical Report. Yan Pei (ypei@cs.utexas.edu) is a graduate research
and GPS.28 Charlottesville, VA, USA, 1994. assistant in the Department of Computer Science at the
4. Becker, A. Kalman filter overview. https://www. University of Texas, Austin, TX, USA.
UKF. When the system dynamics kalmanfilter.net/default.aspx. 2018. Accessed: 2018-11-08.
Swarnendu Biswas (swarnendu@cse.iitk.ac.in) is
and observation models are highly 5. Bergman, K. Nanophotonic interconnection networks
an assistant professor in the Department of Computer
in multicore embedded computing. In 2009 IEEE/
nonlinear, the unscented Kalman fil- LEOS Winter Topicals Meeting Series (2009), 6–7. Science and Engineering at the Indian Institute of
Technology, Kanpur, India.
ter (UKF)15 can be an improvement 6. Cao, L., Schwartz, H.M. Analysis of the Kalman
filter based estimation algorithm: An orthogonal Donald S. Fussell (fussell@cs.utexas.edu) is the Trammell
over the EKF. The UKF is based on the decomposition approach. Automatica 1, 40 (2004), 5–19. Crow Regents Professor in the Department of Computer
unscented transformation, which is a 7. Chui, C.K., Chen, G. Kalman Filtering: With Real-Time Science at the University of Texas, Austin, TX, USA.
Applications, 5th edn. Springer Publishing Company,
method for computing the statistics Incorporated, 2017. Keshav Pingali (pingali@cs.utexas.edu) is a professor in
of a random variable x that undergoes 8. Eubank, R.L. A Kalman Filter Primer (Statistics: Textbooks the Department of Computer Science at the University
and Monographs). Chapman & Hall/CRC, 2005. of Texas, Austin, and the W.A.”Tex” Moncrief Chair of
9. Evensen, G. Data Assimilation: The Ensemble Kalman Computing in the UT Oden Institute of Computational
Filter. Springer-Verlag New York, Inc., Secaucus, NJ, Engineering and Science, Austin, TX, USA.
e The Jacobian matrix is the matrix of all first USA, 2006.
order partial derivatives of a vector-valued 10. Faragher, R. Understanding the basis of the Kalman
function. filter via a simple and intuitive derivation. IEEE Signal © 2019 ACM 0001-0782/19/11 $15.00
Technical
To view the accompanying paper,
visit doi.acm.org/10.1145/3361566 rh
Perspective
A Whitebox Solution for
Blackbox-Like Behaviors
By David G. Andersen
Abstract expected the bus to yield under a set of rare conditions but
Deep learning (DL) systems are increasingly deployed in the bus did not.a
safety- and security-critical domains such as self-driving A Tesla car in autopilot crashed into a trailer because the
cars and malware detection, where the correctness and pre- autopilot system failed to recognize the trailer as an obsta-
dictability of a system’s behavior for corner case inputs are cle due to its “white color against a brightly lit sky” and the
of great importance. Existing DL testing depends heavily on “high ride height”.
manually labeled data and therefore often fails to expose b
Such corner cases were not part of Google’s or Tesla’s
erroneous behaviors for rare inputs. test set and thus never showed up during testing.
We design, implement, and evaluate DeepXplore, the first Therefore, DL systems, just like traditional software,
white-box framework for systematically testing real-world must be tested systematically for different corner cases to
DL systems. First, we introduce neuron coverage for measur- detect and fix ideally any potential flaws or undesired behav-
ing the parts of a DL system exercised by test inputs. Next, we iors. This presents a new system problem as automated and
leverage multiple DL systems with similar functionality as systematic testing of large-scale, real-world DL systems with
cross-referencing oracles to avoid manual checking. Finally, thousands of neurons and millions of parameters for all cor-
we demonstrate how finding inputs for DL systems that ner cases is extremely challenging.
both trigger many differential behaviors and achieve high The standard approach for testing DL systems is to gather
neuron coverage can be represented as a joint optimization and manually label as much real-world test data as possible.
problem and solved efficiently using gradient-based search Some DL systems such as Google self-driving cars also use
techniques. simulation to generate synthetic training data. However,
DeepXplore efficiently finds thousands of incorrect cor- such simulation is completely unguided as it does not con-
ner case behaviors (e.g., self-driving cars crashing into sider the internals of the target DL system. Therefore, for
guard rails and malware masquerading as benign soft- the large input spaces of real-world DL systems (e.g., all pos-
ware) in state-of-the-art DL models with thousands of neu- sible road conditions for a self-driving car), none of these
rons trained on five popular datasets such as ImageNet approaches can hope to cover more than a tiny fraction (if
and Udacity self-driving challenge data. For all tested DL any at all) of all possible corner cases.
models, on average, DeepXplore generated one test input Recent works on adversarial deep learning3 have
demonstrating incorrect behavior within one second while demonstrated that carefully crafted synthetic images by
running only on a commodity laptop. We further show that adding minimal perturbations to an existing image can
the test inputs generated by DeepXplore can also be used to fool state-of-the-art DL systems. The key idea is to cre-
retrain the corresponding DL model to improve the model’s ate synthetic images such that they get classified by DL
accuracy by up to 3%. models differently than the original picture but still
look the same to the human eye. Although such adver-
sarial images expose some erroneous behaviors of a DL
1. INTRODUCTION model, the main restriction of such an approach is that
Over the past few years, Deep Learning (DL) has made it must limit its perturbations to tiny invisible changes
tremendous progress, achieving or surpassing human-level and require ground truth labels. Moreover, just like other
performance for a diverse set of tasks in many application forms of existing DL testing, the adversarial images only
domains. These advances have led to widespread adoption cover a small part (52.3%) of DL system’s logic as shown
and deployment of DL in security- and safety-critical sys- in Section 5. In essence, the current machine learning
tems such as self-driving cars,1 malware detection,4 and air- testing practices for finding incorrect corner cases are
craft collision avoidance systems.6 analogous to finding bugs in traditional software by using
This wide adoption of DL techniques presents new chal-
lenges as the predictability and correctness of such sys-
tems are of crucial importance. Unfortunately, DL systems,
a
http://www.theverge.com/2016/2/29/11134344/google-self-driving-car-
crash-report
despite their impressive capabilities, often demonstrate b
https://electrek.co/2016/07/01/understanding-fatal-tesla-accident-autopilot-
unexpected or incorrect behaviors in corner cases for sev- nhtsa-probe/
eral reasons such as biased training data and overfitting of
the models. In safety- and security-critical settings, such
The original version of this paper was published in
incorrect behaviors can lead to disastrous consequences
Proceedings of the 26th Symposium on Operating Systems
such as a fatal collision of a self-driving car. For example, a
Principles (Shanghai, China, Oct. 28–31, 2017), 1–18.
Google self-driving car recently crashed into a bus because it
test inputs with low code coverage and thus are unlikely to as Udacity self-driving car challenge data, image data from
find many erroneous cases. ImageNet and MNIST, Android malware data from Drebin,
The key challenges in automated systematic testing of and PDF malware data from Contagio/VirusTotal. For all of
large-scale DL systems are twofold: (1) how to generate the tested DL models, on average, DeepXplore generated one
inputs that trigger different parts of a DL system’s logic test input demonstrating incorrect behavior within one sec-
and uncover different types of erroneous behaviors, and ond while running on a commodity laptop. The inputs gener-
(2) how to identify erroneous behaviors of a DL system ated by DeepXplore achieved 34.4 and 33.2% higher neuron
without manual labeling/checking. This paper describes coverage on average than the same number of randomly
and highlights how we design and build DeepXplore to picked inputs and adversarial inputs,3 respectively. We fur-
address both challenges. ther show that the test inputs generated by DeepXplore can
First, we introduce the concept of neuron coverage for be used to retrain the corresponding DL model to improve
measuring the parts of a DL system’s logic exercised by a set classification accuracy as well as identify potentially polluted
of test inputs based on the number of neurons activated by training data. We achieve up to 3% improvement in classifica-
the inputs. At a high level, neuron coverage of DL systems is tion accuracy by retraining a DL model on inputs generated
similar to code coverage of traditional systems, a standard by DeepXplore compared to retraining on the same number
empirical metric for measuring the amount of code exer- of random or adversarial inputs.
cised by an input in the traditional software. However, code A number of follow-up papers after DeepXplore have
coverage itself is not a good metric for estimating coverage expanded the idea of whitebox testing for domain-specific
of DL systems as most rules in DL systems, unlike tradi- transformations in self-driving cars13 and developed exhaus-
tional software, are not written manually by a programmer tive black box testing techniques for a variety of common
but rather learned from training data. In fact, we find that transformations.11 Besides, the metric of neuron coverage
for most of the DL systems that we tested, even a single ran- has also been extended in TensorFlow as an open-source
domly picked test input was able to achieve 100% code cover- tool by the Google Brain team.10 Beyond testing, we have also
age, whereas the neuron coverage was less than 10%. studied and proposed more rigorous verification techniques
Next, we show how multiple DL systems with similar leveraging interval arithmetic to certify the robustness of
functionality (e.g., self-driving cars by Google and Tesla, neural networks15, 16 (Figure 1).
and GM) can be used as cross-referencing oracles to iden-
tify erroneous corner cases without providing ground truth 2. BACKGROUND
labels which require huge manual labeling effort. For exam- 2.1. DL systems
ple, if one self-driving car decides to turn left whereas oth- We define a DL system to be any software system that
ers turn right for the same input, one of them is likely to includes at least one Deep Neural Network (DNN) compo-
be incorrect. Such techniques have been applied success- nent. Note that some DL systems might comprise solely
fully in the past for detecting logic bugs without manual DNNs (e.g., self-driving car DNNs predicting steering angles
specifications in a wide variety of traditional software.2 In without any manual rules), whereas others may have some
this paper, we demonstrate how differential testing can be DNN components interacting with other traditional soft-
applied to DL systems. ware components to produce the final output.
Finally, we demonstrate how the problem of generating As shown in Figure 2, the development process of the
test inputs that maximize neuron coverage of a DL system DNN components of a DL system is fundamentally differ-
while also exposing as many differential behaviors (i.e., dif- ent from traditional software development. Unlike tradi-
ferences between multiple similar DL systems) as possible tional software, where the developers directly specify the
can be formulated as a joint optimization problem. Unlike logic of the system, the DNN components learn their rules
traditional programs, Deep Neural Networks (DNNs) used
by DL systems are differentiable. Therefore, their gradients
with respect to inputs can be calculated accurately given Figure 1. An example of erroneous behavior found by DeepXplore in
whitebox access to the corresponding model. In this paper, Nvidia DAVE-2 self-driving car platform. The DNN-based self-driving
car correctly decides to turn left for image (a) but incorrectly decides
we show how these gradients can be used to efficiently solve to turn right and crashes into the guardrail for image (b), a slightly
the joint optimization problem for large-scale real-world darker version of (a).
DL systems.
We design, implement, and evaluate DeepXplore, to the
best of our knowledge, the first efficient whitebox testing
framework for large-scale DL systems. In addition to maxi-
mizing neuron coverage and behavioral differences between
DL systems, DeepXplore also supports adding custom con-
straints by the users for simulating different types of realis-
tic inputs (e.g., different types of lighting and occlusion for
images/videos). We demonstrate that DeepXplore efficiently
finds thousands of unique incorrect corner case behaviors
(e.g., self-driving cars crashing into guard rails) in 15 state-of- (a) Input 1 (b) Input 2 (darker version of 1)
the-art DL models trained using five real-world datasets such
(k)
(1) (2)
W4 Activation functionσ ... ...
W W I4
/* no bugs */ /* buggy code */ Car ... Face
Function modeled by DNN: ... ... 0.95 0
f (x) = σ(W(2) • σ(W(1) • x)) Individual neurons in layer k (a) A program with a rare branch (b) A DNN for detecting cars and faces
classifiers (with state-of-the-art performance on randomly for erroneous corner case behaviors. The main components
picked testing sets) still incorrectly classify synthetic images of DeepXplore are shown in Figure 5. DeepXplore takes
generated by adding humanly imperceptible perturbations to unlabeled test inputs as seeds and generates new tests that
a test image.3 However, the adversarial inputs, similar to random cover a large number of neurons (i.e., activates them to a
test inputs, also only cover a small part of the rules learned value above a customizable threshold) while causing the
by a DNN as they are not designed to maximize coverage. tested DNNs to behave differently. Specifically, DeepXplore
Moreover, they are also inherently limited to small impercep- solves a joint optimization problem that maximizes both
tible perturbations around a test input as larger perturbations differential behaviors and neuron coverage. Note that both
will visually change the input and therefore will require manual goals are crucial for thorough testing of DNNs and find-
inspection to ensure correctness of the DNN’s decision. ing diverse erroneous corner case behaviors. High neuron
Problems with low-coverage DNN tests. To better under- coverage alone may not induce many erroneous behaviors,
stand the problem of low test coverage of rules learned by a whereas just maximizing different behaviors might simply
DNN, we provide an analogy to a similar problem in testing identify different manifestations of the same underlying
traditional software. Figure 4 shows a side-by-side compari- root cause.
son of how a traditional program and a DNN handle inputs DeepXplore also supports enforcing of custom domain-
and produce outputs. Specifically, the figure shows the simi- specific constraints as part of the joint optimization process.
larity between traditional software and DNNs: in software For example, the value of an image pixel has to be between
program, each statement performs a certain operation to 0 and 255. Such domain-specific constraints can be speci-
transform the output of previous statement(s) to the input fied by the users of DeepXplore to ensure that the generated
to the following statement(s), whereas in DNN, each neuron test inputs are valid and realistic.
transforms the output of previous neuron(s) to the input of We designed an algorithm for efficiently solving the
the following neuron(s). Of course, unlike traditional soft- joint optimization problem mentioned above using gradi-
ware, DNNs do not have explicit branches but a neuron’s ent ascent. First, we compute the gradient of the outputs
influence on the downstream neurons decreases as the neu- of the neurons in both the output and hidden layers with
ron’s output value gets lower. A lower output value indicates the input value as a variable and the weight parameter as a
less influence and vice versa. When the output value of a constant. Such gradients can be computed efficiently for
neuron becomes zero, the neuron does not have any influ- most DNNs. Note that DeepXplore is designed to operate
ence on the downstream neurons. on pretrained DNNs. The gradient computation is efficient
As demonstrated in Figure 4a, the problem of low cover- because our whitebox approach has access to the pretrained
age in testing traditional software is obvious. In this case, DNNs’ weights and the intermediate neuron values. Next,
the buggy behavior will never be seen unless the test input we iteratively perform gradient ascent to modify the test
is 0xdeadbeef. The chances of randomly picking such input toward maximizing the objective function of the joint
a value are very small. Similarly, low-coverage test inputs optimization problem described above. Essentially, we per-
will also leave different behaviors of DNNs unexplored. For form a gradient-guided local search starting from the seed
example, consider a simplified neural network, as shown inputs and find new inputs that maximize the desired goals.
in Figure 4b, that takes an image as an input and classifies Note that, at a high level, our gradient computation is simi-
it into two different classes: cars and faces. The text in each lar to the backpropagation performed during the training of
neuron (represented as a node) denotes the object or prop- a DNN, but the key difference is that, unlike our algorithm,
erty that the neuron detects,c and the number in each neu- backpropagation treats the input value as a constant and the
ron is the real value outputted by that neuron. The number weight parameter as a variable.
indicates how confident the neuron is about its output. A working example. We use Figure 6 as an example to
Note that randomly picked inputs are highly unlikely to show how DeepXplore generates test inputs. Consider that
set high output values for the unlikely combination of we have two DNNs to test—both perform similar tasks, that
neurons. Therefore, many incorrect DNN behaviors will is, classifying images into cars or faces, as shown in Figure 6,
remain unexplored even after performing a large number but they are trained independently with different datasets
of random tests. For example, if an image causes neurons and parameters. Therefore, the DNNs will learn similar but
labeled as “Nose” and “Red” to produce high output values slightly different classification rules. Let us also assume that
and the DNN misclassifies the input image as a car, such
a behavior will never be seen during regular testing as the
Figure 5. DeepXplore workflow.
chances of an image containing a red nose (e.g., a picture
of a clown) are very small. DNNs
Unlabeled under Gradients of Difference-
seed test output & hidden inducing
inputs neurons Objective: maximize inputs
3. OVERVIEW DNN1 differences & neuron coverage
t1 t1'
In this section, we provide a general overview of DeepXplore, t2
Joint optimization with t2'
t3 DNN2 t3'
our whitebox framework for systematically testing DNNs ...
gradient ascent
... ...
DNNn
c
Note that one cannot always map each neuron to a particular task, i.e.,
detecting specific objects/properties. Figure 4b simply highlights that dif- Domain-specific constraints
ferent neurons often tend to detect different features.
Given an arbitrary x as seed that gets classified to the same Our algorithm ensures this property by modifying the gradi-
class by all DNNs, our goal is to modify x such that the modi- ent grad such that xi+1 = xi + s . grad still satisfies the con-
fied input x′ will be classified differently by at least one of straints (s is the step size in the gradient ascent).
the n DNNs. For discrete features, we round the gradient to an inte-
Let Fk(x)[c] be the class probability that Fk predicts x to be ger. For DNNs handling visual input (e.g., images), we add
c. We randomly select one neural network Fj and maximize different spatial restrictions such that only part of the input
the following objective function: images is modified. A detailed description of the domain-
specific constraints that we implemented can be found in
(2) Section 5.2.
Hyperparameters. To summarize, there are four major
where λ1 is a parameter to balance the objective terms hyperparameters that control different aspects of DeepXplore
between the DNNs’ Fk≠j that maintain the same class out- as described below. (1) λ1 balances the objectives between
puts as before and the DNN Fj that produce different class minimizing one DNN’s prediction for a certain label and
outputs. As all of Fk∈1..n are differentiable, Equation 2 can be maximizing the rest of DNNs’ predictions for the same
maximized using gradient ascent by iteratively changing x label. Larger λ1 puts higher priority on lowering the pre-
based on the computed gradient: . diction value/confidence of a particular DNN, whereas
Maximizing neuron coverage. The second objective is smaller λ1 puts more weight on maintaining the other
to generate inputs that maximize neuron coverage. We DNNs’ predictions. (2) λ2 provides balance between find-
achieve this goal by iteratively picking inactivated neurons ing differential behaviors and neuron coverage. Larger
and modifying the input such that the output of that neuron λ2 focuses more on covering different neurons, whereas
goes above the neuron activation threshold. Let us assume smaller λ2 generates more difference-inducing test inputs.
that we want to maximize the output of a neuron n, that is, (3) s controls the step size used during iterative gradient
we want to maximize obj2(x) = fn(x) such that fn(x) > t, where ascent. Larger s may lead to oscillation around the local
t is the neuron activation threshold, and we write fn(x) as optimum, whereas smaller s may need more iterations
the function modeled by neuron n that takes x (the original to reach the objective. (4) t is the threshold to determine
input to the DNN) as the input and produce the output of whether each individual neuron is activated or not. Finding
neuron n (as defined in Equation 1). We can again leverage inputs that activate a neuron becomes increasingly harder
the gradient ascent mechanism as fn(x) is a differentiable as t increases.
function whose gradient is .
Note that we can also potentially jointly maximize mul- 5. EXPERIMENTAL SETUP
tiple neurons simultaneously, but we choose to activate one 5.1. Test datasets and DNNs
neuron at a time in this algorithm for clarity. We adopt 5 popular public datasets with different types of
Joint optimization. We jointly maximize obj1 and fn data—MNIST, ImageNet, Driving, Contagio/VirusTotal, and
described above and maximize the following function: Drebin—and then evaluate DeepXplore on 3 DNNs for each
dataset (i.e., a total of 15 DNNs). We provide a summary of
(3) the five datasets and the corresponding DNNs in Table 1.
The detailed description can be found in the full paper. All
where λ2 is a parameter for balancing between the two objec- the evaluated DNNs are either pretrained (i.e., we use public
tives and n is the inactivated neuron that we randomly pick weights reported by previous researchers) or trained by us
at each iteration. As all terms of objjoint are differentiable, we using public real-world architectures to achieve comparable
jointly maximize them using gradient ascent by modifying x. performance to that of the state-of-the-art models for the
Domain-specific constraints. One important aspect of corresponding dataset. For each dataset, we used DeepXplore
the optimization process is that the generated test inputs to test three DNNs with different architectures.
need to satisfy several domain-specific constraints to be
physically realistic. In particular, we want to ensure that the 5.2. Domain-specific constraints
changes applied to xi during the ith iteration of gradient As discussed earlier, to be useful in practice, we need to
ascent process satisfy all the domain-specific constraints for ensure that the generated tests are valid and realistic by
all i. For example, for a generated test image x, the pixel val- applying domain-specific constraints. For example, gener-
ues must be within a certain range (e.g., 0–255). ated images should be physically producible by a camera.
Although some such constraints can be efficiently Similarly, generated PDFs need to follow the PDF speci-
embedded into the joint optimization process using the fication to ensure that a PDF viewer can open the test file.
Lagrange Multipliers similar to those used in support vec- Below we describe two major types of domain-specific con-
tor machines, we found that the majority of them cannot be straints (i.e., image and file constraints) that we use in this
easily handled by the optimization algorithm. Therefore, we paper. Image constraints (MNIST, ImageNet, and Driving).
designed a simple rule-based method to ensure that the gen- DeepXplore used three different types of constraints for
erated tests satisfy the custom domain-specific constraints. simulating different environmental conditions of images:
As the seed input xseed = x0 always satisfies the constraints by (1) lighting effects for simulating different intensities of
definition, our technique must ensure that after the ith (i > 0) lights, (2) occlusion by a single small rectangle for simulat-
iteration of gradient ascent, xi still satisfies the constraints. ing an attacker potentially blocking some parts of a camera,
Dataset Reported
Dataset description DNN description DNN name # of neurons Architecture Acc. Our Acc.
MNIST Hand-written LeNet variations MNI_C1 52 LeNet-1, LeCun et al. [8] 98.3% 98.33%
digits MNI_C2 148 LeNet-4, LeCun et al. [8] 98.9% 98.59%
MNI_C3 268 LeNet-5, LeCun et al. [8] 99.05% 98.96%
Imagenet General State-of-the-art IMG_C1 14,888 VGG-16, Simonyan et al. [12] 92.6%** 92.6%**
images image classifiers IMG_C2 16,168 VGG-19, Simonyan et al. [12] 92.7%** 92.7%**
from ILSVRC IMG_C3 94,059 ResNet50, He et al. [5] 96.43%** 96.43%**
Driving Driving video Nvidia DAVE DRV_C1 1,560 Dave-orig [1] N/A 99.91%#
frames self-driving DRV_C2 1,560 Dave-norminit## N/A 99.94%#
systems DRV_C3 844 Dave-dropout++ N/A 99.96%#
Contagio/ PDFs PDF malware PDF_C1 402 <200, 200>+ 98.5%− 96.15%
Virustotal detectors PDF_C2 602 <200, 200, 200>+ 98.5%− 96.25%
PDF_C3 802 <200, 200, 200, 200>+ 98.5%− 96.47%
Drebin Android apps Android app APP_C1 402 <200, 200>+, Grosse et al. [4] 98.92% 98.6%
malware APP_C2 102 <50, 50>+, Grosse et al. [4] 96.79% 96.82%
detectors APP_C3 212 <200, 10>+, Grosse et al. [4] 92.97% 92.66%
**
Top-5 test accuracy; we exactly match the reported performance as we use the pretrained networks.
# We report 1-MSE (mean squared error) as the accuracy because steering angle is a continuous value.
+ <x,y,…> denotes three hidden layers with x neurons in first layer, y neurons in second layer, etc.
− Accuracy using SVM as reported by Šrndic et al. [14].
## https://github.com/jacobgil/keras-steering-angle-visualizations.
++ https://github.com/navoshta/behavioral-cloning.
and (3) occlusion by multiple tiny black rectangles for simu- Table 2. Number of difference-inducing inputs found by DeepXplore
lating effects of dirt on camera lens. for each tested DNN obtained by randomly selecting 2000 seeds
Other constraints (Drebin and Contagio/VirusTotal). from the corresponding test set for each run.
For Drebin dataset, DeepXplore enforces a constraint that
only allows modifying features related to the Android mani- Hyperparams
# Differences
fest file and thus ensures that the application code is unaf- DNN name λ1 λ2 s t found
fected. Moreover, DeepXplore only allows adding features MNI_C1 1 0.1 10 0 1073
(changing from zero to one) but does not allow deleting fea- MNI_C2 1968
tures (changing from one to zero) from the manifest files to MNI_C3 827
IMG_C1 1 0.1 10 0 1969
ensure that no application functionality is changed due to
IMG_C2 1976
insufficient permissions. Thus, after computing the gradi- IMG_C3 1996
ent, DeepXplore only modifies the manifest features whose DRV_C1 1 0.1 10 0 1720
corresponding gradients are greater than zero. For Contagio/ DRV_C2 1866
VirusTotal dataset, we follow the restrictions on each feature DRV_C3 1930
as described by Šrndic and Laskkov.14 PDF_C1 2 0.1 0.1 0 1103
PDF_C2 789
PDF_C3 1253
6. RESULTS APP_C1 1 0.5 N/A 0 2000
6.1. Summary APP_C2 2000
Table 2 summarizes the numbers of erroneous behaviors APP_C3 2000
found by DeepXplore for each tested DNN while using 2000
randomly selected seed inputs from the corresponding test
sets. Note that as the testing set has a similar number of in Table 2 in all the experiments unless otherwise specified.
samples for each class, these randomly-chosen 2000 sam- Figure 7 shows some difference-inducing inputs generated
ples also follow that distribution. The hyperparameters for by DeepXplore for MNIST, ImageNet, and Driving dataset
these experiments, as shown in Table 2, are empirically cho- along with the corresponding erroneous behaviors. Table 3
sen to maximize both the rate of finding difference-inducing (Drebin) and Table 4 (Contagio/VirusTotal) show two sample
inputs as well as the neuron coverage. difference-inducing inputs generated by DeepXplore that
For the experimental results shown in Figure 7, we apply caused erroneous behaviors in the tested DNNs. We high-
three domain-specific constraints (lighting effects, occlusion light the differences between the seed input features and
by a single rectangle, and occlusion by multiple rectangles) the features modified by DeepXplore. Note that we only list
as described in Section 5.2. For all other experiments involv- the top three modified features due to space limitations.
ing vision-related tasks, we only use the lighting effects as the
domain-specific constraints. For all malware-related experi- 6.2. Benefits of neuron coverage
ments, we apply all the relevant domain-specific constraints In this subsection, we evaluate how effective neuron cover-
described in Section 5.2. We use the hyperparameters listed age is in measuring the comprehensiveness of DNN testing.
Figure 7. The first row shows the seed test inputs and the second row shows the difference-inducing test inputs generated by DeepXplore.
The left three columns show results under different lighting effects, the middle three are using single occlusion box, and the right three are
using black rectangles as the transformation constraints. For each type of transformation (three pairs of images), the images from left to
right are from self-driving car, MNIST, and ImageNet.
Table 3. The features added to the manifest file for generating two Table 5. Comparison of code coverage and neuron coverage for
malware inputs that Android app classifiers (Drebin) incorrectly 10 randomly selected inputs from the original test set of each DNN.
mark as benign.
Code coverage Neuron coverage
input 1 feature feature:: activity:: service_
Dataset C1 C2 C3 C1 C2 C3
bluetooth .SmartAlertTerms receiver::
.rrltpsi MNIST 100% 100% 100% 32.7% 33.1% 25.7%
before 0 0 0 ImageNet 100% 100% 100% 1.5% 1.1% 0.3%
after 1 1 1 Driving 100% 100% 100% 2.5% 3.1% 3.9%
input 2 feature provider:: permission:: provider:: VirusTotal 100% 100% 100% 19.8% 17.3% 17.3%
xclockprovider CALL_PHONE contentprovider Drebin 100% 100% 100% 16.8% 10% 28.6%
before 0 0 0
after 1 1 1
Avg. no. of
Total neurons activated neurons Avg. overlap
Watch the authors discuss
Diff. class 268 83.6 45.9 their work in this exclusive
Same class 268 84.1 74.2 Communications video.
https://cacm.acm.org/
videos/deepxplore
Boston College tion. Application review is ongoing. Boston Col- consideration. We will accept and review appli-
Non Tenure-Track Positions in Computer lege conducts background checks as part of the cations after this date.
Science hiring process. Applicants can consult http://www.ee.
Submit a cover letter, a detailed CV and teach- columbia.edu for more information about the
The Computer Science Department of Boston ing and research statements. Arrange for three department and http://pa334.peopleadmin.com/
College seeks to fill one or more non-tenure-track confidential letters of recommendation to be postings/4208 for more details on the position
teaching positions, as well as shorter-term visit- uploaded directly to Interfolio. To apply go to and application.
ing teaching positions. All applicants should be https://apply.interfolio.com/68273. Columbia University is an Equal Opportunity
committed to excellence in undergraduate edu- Boston College is a Jesuit, Catholic university Employer / Disability / Veteran.
cation and be able to teach a broad variety of un- that strives to integrate research excellence with
dergraduate computer science courses. Faculty in a foundational commitment to formative liberal
longer-term positions will also participate in the arts education. We encourage applications from Georgia Institute of Technology
development of new courses that reflect the evolv- candidates who are committed to fostering a di- Multiple Tenure Track Faculty Positions
ing landscape of the discipline. verse and inclusive academic community. Boston
Minimum requirements for the title of As- College is an affirmative action/equal opportu- The School of Computer Science at the Georgia
sistant Professor of the Practice, and for the title nity employer. Institute of Technology (Georgia Tech) invites
of Visiting Assistant Professor, include a Ph.D. in applications for several tenure track faculty po-
Computer Science or closely related discipline. sitions at all ranks. We seek candidates in all ar-
Candidates without a Ph.D. would be eligible Columbia University eas who complement and enhance our current
for the title of Lecturer or Visiting Lecturer. Junior Faculty Position in the Department of research strengths and are especially interested
We will begin reviewing applications on Octo- Electrical Engineering this year in candidates whose research focus is in
ber 15, 2019 and will continue considering appli- the broad area of Theoretical Computer Science.
cations until the positions are filled. Applicants Columbia Engineering is pleased to invite appli- Georgia Tech is an equal education/employ-
should submit a cover letter, CV, and a separate cations for a faculty position in the Department of ment opportunity institution dedicated to build-
teaching statement and arrange for three confi- Electrical Engineering at Columbia University in ing a diverse community. We strongly encourage
dential letters of recommendation that comment the City of New York. Applications for Junior rank applications from women, underrepresented
on their teaching performance to be uploaded will be considered. groups, individuals with disabilities, and veterans.
directly to Interfolio. To apply go to https://apply. The Electrical Engineering department wel- Georgia Tech has policies to promote a healthy
interfolio.com/68339. Boston College conducts comes applications in all areas of electrical engi- work-life balance and is aware that attracting fac-
background checks as part of the hiring process. neering (http://www.ee.columbia.edu/ee-research). ulty may require meeting the needs of two careers.
Information about the university and our de- Candidates must have a Ph.D. or its professional The School of Computer Science, one of three
partment is available at https://www.bc.edu and equivalent by the starting date of the appointment. schools in the top-ten ranked College of Comput-
http:// cs.bc.edu. Applicants for this position must demonstrate the ing, focuses on research that makes computing
Boston College is a Jesuit, Catholic university potential to do pioneering research and to teach ef- and communication smart, fast, reliable, and
that strives to integrate research excellence with fectively. The Department is especially interested in secure, with research groups in computer archi-
a foundational commitment to formative liberal qualified candidates who can contribute, through tecture, databases, machine learning, network-
arts education. We encourage applications from their research, teaching, and/or service, to the di- ing, programming languages, security, software
candidates who are committed to fostering a di- versity and excellence of the academic community. engineering, systems, and theory. Faculty in the
verse and inclusive academic community. Boston The successful candidate is expected to con- school are leaders in a variety of Georgia Tech ini-
College is an affirmative action/equal opportu- tribute to the advancement of their field and the tiatives, including: the Algorithms and Random-
nity employer. department by developing an original and lead- ness Center (ARC), the Center for Research into
ing externally funded research program, and Novel Computing Hierarchies (CRNCH), the In-
to contribute to the undergraduate and gradu- stitute for Data Engineering and Science (IDEaS),
Boston College ate educational mission of the Department. and the Institute for Information Security and
Tenure-Track Assistant Professor of Computer Columbia fosters multidisciplinary research Privacy (IISP). The school is in a period of rapid
Science and encourages collaborations with academic growth with eight tenure-track Assistant Profes-
departments and units across Columbia Uni- sors hired in the last two years.
The Computer Science Department of Boston versity. The Department actively participates in Georgia Tech is a top-ranked public research
College seeks a tenure-track Assistant Professor the school-wide Engineering for Humanity ini- university situated in the heart of Atlanta, a di-
beginning in the 2020-2021 academic year. Suc- tiatives that relate to engineering and medicine, verse and vibrant city with multiple universities.
cessful candidates for the position will be expect- autonomous systems, quantum computing and Midtown Atlanta, where Georgia Tech is located,
ed to develop strong research programs that can technology, and sustainability. has been recognized as one of the 2016 Great
attract external funding in an environment that For additional information and to apply, Neighborhoods by the American Planning As-
also values high-quality undergraduate teaching. please see: http://engineering.columbia.edu/ sociation due to its liveliness, walkability, and
Outstanding candidates in all areas of Computer faculty-job-opportunities. Applications should many great cultural and economic strengths. The
Science will be considered, with a preference for be submitted electronically and include the Institute is a member of the University System of
those who demonstrate a potential to contribute following: curriculum vitae including a pub- Georgia, the Georgia Research Alliance, and the
to cross-disciplinary teaching and research in con- lication list, a description of research ac- Association of American Universities. Georgia
junction with the planned Schiller Institute for complishments, statements of research and Tech prides itself on its technology resources,
Integrated Science and Society at Boston College. teaching interests and plans, contact infor- collaborations, high-quality student body, and its
A Ph.D. in Computer Science or a closely re- mation for three experts who can provide let- commitment to diversity, equity, and inclusion.
lated discipline is required. See http://cs.bc.edu ters of recommendation, and up to three pre/ Applications will be considered until open
and https://www.bc.edu/bc-web/schools/mcas/ reprints of scholarly work. All applications positions are filled. For full consideration, appli-
sites/schiller-institute.html for more informa- received by December 1st, 2019 will receive full cants are encouraged to submit their applications
Illinois Institute of Technology completed a successful capital campaign that led State University in University Park, Pennsylvania
Tenure-Track/Tenured Positions in Computer to the creation of multiple endowed positions, in- invites applications for Associate Dean for Under-
Science creased scholarship funding, the Center for Active graduate and Graduate Studies (ADUGS); although
Computational Thinking, and the new Ed Kaplan the position is a full-time administrative job, the
The Department of Computer Science at the Illi- Family Institute for Innovation and Tech Entre- ADUGS will be a Full or Associate Professor, with
nois Institute of Technology invites applications preneurship. In addition to its rigorous research tenure in IST. In this regard, candidates for Associ-
for multiple tenure-track/tenured faculty positions and education programs, Illinois Tech has a long ate Professor should have a strong track record of
at all ranks, appointments to start in Fall 2020. history of strong partnerships and collaborations research, publication, and funding; those for Full
Applicants must have a Ph.D. in computer with local companies, government labs, and non- should have a track record of research, publica-
science or a closely related field, demonstrated profits; the University Technology Park on campus tion, and funding that distinguishes them as na-
excellence in research, a record of attracting ex- is home to many startups who benefit from close tional or international leaders in their fields.
ternal research funding appropriate to their rank, collaboration with faculty and students. We seek candidates with a strong record of:
and a strong commitment to teaching. We seek Review of applications will start on November graduate and/or undergraduate teaching; gradu-
outstanding candidates in all areas of computer 1, 2019; applications will be reviewed continually ate and/or undergraduate advising; curriculum
science; candidates in cybersecurity, data sci- until all available positions are filled. Illinois In- development; faculty leadership; scholarship and
ence, artificial intelligence, parallel and distrib- stitute of Technology is an EEO/AA/Title VI/Title research. We particularly invite candidates who
uted systems, and programming languages are IX/Section 504/ADA/ADEA employer commit- can contribute to these themes: rapid growth;
especially encouraged to apply. ted to enhancing equity, inclusion and diversity technology innovation; interdisciplinary con-
The Department of Computer Science at the within its community. It actively seeks applica- nections; undergraduate research; diverse stu-
Illinois Institute of Technology offers Bachelors, tions from all individuals regardless of race, dent population. Responsibilities will include:
Masters, and Ph.D. degrees in Computer Science, color, sex, marital status, religion, creed, national strengthen ties to relevant University Offices;
as well as Bachelors and Masters degrees in Arti- origin, disability, age, military or veteran status, facilitate collaboration within and across the
ficial Intelligence, a Masters degree in Cybersecu- sexual orientation, and/or gender identity and college and university; program assessment and
rity, and interdisciplinary Masters degrees in Data expression. All qualified applicants will receive curricular reform; academic integrity; and teach-
Science and in Computational Decision Science equal consideration for employment. Applicants ing and research consonant with administrative
and Operations Research. The department is in should apply online at https://academicjobson- responsibilities.
a significant growth phase, with multiple faculty line.org/ajo/jobs/14362. Successful candidates must have a Ph.D. or
hires per year expected for at least the next few terminal degree in a field relevant to our inter-
years. It is also launching diverse new interdisci- disciplinary faculty (e.g., information and com-
plinary research and education programs, and Pennsylvania State University puter science, psychology, sociology) and must
has strong growing partnerships with Chicago’s Associate Dean for Undergraduate and pass a background check. To apply: submit basic
burgeoning tech community. Illinois Institute Graduate Studies (ADUGS) information via http://psu.jobs/job/90506 and
of Technology, a private, technology-focused re- apply via https://academicjobsonline.org/ajo/
search university, is located just 10 minutes from The College of Information Sciences and Technol- jobs/14792, sharing your vision for IST education;
downtown Chicago. The university has recently ogy (IST) (http://ist.psu.edu) at the Pennsylvania a CV; and contact information of 4-6 references.
cations for two or more tenure-track or tenured largest college at Purdue with over 350 faculty and Purdue University
positions in theoretical computer science. These more than 6,000 students. The College is pursu- Department of Computer Science
appointments will be at the level of Assistant or ing significant new initiatives which complement Tenure-Track/Tenured Professors in Computer
Associate Professor. The positions are part of a campus-wide plans, including an Integrative Science - Artificial Intelligence
continued expansion in a large-scale hiring effort Data Science Initiative. Opportunities for col-
across key strategic areas in the College of Science. laboration exist across mathematics, probability, The Department of Computer Science in the
statistics, and the physical and life sciences. Pur- College of Science at Purdue University invites
Qualifications: The Department is interested in due itself is one of the nation’s leading land-grant applications for two or more tenure-track or ten-
candidates whose work in theoretical computer universities, with an enrollment of over 41,000 ured positions in the broad area of artificial intel-
science focuses on the design and analysis of al- students primarily focused on STEM subjects. ligence. These appointments will be at the level
gorithms, quantum computing, randomness in of Assistant or Associate Professor. The positions
computation, as well as computational science Application Procedure: Please visit http://www. are part of a continued expansion in a large-scale
and engineering. Highly qualified applicants in cs.purdue.edu/hiring to apply. Applications hiring effort across key strategic areas in the Col-
other areas of theoretical computer science will need to include (1) a complete curriculum vi- lege of Science.
be considered. Applicants should hold a PhD in tae, (2) a statement of research and a state-
Computer Science or a related discipline, have ment of teaching, and (3) at least three names Qualifications: The Department is broadly in-
demonstrated excellence in research, and have of reference. Purdue University’s Department terested in candidates from all areas of Artificial
a strong commitment to teaching. Successful of Computer Science is committed to advanc- Intelligence. To expand and enhance our exist-
candidates will be expected to conduct research ing diversity in all areas of faculty effort, includ- ing strengths, we are particularly interested in
in their fields of expertise, teach courses in com- ing scholarship, instruction, and engagement. machine learning, natural language processing,
puter science, and participate in department and Candidates should address at least one of these human-computer interaction, vision, and reason-
university activities. areas in their cover letter, indicating their past ing/decision making. Applicants should hold a
experiences, current interests or activities, and/ PhD in Computer Science or a related discipline,
The Department and College: The Department of or future goals to promote a climate that values have demonstrated excellence in research, and
Computer Science offers a stimulating academic diversity and inclusion. have a strong commitment to teaching. Success-
environment with active research programs in A background check will be required for ful candidates will be expected to conduct re-
most areas of computer science. The department employment in this position. Review of appli- search in their fields of expertise, teach courses in
offers undergraduate programs in Computer cations and interviews will begin in Novem- computer science, and participate in department
Science and Data Science, and graduate MS and ber 2019 and will continue until positions are and university activities.
PhD programs, including a Professional MS in filled. Inquiries can be sent to TA-search@
Information Security. For more information, see cs.purdue.edu. The Department and College: The Department of
https://www.cs.purdue.edu. Purdue University is an EOE/AA employer. Computer Science offers a stimulating academic
Computer Science is part of the College of All individuals, including minorities, environment with active research programs in
Science, which comprises the computing, physi- women, individuals with disabilities, and most areas of computer science. The department
cal, and life sciences at Purdue. It is the second- veterans are encouraged to apply. offers undergraduate programs in Computer
The School of Computer and Communication Sciences (IC) at The following documents are requested in PDF format:
EPFL invites applications for tenure track faculty positions in cover letter, curriculum vitae including a publication list,
all areas of computer and communication sciences. brief statements of research and teaching interests, and
Areas of particular interest are systems, programming contact information (name, postal address, and email) of
languages and verification, unconventional computing, and 3 references. Screening will start on December 1, 2019.
intelligent systems. In rare cases, senior faculty Further questions can be addressed to:
appointments may be possible. Profs. George Candea and Rüdiger Urbanke
We seek candidates with an outstanding academic record Co-Chairs of the Faculty Recruiting Committee
and a strong commitment to teaching and mentoring recruiting.ic@epfl.ch
students. EPFL offers internationally competitive salaries,
generous research support, attractive start-up resources, For additional information on EPFL and IC, please visit:
and outstanding research infrastructure. http://www.epfl.ch or http://ic.epfl.ch
Academics in Switzerland enjoy many research funding
opportunities, as well as an exceptionally high standard of EPFL is an equal opportunity employer and family friendly
living. To apply, please follow the application procedure at university. It is committed to increasing the diversity of its
https://facultyrecruiting.epfl.ch/position/18186243 faculty. It strongly encourages women to apply.
or future goals to promote a climate that values should demonstrate an excellent potential to coeacademicaffairs@purdue.edu. Review of ap-
diversity and inclusion. build an independent research program at the plications will begin on September 16, 2019. Ap-
A background check will be required for em- forefront of their field, and to educate and men- plications received after the date will continue to
ployment in this position. Review of applications tor students. Successful candidates will conduct be reviewed until the positions are filled. A back-
and interviews will begin in November 2019 and original research, advise graduate students, ground check will be required for employment in
will continue until positions are filled. Inquiries teach undergraduate and graduate level courses, this position.
can be sent to systems-search@cs.purdue.edu. and perform service both at the School and Uni- Purdue University is an EOE/AA employer.
Purdue University is an EOE/AA employer. versity levels. All individuals, including minorities,
All individuals, including minorities, These positions are part of a continued ex- women, individuals with disabilities, and
women, individuals with disabilities, and pansion in a large-scale hiring effort across key veterans are encouraged to apply.
veterans are encouraged to apply. strategic areas in the College of Engineering.
Purdue Engineering is pursuing significant new
growth and initiatives in Computer & Informa- San José State University - San José,
Purdue University tion Systems Engineering within ECE. These are California
School of Electrical and Computer Engineering evidenced by recent strategic investments by the Assistant/Associate Professor (Tenure-Track)
Assistant or Associate Professor of Computer college, ECE, and external sponsors in centers
Engineering such as C-BRIC, PurPL, and CRISP. San José State University - San José, California
The School is an integral part of Purdue’s Col- POSITION AVAILABILITY
The School of Electrical and Computer Engi- lege of Engineering. Purdue Engineering is one Subject to Budgetary Approval
neering at Purdue University is seeking applica- of the largest and highest-ranked engineering
tions for tenured or tenure-track positions at the colleges in the nation (8th for graduate programs Specialization: Computer/Software Engineering
Assistant or Associate Professor level in any area and 9th for undergraduate per US News and Job Opening ID (JOID): 25110
of Computer Engineering. We are particularly World Report, 2019) and renowned for top-notch Rank: Assistant/Associate Professor (Tenure-
interested in candidates in computer systems faculty, students, unique research facilities, and a Track)
and computer security. All aspects of computer culture of collegiality and excellence. The College
systems will be considered such as computer goal of Pinnacle of Excellence at Scale is guiding The Computer Engineering Department at San
networks, mobile computing, operating sys- strategic growth in new directions, by investing in José State University (SJSU) invites applications
tems, dependability, and embedded systems. people, exciting initiatives, and facilities. for two tenure-track faculty positions at the rank
Similarly, all aspects of computer security will be Submit applications online at https://tinyurl. of Assistant or Associate Professor. Areas of par-
considered including data security and privacy, com/purdue-ecesystems2019, including curricu- ticular interest include machine learning and
network security, software security, and systems lum vitae, teaching and research plans, names artificial intelligence, virtual and augmented re-
security. of three references, and copies of the two most ality, robotics, data mining and big data, cloud
Successful candidates must hold a Ph.D. de- significant publications. For information/ques- computing and virtualization, networking and
gree in Electrical and Computer Engineering, tions regarding applications, contact the Office mobile systems, computer systems architecture,
Computer Science, or a related discipline. They of Academic Affairs, College of Engineering, at FPGA, and embedded systems, but other areas
in the last five years. Presently, we are one of the Teaching load is four courses per year for the first as they are received and continue until a suitable
most popular majors at the College and expect to two years and five courses per year thereafter, candidate is found.
have over 70 Computer Science majors graduat- with a one-semester leave every four years. We The University of Alabama in Huntsville is an
ing this year (2020). offer a competitive salary and benefits package, affirmative action/equal opportunity employer/mi-
plus a start-up expense fund. For information norities/females/veterans/disabled.
Qualifications: about the Computer Science Department, visit: Please refer to log number: 20/21-549
Applicants must have a Ph.D. in Computer Sci- http://www.cs.trincoll.edu/.
ence or expected by fall 2020. Applicants strong in Applicants should submit a curriculum vitae
any area of computer science will be considered. and teaching and research statements and ar- University of Central Missouri
range for three letters of reference to be sent to: Assistant Professor in Computer Science -
Institutional Statement on Teaching Diverse https://trincoll.peopleadmin.com/postings/2020. Multiple Positions
Audiences: Consideration of applications will begin on
The strongest candidates will be expected to dem- December 15, 2019, and continue until the posi- The School of Computer Science and Mathemat-
onstrate a commitment to creative teaching and tion is filled. ics at the University of Central Missouri is accept-
an active research program that speaks to and Trinity College is an Equal-Opportunity/Affir- ing applications for four tenure-track positions in
motivates undergraduates from diverse back- mative-Action employer. Women and members Computer Science at the rank of Assistant Profes-
grounds. of minority groups are encouraged to apply. sor. The appointment will begin August 2020. We
are looking for faculty excited by the prospect of
Applicant Instructions: shaping our school’s future and contributing to
Applicants should include a cover letter, a cur- The University of Alabama in Huntsville its sustained excellence.
riculum vitae, a research statement, a teaching Assistant Professor
statement and three letters of recommendation, The Position: Duties will include teaching un-
including at least one letter specifically com- The Department of Computer Science at The Uni- dergraduate and graduate courses in computer
menting on teaching. Applications will not be versity of Alabama in Huntsville (UAH) invites ap- science and/or cybersecurity and developing new
considered until letters of recommendation have plicants for a tenure-track faculty position at the courses depending upon the expertise of the ap-
been submitted. Please address any questions Assistant Professor level beginning August 2020 plicant and school needs, conducting research
you may have to Kathy Reinersmann, Computer to support the gaming and entertainment com- which leads toward peer-reviewed publications
Science Department at kreiner1@swarthmore. puting program. and/or externally funded grants, and program ac-
edu. A Ph.D. in computer science or a closely relat- creditation/assessment. Faculty are expected to
Applications received by November 15, 2019 ed area is required. The successful candidate will assist with school and university committee work
for the Tenure Track position will receive full have a strong academic background and be able and service activities and advising majors.
consideration - Apply at https://apply.interfolio. to secure and perform funded research in areas
com/67943. typical for publication in well-regarded academic Required Qualifications:
Applications received by January 15, 2020 for conference and journal venues. In addition, the ˲˲ Ph.D. in Computer Science by August 2020
the visiting assistant professor position will re- candidate should embrace the opportunity to ˲˲ Research expertise and/or industrial experi-
ceive full consideration – Apply at https://apply. provide undergraduate education. ences in Cybersecurity, Bioinformatics, Game
interfolio.com/68448. The department has a strong commitment Development or Software Engineering
Review of all applications will continue until to excellence in teaching, research, and service; ˲˲ Demonstrated ability to teach existing courses
the positions are filled. the candidate should have good communication at the undergraduate and graduate levels
Swarthmore College actively seeks and wel- skills, strong teaching potential, and research ac- ˲˲ Ability to develop a quality research program
comes applications from candidates with ex- complishments. and secure external funding
ceptional qualifications, particularly those with UAH is located in an expanding, high-tech- ˲˲ Commitment to engage in curricular devel-
demonstrable commitments to a more inclusive nology area, in close proximity to Cummings opment/assessment at the undergraduate and
society and world. Swarthmore College is an Research Park, the second largest research park graduate levels
Equal Opportunity Employer. Women and mi- in the nation and the fourth largest in the world. ˲˲ A strong commitment to excellence in teach-
norities are encouraged to apply. Nearby are the NASA Marshall Space Flight Center, ing, research, and continued professional growth
the Army’s Redstone Arsenal, numerous Fortune ˲˲ Excellent verbal and written communication
500 and high tech companies. UAH also has an skills
Trinity College, Hartford, Connecticut array of research centers, including information The Application Process: To apply online, go
Assistant Professor of Computer Science technology and cybersecurity. In short, collab- to https://jobs.ucmo.edu. Apply to positions
orative research opportunities are abundant, and #997516, #997517, #998332 or #998446. The
Applications are invited for a tenure-track posi- many well-educated and highly technically skilled following items should be attached: a letter of
tion in computer science at the rank of Assistant people are in the area. There is also access to excel- interest, a curriculum vitae, a teaching and re-
Professor to start in the fall of 2020. Candidates lent public schools and inexpensive housing. search statement, copies of transcripts, and a list
must hold a Ph.D. in computer science at the time UAH has an enrollment of approximately of at least three professional references including
of appointment. 9,900 students. The Computer Science depart- their names, addresses, telephone numbers and
We are seeking candidates with teaching and ment offers BS, MS, and PhD degrees in Com- email addresses. Official transcripts and three
research interests in applied areas associated puter Science and contributes to interdisciplin- letters of recommendation will be requested for
with data analytics, such as database and infor- ary degrees. Faculty research interests are varied candidates invited for on-campus interview.
mation systems, data mining and knowledge and include cybersecurity, mobile computing,
discovery, machine learning, and artificial intel- data science, software engineering, visualization, For more information, contact:
ligence, but other related areas will also be seri- graphics and game computing, multimedia, AI, Dr. Songlin Tian, Search Committee Chair
ously considered. image processing, pattern recognition, and dis- School of Computer Science and
Trinity College is a coeducational, indepen- tributed systems. Recent NSF figures indicate the Mathematics
dent, nonsectarian liberal arts college located in, university ranks 30th in the nation in overall fed- University of Central Missouri
and deeply engaged with, Connecticut’s capital eral research funding in computer science. Warrensburg, MO 64093
city of Hartford. Our approximately 2,200 stu- Interested parties must submit a detailed (660) 543-4930
dents come from all socioeconomic, racial, reli- resume with references to info@cs.uah.edu or tian@ucmo.edu
gious, and ethnic backgrounds across the United Chair, Search Committee, Department of Com-
States, and seventeen percent are international. puter Science, The University of Alabama in Initial screening of applications begins No-
We emphasize excellence in both teaching and Huntsville, Huntsville, AL 35899. Qualified fe- vember 15, 2019 and continues until position is
research, and our intimate campus provides an male and minority candidates are encouraged filled. AA/EEO/ADA. Women and minorities are
ideal setting for interdisciplinary collaboration. to apply. Initial review of applicants will begin encouraged to apply.
lished labs and many of these areas are currently ents (e.g., CSE has 10). Teaching load is very attrac- your teaching philosophy and accomplishments,
funded by federal agencies and industries. tive. The AI Institute has exceptional infrastructure your future teaching plans and interests, sample
and resources including 20,000 sq. ft. space. course syllabi and materials, and teaching
Qualifications: Review of applications will begin November evaluations. Applicants must arrange for three
Qualified candidates must have earned a Ph.D. 1, 2019 and continue until positions are filled. Ex- letters of reference, including at least one primarily
degree in computer science or a closely related pected start date January 1, 2020 or later. All appli- addressing the candidates teaching, to be sent
discipline by September 1, 2020. Candidates cants must apply online at http://uscjobs.sc.edu/ directly by the referees (on letterhead, signed
will be expected to do scholarly and sponsored postings/67450. Qualified candidates must in- and scanned), by email to the ECE department at
research, as well as teaching at both the under- clude: (1) letter of intent, (2) curriculum vitae, (3) search2019@ece.utoronto.ca. Applications without
graduate and graduate levels. concise description of research plans, (4) teaching any reference letters will not be considered; it is
plan, and (5) names and contact information of your responsibility to make sure your referees send
Applications: 3 references for a junior faculty rank and 5 refer- us the letters while the position remains open.
Applicants should send a cover letter, curriculum ences for a senior faculty rank (references can be You must submit your application online
vitae, statements of teaching and research inter- contacted later in the process for a senior posi- while the position is open, by following the sub-
ests, evidence of teaching performance (if any), tion). For questions or further information, please mission guidelines given at http://uoft.me/how-
and a list of three references through Interfolio at: contact Dr. Amit Sheth (amit@sc.edu). to-apply. Applications submitted in any other way
http://apply.interfolio.com/68333 for the po- The University of South Carolina does not dis- will not be considered. We recommend combin-
sition in software engineering; criminate in educational or employment opportu- ing attached documents into one or two files in
http://apply.interfolio.com/68336 for the po- nities on the basis of race, sex, gender, age, color, PDF/MS Word format. If you have any questions
sition in emerging systems or any other area of religion, national origin, disability, sexual orienta- about this position, please contact the ECE de-
computer science. tion, genetics, protected veteran status, pregnancy, partment at search2019@ece.utoronto.ca.
Review of applications will begin immediately childbirth or related medical conditions. The University of Toronto is strongly com-
and continue until suitable candidates are ap- mitted to diversity within its community and es-
pointed. pecially welcomes applications from racialized
The University of Michigan-Dearborn, as an University of Toronto persons / persons of colour, women, Indigenous /
equal opportunity/affirmative action employer. Assistant Professor, Teaching Stream Aboriginal People of North America, persons with
disabilities, LGBTQ persons, and others who may
The Edward S. Rogers Sr. Department of Electri- contribute to the further diversification of ideas.
University of South Carolina cal and Computer Engineering (ECE) at the Uni- As part of your application, you will be asked
Artificial Intelligence Institute versity of Toronto invites applications for a full- to complete a brief Diversity Survey. This survey is
Multiple Open-Rank Faculty Positions time teaching stream faculty appointment at the voluntary. Any information directly related to you is
rank of Assistant Professor, Teaching Stream, in confidential and cannot be accessed by search com-
The Artificial Intelligence (AI) Institute (http:// the general area of Computer Systems and Soft- mittees or human resources staff. Results will be
ai.sc.edu) is a new university-wide institute en- ware. The appointment will commence on July 1, aggregated for institutional planning purposes. For
gaged in core AI research, as well as high-impact 2020, or shortly thereafter. more information, please see http://uoft.me/UP.
interdisciplinary research involving AI imple- Applicants must have a Ph.D. in Electrical and All qualified candidates are encouraged to
mentations and applications. It is an outcome of Computer Engineering, or a related field, at the apply; however, Canadians and permanent resi-
the university’s Presidential Excellence Initiative, time of appointment or soon after. dents will be given priority.
which seeks to bring national prominence to our The successful candidate will have demon-
college and university through AI research and its strated excellence in teaching and pedagogical in-
economic impact. We seek multiple tenured and quiry, including in the development and delivery University of Toronto
tenure-track faculty members at all ranks in core- of undergraduate courses and laboratories, cur- Assistant Professor – Tenure Stream
AI and in interdisciplinary fields at the intersec- riculum development, and supervision of under-
tion with engineering disciplines. graduate design projects. This will be demonstrat- The Edward S. Rogers Sr. Department of Electrical
˲˲ Applicant is required to possess a Ph.D. degree ed by strong communication skills, a compelling and Computer Engineering (ECE) at the University
in computer science or a closely related field by statement of teaching submitted as part of the of Toronto invites applications for up to three full-
the beginning date of employment and have a application highlighting areas of interest, awards time tenure stream faculty appointments at the
demonstrated superior record of research accom- and accomplishments and teaching philosophy; rank of Assistant Professor. The appointments will
plishments. sample course syllabi and materials; and teaching commence on July 1, 2020, or shortly thereafter.
˲˲ The successful applicant is expected to develop evaluations, as well as strong letters of reference Within the general field of electrical and com-
internationally recognized, externally-funded re- from referees of high standing endorsing excel- puter engineering, we seek applications from
search programs that broaden the institution’s lent teaching and commitment to excellent peda- candidates with expertise in one or more of the
strengths, leverage interdisciplinary collabora- gogical practices and teaching innovation. following strategic research areas: 1. Computer
tions (http://bit.ly/AIInst), and align with vital Eligibility and willingness to register as a Pro- Systems and Software; 2. Electrical Power Sys-
cross-cutting research themes (eg. smart & con- fessional Engineer in Ontario is highly desirable. tems; 3. Systems Control, including but not lim-
nected communities, healthcare transforma- Salary will be commensurate with qualifica- ited to autonomous and robotic systems.
tions, and agile manufacturing). tions and experience. Applicants must have a Ph.D. in Electrical and
The Edward S. Rogers Sr. Department of Elec- Computer Engineering, or a related field, at the
Research areas of special interest include: trical and Computer Engineering at the Univer- time of appointment or soon after.
˲˲ Human in the loop or knowledge-enhanced AI, sity of Toronto ranks among the best in North Successful candidates will be expected to
deep learning/MMML, NLP, QA/conversational America. It attracts outstanding students, has ex- initiate and lead an outstanding, innovative, in-
AI, brain-inspired computing; cellent facilities, and is ideally located in the mid- dependent, competitive, and externally funded
˲˲ AI and Big data (incl. sensor, social, health, bio- dle of a vibrant, artistic, diverse and cosmopoli- research program of international calibre, and
logical); tan city. Additional information may be found at to teach at both the undergraduate and graduate
˲˲ AI and computer vision, robotics, CPS, human- http://www.ece.utoronto.ca. levels. Candidates must have demonstrated ex-
computer interaction, autonomous vehicles, etc. Review of applications will begin after Oc- cellence in research and teaching. Excellence in
tober 9, 2019, however, the position will remain research is evidenced primarily by publications
The faculty will have the appointment with the open until December 2, 2019. or forthcoming publications in leading journals
new AI Institute with tenure-track or tenured ap- As part of your online application (https:// or conferences in the field, presentations at sig-
pointment in CSE (http://cse.sc.edu) or another de- utoronto.taleo.net/careersection/jobdetail.ftl?job= nificant conferences, awards and accolades, and
partment in the college (http://cec.sc.edu/). CEC is 1903901&lang=en), please include a cover letter, a strong endorsements by referees of high interna-
ranked among top 100 engineering colleges in the curriculum vitae, and a teaching dossier including tional standing. Evidence of excellence in teach-
nation, and has many NSF CAREER Award recipi- a summary of your previous teaching experience, ing will be demonstrated by strong communica-
[ C ONTI N U E D FRO M P. 160] Finally, lower rate than the higher-fidelity AR/
and most important, EEG sensing, VR network audience. Of course, ASI-
passive for the live audience, active for Some of style tech found other uses, notably
me and the VR/AR audience and for the audience knows for prostitution, but we won’t go into
any live audience members who “en- any detail on that. Some guitar players
rolled” their sensory implants. Some what’s happening: tried hacking ENM (EmotoNeuroMus-
of the audience know what’s happen- manipulation cular) interfaces into their arms and
ing: manipulation of their emotions, hands, with interesting results, but
via my voice, which they are steering of their emotions, not all good. One poor guy put surplus
with their emotions. They don’t care, via my voice, which leg muscle actuators in his hands, and
though. They’re there to feel some- was quite amazing, until two of his fin-
thing, and the more they give in, give they are steering gers tore off, flying into the audience
up, give, the better. with their emotions. during a particularly enthusiastic gui-
But is it Really Music? tar shred. That was really funn …
Maybe not, or maybe more than Hey!! Pay attention! PAIN/Itch …
ever in history. All I know is that it sure I’ve caught the eye of one particu-
beats the heck out of what I was doing lar girl on the front row. Actually, she’s
for a living. I have three degrees: Music, caught my focus. I can’t do that. Any
AI, and CyberEthics, yet there I was, one-on-one connection messes up the
grinding BlockCoin in the VGame in- audience biometrics. There’s pain, and
dustry. eSports, my ass; just cramps in it’s how I will say it. I’m engineered to lots of that itch ... The bad itch. I need
my hands and fingers, wearing out my connect, to persuade. to climb back into the ASI furrow and
tendons and many VR controllers, just There aren’t many ASIngers; the do my job.
to earn a few µB¢s per minute. I have a market can only support a few at a OK, a little better now …
fairly good voice, and grew up singing, time. Certainly not as many as the But not for the audience. My con-
so becoming an AVeC was a far prefer- castrati (my bio-altered singer an- nection with them is now broken by
able career choice. cestors) in their heyday, or the robot my distraction with that front-row
The surgeries for the implants drummers that were briefly a craze girl. They’re not responding correctly.
didn’t hurt much and took only about during the last-gasp days of meth- They’re jealous, envious. Some are at-
six weeks to heal. That also gave me metal. The socioeconomics of all tracted to her. Like that fateful Court-
time to learn some AI-composed pop this is, of course, quite bizarre. Just ney Cox and Springsteen incident, the
songs that MusiCorp™ fed to me. as SnapGram photo filters caused an audience is now emoting at and with
Having all that stuff installed wasn’t epidemic of face and body dysmor- her, not me. I feel that strongly. ASIAI
so hard, but ripping it out would be a phia, so did CyberTune, RoboDyne, is unhappy. The audience is unhappy.
far different matter; it could destroy and other voice perfection technolo- Oh, no!
my voice, leaving me unable to speak gies create a rash of personal vocal Most important, MusiCorp™ is un-
normally ever again. Also, AVeCs be- dissatisfaction. People felt hopeless happy. Two RoBouncers have picked
come dependent on it, the feel of the to ever try to sing. They’d never be any up the front-row girl and are “usher-
extra hardware, but more important- good at it, not like those huge AVR- ing” (carrying) her out of the concert
ly, that direct emotional feedback Tube stars like Gr3tch@n, Cheetθh, venue. Within two seconds, I feel side-
from other humans. There’s more and k!dCRAP. ways motion. The AI concert manager
than one story of an AVeC having Voice perfection tech meant anyone is rotating the stage to reveal the next
their “rig” removed, and committing performing live had to be better than “act” early. My voice fades and the new
suicide within a couple of months, the best singers from before. Direct star’s voice replaces mine, my backing
from the pain of the lost (unnatural) emotion manipulation was a fairly re- track morphing into hers.
human connection. liable means to that end. As ASI tech OK then. My ASInging career is over.
No, thanks, I’ll keep my AVI. Sure, spread, the public quickly grew tired of Much more quickly than I rose to cyber
I probably only have another year or their AVRTube experiences, and tired singer “stardom,” I have fallen, and
two before the next ASIStar replaces of the pop stars that lived there. AR/ will never rise again.
me. but after that I can still be an (in- VR Video channels soon degenerated But … that girl was really cute. Maybe I
humanly) effective salesman, or politi- back to spectacular sports wipe-outs, can duck out the back door into the alley
cian.b Having one of the most influen- puppies, kitties, hedgehogs, and simi- and find her outside on the street (that
tial voices in history is worth a lot, and lar content. But the music and music is, if the RoBouncers haven’t whisked
it won’t much matter what I’m saying: personalities left. her off to a new acting career).
So as AR/VR Music Video collapsed,
there was a huge uptick in live+VR mu- *P-Ray is the creative/artistic moniker of Perry R. Cook,
b There were attempts to restrict proliferation of who is professor emeritus of computer science (also
sic concerts, and that rocked the mu- music) at Princeton University. Cook is advisor and IP
bio-assisted persuasion technology, especially Strategist to social music company Smule, and co-founder
its use by politicians. As expected, “forces”
sic industry (again). Revenues shifted of online arts education company Kadenze.
were too strong for any meaningful anti-ASI to per-minute billing for live concerts,
legislation to succeed. with venue attendees paying a slightly © 2019 ACM 0001-0782/19/11 $15.00
DOI:10.1145/3363232 P-Ray*
Future Tense
Cantando con la Corrente
(Singing with Current)
An augmented singer gets some
unexpected feedback from his audience.
AHH … T H E R E IT IS …
That familiar warm burn, actually
more of a sweet pain+itch, guiding me,
into the groove, into …
The Flow …
Just go with it. Sing the song. Don’t
worry about the lyrics. They don’t mat-
ter much anyway.
Emotion => Affect => Influence.
The slight scoops into certain notes.
A touch of vocal fry at the ends of key
phrases. Correct pitch, but not that
annoying CyberTune™. Just the right
amount of breathiness at every instant.
Perfect or, actually, maximally influ-
ential prosody. My voice, but not com-
pletely in my control. I sing the song,
sort of. The result: my deep connection
with listeners, and theirs with me…
ASIBOVa takes care of all that
Beginning a show, from the first
song, the warm itch is strong, as ASI
helps me do the right things. Bio-
actuators ad/abduct my cricothyroid,
raise and lower my larynx, flex/pulse right times, to create an “optimal” grow ever smaller. The audience yield-
my diaphragm, agonist and antago- performance. What I can’t do physi- ed long before I did; the AI and DSP
nist, tensing and relaxing all the im- cally, ASI takes care of via real-time took care of that. We all find the flow,
portant parts of my vocal mechanics DSP audio effects. I wear a headset in the song, signals, and sensations.
into just the right places, at just the mic anyway, and the audience is far Neural nets of silicon and tissue, syn-
enough away so they hear and feel chronizing. Layers of machine intelli-
only perfect, emotional ... gence grind on bio-emotion signals
a ASIBOV = Audience+Singer(Speaker) Influ-
enced Bio-feedback Optimized Voice (ASI for
Connection … gathered from the audience: their
short). Invented in 2023 by J.R. Coupling at As I let it happen, I feel it, or rather, smiles, open/closed eyes, eye-blinks
I don’t feel the itch any more. I am do- and rates, breathing rates, body poses,
IMAGE BY TA ND EM BRA NDING
:
Early Submission: 12 / 6 / 2019
Regular Submission: 2 / 5 / 2020
Contact:
confs@servicessociety.org / icws.org