ai matters, volume 4, issue 34(3) 2018 ai...
TRANSCRIPT
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
AI MattersAnnotated Table of Contents
Welcome to AI Matters 4(3)Amy McGovern, Editor
Full article: http://doi.acm.org/10.1145/3284751.3284752
Welcome and summary
2018 ACM SIGAI Student Essay Con-test on Artificial Intelligence Tech-nologiesNicholas Mattei
Full article: http://doi.acm.org/10.1145/3284751.3284753
SIGAI Student essay contest
ACM SIGAI Activity ReportSven Koenig, Sanmay Das, RosemaryParadis, John Dickerson, Yolanda Gil,Katherine Guo, Benjamin Kuipers, HangMa, Nicholas Mattei, Amy McGovern,Larry Medsker, Todd Neller, PlamenPetrov, Michael Rovatsos & David G.Stork
Full article: http://doi.acm.org/10.1145/3284751.3284754
Annual report for ACM SIGAI
EventsMichael Rovatsos
Full article: http://doi.acm.org/10.1145/3284751.3284755
Upcoming AI events
AI Education Matters: Teaching withDeep Learning Frameworks in Intro-ductory Machine Learning CoursesMichael Guerzhoy
Full article: http://doi.acm.org/10.1145/3284751.3284756
Teaching deep learning
AI Education Matters: TeachingSearch AlgorithmsNathan R. Sturtevant
Full article: http://doi.acm.org/10.1145/3284751.3284757
Teaching search
AI Profiles: An Interview with KristianKerstingMarion Neumann
Full article: http://doi.acm.org/10.1145/3284751.3284758
Interview with Kristian Kersting
AI Policy MattersLarry Medsker
Full article: http://doi.acm.org/10.1145/3284751.3284759
Updates on AI and automation
The Partnership on AIJeffrey Heer
Full article: http://doi.acm.org/10.1145/3284751.3284760
Partnership on AI to benefit people and society
Mechanism Design for Social GoodRediet Abebe & Kira Goldner
Full article: http://doi.acm.org/10.1145/3284751.3284761
Designing for social good
Links
SIGAI website: http://sigai.acm.org/Newsletter: http://sigai.acm.org/aimatters/Blog: http://sigai.acm.org/ai-matters/Twitter: http://twitter.com/acm sigai/Edition DOI: 10.1145/3284751
Join SIGAI
1
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Students $11, others $25For details, see http://sigai.acm.org/Benefits: regular, student
Also consider joining ACM.
Our mailing list is open to all.
Notice to Contributing Authorsto SIG Newsletters
By submitting your article for distribution in thisSpecial Interest Group publication, you herebygrant to ACM the following non-exclusive, per-petual, worldwide rights:
• to publish in print on condition of acceptanceby the editor
• to digitize and post your article in the elec-tronic version of this publication
• to include the article in the ACM Digital Li-brary and in any Digital Library related ser-vices
• to allow users to make a personal copy ofthe article for noncommercial, educationalor research purposes
However, as a contributing author, you retaincopyright to your article and ACM will referrequests for republication directly to you.
Submit to AI Matters!
We’re accepting articles and announce-ments now for the next issue. Details onthe submission process are available athttp://sigai.acm.org/aimatters.
AI Matters Editorial Board
Amy McGovern, Editor-in-Chief, U. OklahomaSanmay Das, Washington Univ. in Saint LouisAlexei Efros, Univ. of CA BerkeleySusan L. Epstein, The City Univ. of NYYolanda Gil, ISI/Univ. of Southern CaliforniaDoug Lange, U.S. NavyKiri Wagstaff, JPL/CaltechXiaojin (Jerry) Zhu, Univ. of WI Madison
Contact us: [email protected]
Contents Legend
Book Announcement
Ph.D. Dissertation Briefing
AI Education
Event Report
Hot Topics
Humor
AI Impact
AI News
Opinion
Paper Precis
Spotlight
Video or Image
Details at http://sigai.acm.org/aimatters
2
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Welcome to AI Matters 4(3)Amy McGovern, Editor (University of Oklahoma; [email protected])DOI: 10.1145/3284751.3284752
Issue overview
Welcome to the third issue of the fourth vol-ume of the AI Matters Newsletter. In this is-sue, we have some great news: we are bring-ing back the AI student essay contest! Pleasesee the first article for details on this fun con-test which is due January 10, 2019. In addi-tion to the prizes discussed in the article, thewinners will have their articles features in AIMatters in the issue following their selection.
Our second big news item is an in-depth sum-mary report of what ACM SIGAI has beenworking on for the last year. This report isin-depth and shows the wide extent that ACMSIGAI has been reaching out, from starting anew conference, helping with a variety of ex-isting conferences, creating new awards, sup-porting members, adding new content to thisnewsletter, and increasing our social mediapresence.
In our regular columns, don’t forget to checkout upcoming AI related events, summarizedwell in Michael Rovatsos’ events column. Wegot extra lucky this issue with two educationalsubmissions! Michael Guerzhoy contributesan article on using deep learning and tensor-flow in machine learning classes and NathanSturtevant discusses how to teach search al-gorithms to students in AI classes.
Our interview series is back! Marion Neumannhas joined our contributing authors and willbe working on interviews as well as other se-ries. In this issue, Marion interviews KristianKersting, Professor in Computer Science andDeputy Director of the Centre for CognitiveScience at the Technical University of Darm-stadt, Germany.
Larry Medsker’s policy column looks into therecent report from the WEF on the future ofjobs with increasing automation as well as ve-hicle automation. Finally, we close with twopaper contributions. The first is a discussionof the Partnership on AI to Benefit Peopleand Society. The second is a thoughtful pa-
Copyright c© 2018 by the author(s).
per on the Mechanism Design for Social Goodagenda.
Submit to AI Matters!Thanks for reading! Don’t forget to sendyour ideas and future submissions to AIMatters! We’re accepting articles and an-nouncements now for the next issue. De-tails on the submission process are avail-able at http://sigai.acm.org/aimatters.
Amy McGovern is chiefEditor of AI Matters. Sheis a Professor of com-puter science at the Uni-versity of Oklahoma andan adjunct professor ofmeteorology. She directsthe Interaction, Discovery,Exploration and Adapta-tion (IDEA) lab. Her re-search focuses on ma-
chine learning and data mining with applica-tions to high-impact weather.
3
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
2018 ACM SIGAI Student Essay Contest on Artificial IntelligenceTechnologiesNicholas Mattei (IBM Research; [email protected])DOI: 10.1145/3284751.3284753
Abstract
Win one of several prizes including$500USD or a chat with a leading AI Re-searcher. For complete details, includ-ing submission requirements, formats,and judges please see: www.tinyurl.com/SIGAIEssay2018
2018 Topic
The ACM Special Interest Group on Artifi-cial Intelligence (ACM SIGAI) supports the de-velopment and responsible application of Ar-tificial Intelligence (AI) technologies. Fromintelligent assistants to self-driving cars, anincreasing number of AI technologies now(or soon will) affect our lives. Examples in-clude Google Duplex (Link) talking to humans,Drive.ai (Link) offering rides in US cities,chatbots advertising movies by impersonat-ing people (Link), and AI systems making de-cisions about parole (Link) and foster care(Link). We interact with AI systems, whetherwe know it or not, every day.
Such interactions raise important questions.ACM SIGAI is in a unique position to shapethe conversation around these and related is-sues and is thus interested in obtaining inputfrom students worldwide to help shape the de-bate. We therefore invite all students to enteran essay in the 2018 ACM SIGAI Student Es-say Contest, to be published in the ACM SIGAInewsletter AI Matters, addressing one or bothof the following topic areas (or any other ques-tion in this space that you feel is important)while providing supporting evidence:
• What requirements, if any, should be im-posed on AI systems and technology wheninteracting with humans who may or may notknow that they are interacting with a ma-chine? For example, should they be re-quired to disclose their identities? If so,
Copyright c© 2018 by the author(s).
how? See, for example,“Turing?s Red Flag”in CACM (Link).
• What requirements, if any, should be im-posed on AI systems and technology whenmaking decisions that directly affect hu-mans? For example, should they be re-quired to make transparent decisions? If so,how? For example, the IEEE’s discussion ofEthically Aligned Design (Link).
Each of the above topic areas raises furtherquestions, including
• Who is responsible for the training andmaintenance of AI systems? See, for ex-ample, Google?s (Link), Microsoft?s (Link),and IBM?s (Link) AI Principles.
• How do we educate ourselves and othersabout these issues and possible solutions?See, for example, new ways of teaching AIethics (Link).
• How do we handle the fact that different cul-tures see these problems differently? See,for example, Joi Ito?s discussion in Wired(Link).
• Which steps can governments, industries,or organizations (including ACM SIGAI) taketo address these issues? See, for example,the goals and outlines of the Partnership onAI (Link).
All sources must be cited. However, we arenot interested in summaries of the opinionsof others. Rather, we are interested in theinformed opinions of the authors. Writing anessay on this topic requires some backgroundknowledge. Possible starting points for acquir-ing such background knowledge are: the re-vised ACM Code of Ethics (Link), especiallySection 3.7, and a discussion of why the re-vision was necessary (Link), IEEE?s EthicallyAligned Design (Link), and the One HundredYear Study on AI and Life in 2030 (Link).
4
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Format and Eligibility
The ACM SIGAI Student Essay Contest isopen to all ACM SIGAI student members atthe time of submission. (If you are a studentbut not an ACM SIGAI member, you can joinACM SIGAI before submission for just US $11at https://goo.gl/6kifV9 by selecting Option 1,even if you are not an ACM member.) Es-says can be authored by one or more ACMSIGAI student members but each ACM SIGAIstudent member can (co-)author one essay.
All authors must be SIGAI members at thetime of submission. All submissions not meet-ing this requirement will not be reviewed.
Essays should be submitted as pdf documentsof any style with at most 5,000 words via emailto AI Matters Easychair.
The deadline for submissions is January10th, 2019.
Prizes
All winning essays will be published in theACM SIGAI newsletter AI Matters. ACM SIGAIprovides five monetary awards of USD 500each as well as 45-minute Skype sessionswith one of the following AI researchers:
• Joanna Bryson, Reader (Assoc. Prof) in AI,University of Bath
• Murray Campbell, Senior Manager, IBM Re-search AI
• Eric Horvitz, Managing Director, MicrosoftResearch
• Peter Norvig, Director of Research, Google• Iyad Rahwan, Associate Professor, MIT Me-
dia Lab and Head of Scalable Corp.• Francesca Rossi, AI and Ethics Global
Lead, IBM Research AI• Toby Walsh, Scientia Professor of Artificial
Intelligence, UNSW Sydney, Data61 and TUBerlin
Judges and Judging Criteria
Winning entries from last year’s essay con-test can be found in recent issues of the ACMSIGAI newsletter AI Matters, specifically Vol-ume 3, Issue 3 and Volume 3, Issue 4.
Entries will be judged by the following panelof leading AI researchers and ACM SIGAI of-ficers. Winning essays will be selected basedon depth of insight, creativity, technical merit,and novelty of argument. All decisions by thejudges are final.
• Rediet Abebe, Cornell University• Emanuelle Burton, University of Illinois at
Chicago• Sanmay Das, Washington University in St.
Louis• John P. Dickerson, University of Maryland• Virginia Dignum, Delft University of Technol-
ogy• Tina Eliassi-Rad, Northeastern University• Judy Goldsmith, University of Kentucky• Amy Greenwald, Brown University• H. V. Jagadish, University of Michigan• Sven Koenig, University of Southern Califor-
nia• Benjamin Kuipers, University of Michigan• Nicholas Mattei, IBM Research• Alexandra Olteanu, Microsoft Research• Rosemary Paradis, Leidos• Kush Varshney, IBM Research• Roman Yampolskiy, University of Louisville• Yair Zick, National University of Singapore
Main organization by Nicholas Mattei (IBM Re-search), AI and Society Officer with involve-ment from Sven Koenig (University of South-ern California), ACM SIGAI Chair; SanmayDas (Washington University in St. Louis),ACM SIGAI Vice Chair; Rosemary Paradis(Leidos), ACM SIGAI Secretary/Treasurer;Benjamin Kuipers (University of Michigan),ACM SIGAI Ethics Officer; and Amy McGov-ern (University of Oklahoma), ACM SIGAI AIMatters Editor-in Chief.
In case of questions, please first check theACM SIGAI blog for announcements and clar-ifications Link. You can also contact the ACMSIGAI Student Essay Contest Organizers [email protected].
5
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Nicholas Mattei is aResearch Staff Memberin the IBM Research AIgroup at the IBM TJ Wat-son Research Laboratory.His research focuses onthe theory and practiceof AI, developing systemsand algorithms to supportdecision making.
6
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
ACM SIGAI Activity ReportSven Koenig (elected; ACM SIGAI Chair)Sanmay Das (elected; ACM SIGAI Vice-Chair)Rosemary Paradis (elected; ACM SIGAI Secretary/Treasurer)John Dickerson (appointed; ACM SIGAI Labor Market Officer)Yolanda Gil (appointed; ACM SIGAI Past Chair)Katherine Guo (appointed; ACM SIGAI Membership and Outreach Officer)Benjamin Kuipers (appointed; ACM SIGAI Ethics Officer)Hang Ma (appointed; ACM SIGAI Information Officer)Nicholas Mattei (appointed; ACM SIGAI AI and Society Officer)Amy McGovern (appointed; ACM SIGAI Newsletter Editor-in-Chief)Larry Medsker (appointed; ACM SIGAI Public Policy Officer)Todd Neller (appointed; ACM SIGAI Education Activities Officer)Plamen Petrov (appointed; ACM SIGAI Industry Liaison Officer)Michael Rovatsos (appointed; ACM SIGAI Conference Coordination Officer)David G. Stork (appointed; ACM SIGAI Award Officer)DOI: 10.1145/3284751.3284754
Abstract
We are happy to present the annual activityreport of ACM SIGAI, covering the period fromJuly 2017 to June 2018.
The scope of ACM SIGAI consists of thestudy of intelligence and its realization incomputer systems (see also our website atsigai.acm.org). This includes areas suchas
autonomous agents, cognitive modeling,computer vision, constraint programming, hu-man language technologies, intelligent userinterfaces, knowledge discovery, knowledgerepresentation and reasoning, machine learn-ing, planning and search, problem solving,and robotics.
Our members come from academia, industry,and government agencies worldwide. We arethrilled to be able to report that our member-ship numbers increased by about 10 percentover the past year!
The terms of three of our officers (one of thetwo education activities officers, one of the twonewsletter co-editors in chief, and the informa-tion officer) came to an end. We thank them
Copyright c© 2018 by the author(s).
for their valuable service and have now startedto provide certificates of appreciation for out-going officers. We appointed a new informa-tion officer and are currently looking for a newnewsletter co-editor in chief and additional col-umn editors. We also created and filled twonew officer positions to be able to serve ourmembers even better, namely an AI and soci-ety officer and a labor market officer.
Conferences
We helped to found the AAAI/ACM AI, Ethics,and Society (AIES) conference, a high-profile,multi-disciplinary meeting that addresses theimpact of AI on society (including aspectssuch as value alignment, data handling andbias, regulations, and workforce impact) in ascientific context. We have a 50 percent fi-nancial stake in AIES, which we expect to be-come the prime international conference in thefield. The inaugural AIES, held on February13, 2018 in New Orleans directly before theAAAI conference, was a great success, at-tracting 162 submissions and selling out (seea recent summary in our newsletter AI Mat-ters). The AI and society officer organizedthe AIES doctoral consortium, which receivedover 60 submissions but could only accept 20students. AIES 2019 is now in the planning
7
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
phase.
We sponsored the following conferences inaddition to AIES 2018:
• The IEEE/WIC/ACM International Confer-ence on Web Intelligence (WI 2017),Leipzig, Germany, August 23-26, 2017
• The 32nd International Conference on Au-tomated Software Engineering, Urbana-Champaign, USA, October 30-November 3,2017
• The 9th International Conference on Knowl-edge Capture (K-CAP 2017), Austin, TX,USA, December 4-6, 2017
• The CRA Summit on Technology and Jobs,Washington D.C., USA, December 12, 2017
• The 23rd International Conference on In-telligent User Interfaces (IUI 2018), Tokyo,Japan, March 7-11, 2018
• The 13th Annual ACM/IEEE InternationalConference on Human Robot Interaction(HRI 2018), Chicago, USA, March 5-8, 2018
We approved the following in-cooperation andsponsorship requests from events covering awide thematic and geographical range acrossthe international AI community:
• The 20th International Conference on En-terprise Information Systems (ICEIS 2018),Funchal, Portugal, March 21-24, 2018
• The 17th International Conference on Au-tonomous Agents and Multiagent Systems(AAMAS 2018), Stockholm, Sweden, July10-15, 2018
• The 33rd IEEE/ACM International Confer-ence on Automated Software Engineering(ASE 2018), Montpellier, France, Septem-ber 3-7, 2018
• The 18th International Conference on Intel-ligent Virtual Agents (IVA 2018), Sydney,November 5-8, 2018
• The 2nd ACM Computer Science in CarsSymposium (CSCS 2018), Munich, Ger-many, September 13-14, 2018
• The 10th International Joint Conferenceon Knowledge Discovery, Knowledge Engi-neering and Knowledge Management (IC3K2018), Seville, Spain, September 18-20,2018
• International Conference on the Founda-tions of Digital Games (FDG 2018), Malm,Sweden, August 7-10, 2018
• The 10th International Joint Conference onComputational Intelligence, Seville, Spain,September 18-20, 2018
• The 15th International Conference on Infor-matics in Control, Automation, and Robotics(ICINCO 2018), Porto, Portugal, July 29-31,2018
• The First IEEE International Conferenceon Artificial Intelligence and Virtual Reality(IEEE-AIVR 2018), Taichung, Taiwan, De-cember 10-12, 2018
ACM SIGAI membership benefits include re-duced registration fees for many of the in-cooperation and sponsored conferences andaccess to many proceedings in the ACM Digi-tal Library.
We also have an agreement with the Asso-ciation for the Advancement of AI (AAAI) tojointly organize the annual joint job fair atthe AAAI conference, where attendees canfind out about job and internship opportuni-ties from representatives from industry, univer-sities, and other organizations. The job fairin 2018 was run by two of our officers andattracted over 30 organizations offering jobsand hundreds of job seekers (see the recentsummary in AI Matters). Similarly, we havean agreement with AAAI to jointly sponsor theannual joint doctoral consortium at the AAAIconference, which provides an opportunity forPh.D. students to discuss their research in-terests and career objectives with the otherparticipants and a group of established AI re-searchers that act as their mentors.
Awards
We started a new award, the ACM SIGAI In-dustry Award for Excellence in AI, which willbe given annually to an individual or team inindustry who created a fielded AI applicationin recent years that demonstrates the power ofAI techniques via a combination of the follow-ing features: novelty of application area, nov-elty, and technical excellence of the approach,importance of AI techniques for the approach,and actual and predicted societal impact of theapplication. The award will be accompaniedby a prize of US$5,000 and be presented at
8
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
the International Joint Conference on AI (IJ-CAI), starting in 2019 - thanks to a collabora-tion with the IJCAI Board of Trustees.
The ACM SIGAI Autonomous Agents Re-search Award was presented at the Interna-tional Conference on Autonomous Agents andMultiagent Systems (AAMAS) 2018 - thanksto a collaboration with the International Foun-dation for Autonomous Agents and MultiagentSystems (IFAAMAS) - to Craig Boutilier, Prin-cipal Research Scientist at Google, for hisseminal contributions to research on decision-making under uncertainty, game theory, andcomputational social choice.
Member Support
We support our members in different ways.For example, we nominate them for awards orsupport their nominations. We also nominatepublications of recent, significant, and excit-ing AI research results that are of general in-terest to the computer science research com-munity to the Research Highlights track of theCommunications of the ACM (CACM) to makeimportant AI research results visible to manycomputer scientists. We concentrate our fi-nancial support on our student members andprovided funding to conferences so that theycan award scholarships to their student atten-dees. We also supported the review processfor the applications of students to attend theHeidelberg laureate forum to meet the recipi-ents of the most prestigious awards in math-ematics and computer science, including theACM A.M. Turing Award and the ACM Prize inComputing.
To understand the interests of our membersbetter, we conducted a membership surveyand learned, among other things, that the ma-jority of our members have been SIGAI mem-bers for two or more years. They are inter-ested in AI resources (such as digital con-tent access and webinars) as well as network-ing, and consider it important that we cover(among other things) hot AI topics, AI ethics,and AI applications with an industry focus.We also reached out to other groups to learnabout their outreach activities, including to re-gional groups (such as the Northeast OhioACM chapter) and larger ACM Special Inter-est Groups (such as SIGCHI).
Member Communication
We communicate with our members via emailannouncements, the SIGAI newsletter AI Mat-ters, the AI Matters blog, and webinars.
AI Matters
We continued to expand the scope of ournewsletter AI Matters this year, introducedthe EasyChair system for the submissionof manuscripts and added additional columneditors to the editorial team. We publishfour issues of AI Matters per year that areopenly available on the ACM SIGAI website atsigai.acm.org/aimatters/ and featurearticles of general interest to our members.Recurring columns have included:
• AI Interviews (with interesting people fromacademia, industry, and government),
• AI Amusements (including AI humor, puzzles,and games),
• AI Education (led by the education activities offi-cer),
• AI Policy Issues (led by the public policy officer),• AI Buzzwords (which explains new AI concepts
or terms),• AI Events (which includes conference an-
nouncements and reports, led by the conferencecoordination officer),
• AI Dissertation Abstracts and• News from AI Groups and Organizations.
The Symposium on Educational Advances inAI (EAAI) will feature an undergraduate re-search track for the Birds of a Feather faculty-mentored undergraduate research challengebased on one of the AI Education columns. AIMatters also published the 8 winning entriesfrom the ACM SIGAI Student Essay Conteston the Responsible Use of AI Technologies.One of these essays had more than 2,000 ac-cesses (including accesses to the pdf file ofthe complete issue that contained the article).
AI Matters Blog
The AI Matters Blog is openly avail-able on the ACM SIGAI website atsigai.acm.org/aimatters/blog/ andserves as a forum for important announce-ments and news. For example, we post new
9
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
information every two weeks in the blog tosurvey and report on current AI policy issuesand raise awareness about the activities ofother organizations that share interests withACM SIGAI. We are also open to posts by re-gional groups and have invited the NortheastOhio ACM chapter to contribute.
Webinars
We extended our webinars on AI topics aspart of our commitment to lifelong learning. Inparticular, we offered monthly webinars fromNovember 2017 to June 2018 that focusedon the application of AI technology to real-world problems and were presented by speak-ers typically involved with both academic re-search and industrial implementations:
• November 10, 2017Dan MoldovanFounder of Lymba Corporation, Professorin the Computer Science Department atthe University of Texas at Dallas and Co-Director of its Human Language TechnologyResearch InstituteTopic: “On the Evolution of NLP, QA, andIE, and Current Research and CommercialTrends”
• December 15, 2017Peter ElkinProfessor and Chair of the University at Buf-falo Department of Biomedical InformaticsTopic: “HTP-NLP: A New NLP System forHigh Throughput Phenotyping”
• January 12, 2018Lionel JouffeCo-founder and CEO of France-basedBayesia S.A.S.Topic: “Data Mining, Knowledge Modeling,and Causal Analysis with Bayesian Net-works”
• February 23, 2018Kristian HammondProfessor of Computer Science at North-western University and co-founder of Narra-tive ScienceTopic: “Communicating with the New Ma-chine: Human Insight at Machine Scale”
• March 15, 2018Tomek StrzalkowskiDirector of the Institute for Informatics, Log-ics, and Security Studies and Professor at
SUNY AlbanyTopic: “Advances in Socio-Behavioral Com-puting”
• May 7, 2018Jussi KarlgrenKTH Royal Institute of Technology andHelsinki University and founding partner ofGavagi, a text analysis companyTopic: “Explicitly Encoded High-Dimensional Semantic Spaces”
• June 4, 2018Maja MataricProfessor and Chan Soon-Shiong Chair inthe Computer Science Department, Neuro-science Program, and the Department ofPediatrics at the University of Southern Cal-iforniaTopic: “Socially Assistive Robotics”
The webinars were streamed live butcan still be watched on demand atlearning.acm.org/webinar/. Theyturned out to be very popular, typicallyreaching a thousand or more viewers.
Public Policy and AI Ethics
Within ACM, we work with the ACM US PublicPolicy Council (USACM) through the member-ship of our public policy officer in USACM andthe participation of our members in US pub-lic policy issues related to computing and in-formation technology. For example, our pub-lic policy officer works with the USACM lead-ership and the Electronic Privacy InformationCenter to petition the Office of Science andTechnology Policy of the White House to con-struct and publicize a formal process by whichthe public might have input into the work ofthe recently-named Select Committee on Arti-ficial Intelligence. He also studies how organi-zations collect and analyze data and whetherthese practices are consistent with recom-mendations by the USACM working groups onalgorithmic accountability, transparency, andbias. Finally, he works on recommendationsfor possible changes to the ACM policy re-garding data privacy of ACM SIGAI and ACMmembers who use EasyChair to submit arti-cles for publication, including to AI Matters.
Outside of ACM, we helped to found theAAAI/ACM AI, Ethics, and Society (AIES) con-ference (as detailed above). We also partici-
10
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
pate in the executive committee of the IEEEGlobal Initiative on Ethics of Autonomous andIntelligent Systems to ensure that every tech-nologist is educated, trained, and empoweredto prioritize ethical considerations in the de-sign and development of autonomous and in-telligent systems.
Planning for the Future
In our last progress report, we promised to in-crease our understanding of our current mem-bership. To further this objective, we per-formed a membership survey this year. Wepromised to intensify our activities to reachindustry professionals. To further this objec-tive, we substantially extended our webinarsthat focus on the application of AI technol-ogy to real-world problems and started theACM SIGAI Industry Award for Excellence inAI. We promised to reach out to more AIgroups worldwide that could benefit from ACMsupport, such as providing financial support,making the proceedings widely accessible inthe ACM Digital Library, and providing speak-ers via the ACM Distinguished Speakers pro-gram. To further this objective, our mem-bership officer contacted a number of suchgroups. We promised to expand the number ofco-sponsored and in-cooperation conferencesand to continue our efforts to further the dis-cussion on the responsible use of AI technolo-gies. To further these objectives, we helpedto found the AAAI/ACM AI, Ethics, and Soci-ety conference. In the next year, we intend tocontinue the reorganization of AI Matters to beable to provide more content while spreadingthe production effort among more editors. Wealso intend to intensify our activities to supportstudents and regional chapters and becomeeven more international in our news coverage.Furthermore, the election of the next SIGAIleadership team is coming up and we are look-ing for strong candidates who are interested inshaping the future of ACM SIGAI!
11
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
EventsMichael Rovatsos (University of Edinburgh; [email protected])DOI: 10.1145/3284751.3284755
This section features information about up-coming events relevant to the readers of AIMatters, including those supported by SIGAI.We would love to hear from you if you are areorganizing an event and would be interestedin cooperating with SIGAI, or if you haveannouncements relevant to SIGAI. For moreinformation about conference support visitsigai.acm.org/activities/requesting sponsor-ship.html.
International Summit of ArtificialIntelligenceSophia Antipolis, France, November 7-9, 2018sophia-summit.com/sophia2018/enEurope’s first science and technology parkand digital cradle, Sophia Antipolis, will host,from November 7th to 9th, 2018, the top Arti-ficial Intelligence specialists for a summit ded-icated to researchers, businesses, economicpartners, students and the general public.This summit will be the culmination of otherevents that, throughout 2018, have promotedand challenged Artificial Intelligence and thenumerous applications of AI. The summit isjointly organized by the Sophia Antipolis Ur-ban Community and University Cote d’Azur.
18th ACM International Conference onIntelligent Virtual Agents (IVA 2018)Sydney, Australia, November 5-8, 2018iva2018.westernsydney.edu.auIVA 2018 is the 18th meeting of an interdisci-plinary annual conference and the main lead-ing scientific forum for presenting researchon modeling, developing and evaluating intel-ligent virtual agents (IVAs) with a focus oncommunicative abilities and social behavior.IVAs are interactive digital characters that ex-hibit human-like qualities and can communi-cate with humans and each other using nat-ural human modalities like facial expressions,speech and gesture. They are capable of real-time perception, cognition, emotion and actionthat allow them to participate in dynamic so-
Copyright c© 2018 by the author(s).
cial environments. In addition to presentationson theoretical issues, the conference encour-ages the showcasing of working applications.Students, academics and industry profession-als with an interest in learning about and pre-senting the most cutting-edge research con-ducted today in the multi-disciplinary field ofintelligent virtual agents. Advances in IVA re-search are enabling increasingly autonomousagents that are now being utilized across agrowing number of fields, including counsel-ing, entertainment, medicine, the military, psy-chology and teaching. Researchers from thefields of human-human and human-robot in-teraction are also encouraged to share workwith a relevance to IVAs.
1st IEEE International Conference onArtificial Intelligence and VirtualReality (AIVR 2018)Taichung, Taiwan, December 10-12, 2018aivr.asia.edu.tw/2018/The First IEEE International Conference onArtificial Intelligence and Virtual Reality (AIVR2018) is a unique event, addressing re-searchers and industries from all areas of AIas well as Virtual, Augmented, and Mixed Re-ality. It provides an international forum for theexchange between those fields, to present ad-vances in the state of the art, identify emerg-ing research topics, and together define thefuture of these exciting research domains. Weinvite researchers from Virtual, as well as Aug-mented Reality (AR) and Mixed Reality (MR)to participate and submit their work to the pro-gram. Likewise, any work on AI that has a re-lation to any of these fields or potential for theusage in any of them is welcome.
24th International Conference onIntelligent User Interfaces (IUI 2019)Los Angeles, CA, USA, March 17-20, 2019iui.acm.org/2019ACM IUI 2019 is the 24th annual meetingof the intelligent interfaces community andserves as a premier international forum for
12
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
reporting outstanding research and develop-ment on intelligent user interfaces. ACM IUI iswhere the Human-Computer Interaction com-munity meets the Artificial Intelligence com-munity. We are also very interested in contri-butions from related fields, such as psychol-ogy, behavioral science, cognitive science,computer graphics, design, the arts, etc.
32nd International Conference onIndustrial, Engineering & OtherApplications (IEA/AIE 2019)Graz, Austria, July 9-11, 2019ieaaie2019.ist.tugraz.atThe 32nd International Conference on Indus-trial, Engineering & Other Applications of Ap-plied Intelligent Systems will continue the tra-dition of emphasizing on applications of ap-plied intelligent systems to solve real-life prob-lems in all areas. These areas include: en-gineering, science, industry, automation &robotics, business & finance, medicine andbiomedicine, bioinformatics, cyberspace, andhuman-machine interactions. IEA/AIE 2019will have a special focus on automated drivingand autonomous systems such that contribu-tions dealing with such systems or their verifi-cation and validation are especially welcome.
Michael Rovatsos is theConference CoordinationOfficer for ACM SIGAI,and a faculty memberof the School of Infor-matics at the Universityof Edinburgh, UK. Hisresearch in in multia-gent systems and human-friendly AI. Contact him [email protected].
13
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
AI Education Matters: Teaching with Deep Learning Frameworksin Introductory Machine Learning CoursesMichael Guerzhoy (Princeton University, University of Toronto, and the Li Ka Shing Knowl-edge Institute, St. Michael’s Hospital; [email protected])DOI: 10.1145/3284751.3284756
Introduction
In this article, we demonstrate an assignment1in which students use TensorFlow to build aface recognition system. Students build shal-low and deep neural networks for face recog-nition in TensorFlow and use transfer learningto obtain near-perfect performance on a sim-ple face recognition task. Visualizing neuralnetworks in order to explain how they work iscentral to the assignment.
The emergence of frameworks such as Ten-sorFlow (Abadi et al., 2016) and PyTorch2
and the explosive growth of ready-made opensource machine learning systems on the Webchanged the practice of machine learning.The pedagogy of machine learning faceschallenges that are analogous to the chal-lenges introductory computer science peda-gogy started facing with the advent of high-level programming languages and GUI li-braries (Astrachan, Bruce, Koffman, Kolling,& Reges, 2005): we would like for studentsto use the tools practitioners use to build real-world systems, but at the same time we wantstudents to have a thorough understanding ofhow systems can be built “from scratch.”
We resolve this dilemma by advocating for anassignment style that approximates buildinga new machine learning system prototype asclosely as possible. We avoid giving out in-complete starter code and asking students to“fill in the blanks.” Instead, we give studentsworking code that is relevant to the task theyare working on and that is similar to code thatthey might find on the Web. Students use asmuch or as little of the code as they want tocomplete the assignment.
Copyright c© 2018 by the author(s).1http://modelai.gettysburg.edu/
2018/nnfaces/2Excellent introductory tutorials are available at
http://www.tensorflow.org and http://www.pytorch.org
Figure 1: Demonstrating overfitting by visualizingthe weights of a one-hidden-layer neural network
We ask students to visualize the modelsthat they are working with, building on re-cent research work on visualizing and un-derstanding neural networks (Zeiler & Fer-gus, 2014) (Springenberg, Dosovitskiy, Brox,& Riedmiller, 2014). Visualizing ML modelsis a nontrivial technical task which allows stu-dents to practice developing machine learningcode, it is helpful when debugging ML models,and, finally, it allows students to gain a betterintuition for why the models work and whenthey can fail. Pedagogically, this task is bothnontrivial and feasible for students to accom-plish.
Model Assignment: Neural Networksfor Face Recognition with TensorFlow
In the assignment, students train both shallowand deep networks to classify faces of six fa-mous actors. Students adapt handout Tensor-Flow code in order to build and train the net-works.
Students train a one-hidden-layer neural net-work for face classification. Students then im-prove the performance of their system by train-
14
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
ing a convolutional neural network to performface classification; as a first step, students usethe activations of a pre-trained AlexNet net-work (Krizhevsky, Sutskever, & Hinton, 2012)as the features for their face classifier.
Students visualize the shallow and the deepnetworks that they train.
Introducing TensorFlow
TensorFlow is introduced by giving studentsworking code that trains a one-hidden-layerneural network on the MNIST dataset. Wealso provide a beginner-friendly implemen-tation of AlexNet in TensorFlow, along withweights pretrained on ImageNet3. The hand-out is designed to be very easily readableand only uses the most elementary Tensor-Flow features. This makes it possible for stu-dents to modify the code to suit their needswithout having to immediately understand thehighly-engineered neural network implemen-tations that are included with TensorFlow. Stu-dents can (and do) reuse the generic handoutcode later on in their career.
Visualizing Feedforward Networks
Visualizing the weights of shallow neural net-works allows students to understand one-hidden-layer neural networks in terms of tem-plate matching, and allows students to ex-plore overfitting. Students can demonstratethat, with enough hidden units, a one-hidden-layer neural network will “memorize” faces(see Fig. 1).
Students start exploring the visualization ofdeep neural networks by computing the gra-dient of a network’s output with respect toits input using TensorFlow and visualizing it.For the bonus component of the assignment,students are asked to explore a visualiza-tion strategy of their choice for understandingthe convolutional network they train. GuidedBackpropagation (Springenberg et al., 2014))is a common technique for students to attemptto implement.
3http://github.com/guerzh/tfweights
Conclusion
Deep Learning frameworks should be intro-duced early in machine learning courses.Handout code that trains simple networks canbe given out. Performing transfer learning andvisualizing the neural networks that studentstrain are two interesting tasks that can be as-signed to students.
References
Abadi, M., Barham, P., Chen, J., Chen, Z.,Davis, A., Dean, J., . . . Zheng, X. (2016).TensorFlow: A System for Large-ScaleMachine Learning. In USENIX Sympo-sium on Operating Systems Design andImplementation (Vol. 16, pp. 265–283).
Astrachan, O., Bruce, K., Koffman, E., Kolling,M., & Reges, S. (2005). Resolved: ob-jects early has failed. In ACM SIGCSEBulletin (Vol. 37, pp. 451–452).
Krizhevsky, A., Sutskever, I., & Hinton, G. E.(2012). Imagenet classification with deepconvolutional neural networks. In Ad-vances in Neural Information ProcessingSystems (pp. 1097–1105).
Springenberg, J. T., Dosovitskiy, A., Brox, T.,& Riedmiller, M. (2014). Striving for sim-plicity: The all convolutional net. arXivpreprint arXiv:1412.6806.
Zeiler, M. D., & Fergus, R. (2014). Visu-alizing and Understanding ConvolutionalNetworks. In European Conference onComputer Vision (pp. 818–833).
Michael Guerzhoy is aLecturer in the Centerfor Statistics and Ma-chine Learning at Prince-ton University, an As-sistant Professor (StatusOnly) in the Dept. ofStatistical Sciences at theUniversity of Toronto, anda Scientist at the Li Ka
Shing Knowledge Institute, St. Michael’s Hos-pital. His professional interests are in com-puter science and data science education andin applications of machine learning to health-care.
15
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
AI Education Matters: Teaching Search AlgorithmsNathan R. Sturtevant (University of Alberta; [email protected])DOI: 10.1145/3284751.3284757
Introduction
In this column, we share advice and resourcesfor teaching and learning about heuristicsearch algorithms. These are algorithms likeA* (Hart, Nilsson, & Raphael, 1968), IDA*(Korf, 1985) and others that use admissibleheuristics to find optimal and suboptimal so-lutions to shortest path problems.
While A* and IDA* are well-known and well-covered in many courses, they can still be con-fusing to students studying them for the firsttime. But, there is a significant body of workthat goes beyond these basic algorithms thatcan be used as additional projects and to en-courage advanced students. These includestopics such as:
• Improving search with better heuristics(Culberson & Schaeffer, 1996; Sturtevant,Felner, Barer, Schaeffer, & Burch, 2009) andconstraints (Goldberg, Kaplan, & Werneck,2006; Rabin & Sturtevant, 2016)
• Exploiting path symmetries on grids(Harabor & Grastien, 2011; Sturtevant& Rabin, 2016)
• Any-angle path planning (Nash & Koenig,2013)
• 3D Path Planning (Brewer & Sturtevant,2018)
• Suboptimal path planning (Hatem & Ruml,2014)
• Bidirectional path planning (Sturtevant &Felner, 2018)
• Multi-agent path planning (Felner et al.,2017)
Textbooks
There are two common textbooks that providean introduction to heuristic search.
Russell and Norvig’s Artificial Intelligence: AModern Approach (Russell & Norvig, 2009)covers the basics of search with heuristics andCopyright c© 2018 by the author(s).
gives several examples of different heuristicsfor different domains. This book is relativelyhigh-level with mostly introductory material.
On the opposite spectrum is StefanEdelkamp’s Textbook, Heuristic Search: The-ory and Applications (Edelkamp & Schrodl,2012). This excellent book goes into greatdetail about most aspects of heuristic search,however it may be too advanced for mostundergraduates.
Video Introductions
There are many videos available on YouTubewhich demonstrate the A* algorithm and gothrough examples of how it works. Our MovingAI web pages1 feature videos on different as-pects of search algorithms that are both avail-able on YouTube or can be download for offlineuse.
Interactive Web Demos
More useful than static videos are interactiveweb pages that allow users to interact withthe search process by changing problem in-stances, search algorithms, and other param-eters.
Amit Patel has developed a broad range oftutorials for subjects of interest to game de-velopers. Among these, he has many differ-ent pages that explore and explain search al-gorithms, some statically2 and some interac-tively.3
Two other notable pages have been developedby Xueqiao (Joe) Xu4 and Dave Churchill5.
1https://www.movingai.com/2http://theory.stanford.edu/
˜amitp/GameProgramming/3https://www.redblobgames.com/
pathfinding/a-star/introduction.html4https://qiao.github.io/
PathFinding.js/visual/5http://www.cs.mun.ca/˜dchurchill/
search/
16
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Each of these pages allow you to interact witha variety of search algorithms and search pa-rameters, visualizing how the algorithms run.
Finally, we have recently released a broadset of demos6 that we have been usingfor our cross-listed undergraduate/graduatecourse on heuristic search. In addition to theregular A*-like algorithms, this page also hasdemos of IDA*, bidirectional search, heuris-tics, constraints, and other search techniques.These demos are still under active develop-ment, particularly when we teach our heuristicsearch course. In addition to providing thesedemos, we give them in class with challengequestions, such as ‘Find a problem where A*updates the g−cost of a state on the open list’.This gives student more purpose when explor-ing them.
Model AI Assignments
Model AI Assignments are free, peer-reviewedassignment materials made available in or-der to advance AI education. There are sev-eral model assignments that explore aspectsof heuristic search.
Sven Koenig’s has two Model AI assignmentsavailable. “Any-Angle Path Planning”7 isfor undergraduate or graduate artificial intelli-gence classes and covers algorithms that usea grid representation of the obstacles in theworld when agents are able to move throughfree space.
“Fast Trajectory Replanning”8 is also for un-dergraduate and graduate students and cov-ers algorithms for re-planning when the worldrepresentation changes or the agents discoverobstacles they were not previously aware of.
“The Pac-Man Projects” by John DeNero andDan Klein also has a heuristic search project,“Project #1: Search Project”9 in which “stu-dents implement depth-first, breadth-first, uni-form cost, and A* search algorithms. Thesealgorithms are used to solve navigation and
6https://movingai.com/SAS/7http://idm-lab.org/project-m/
project2.html8http://idm-lab.org/project-m/
project1.html9http://modelai.gettysburg.edu/
2010/pacman/projects/search/search.html
traveling salesman problems in the Pac-Manworld. ”
Conferences for Heuristic Search
For those looking for recent research in thearea of heuristic search, there are severalvenues that regularly publish work in heuris-tic search. These include the Symposium onCombinatorial Search (SoCS)10, which spe-cializes particularly in search. The Interna-tional Conference on Planning and Schedul-ing (ICAPS)11 has a broader scope, but regu-larly contains significant research in heuristicsearch. Finally, major conferences like the In-ternational Joint Conference on Artificial Intel-ligence (IJCAI)12 and the AAAI Conference onArtificial Intelligence13 also publish research inthis subfield of Artificial Intelligence.
References
Brewer, D., & Sturtevant, N. R. (2018).Benchmarks for pathfinding in 3d voxelspace. Symposium on CombinatorialSearch (SoCS), 143 – 147.
Culberson, J. C., & Schaeffer, J. (1996).Searching with pattern databases. Ad-vances in Artificial Intelligence, 402-416.
Edelkamp, S., & Schrodl, S. (2012). Heuris-tic search - theory and applications.Academic Press. Retrieved fromhttp://www.elsevierdirect.com/product.jsp?isbn=9780123725127
Felner, A., Stern, R., Shimony, S. E., Boyarski,E., Goldenberg, M., Sharon, G., . . .Surynek, P. (2017). Search-based op-timal solvers for the multi-agent pathfind-ing problem: Summary and challenges..
Goldberg, A. V., Kaplan, H., & Werneck, R. F.(2006). Reach for A*: Efficient point-to-point shortest path algorithms. InALENEX (pp. 129–143).
Harabor, D. D., & Grastien, A. (2011). On-line graph pruning for pathfinding on gridmaps. In AAAI (pp. 1114–1119).
10http://search-conference.org11http://www.icaps-conference.org12http://www.ijcai.org/13http://www.aaai.org/Conferences/
conferences.php
17
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Hart, P., Nilsson, N. J., & Raphael, B. (1968).A formal basis for the heuristic deter-mination of minimum cost paths. IEEETransactions on Systems Science andCybernetics, 4, 100–107.
Hatem, M., & Ruml, W. (2014). Simplerbounded suboptimal search. In AAAI(pp. 856–862).
Korf, R. E. (1985). Iterative-Deepening-A*: Anoptimal admissible tree search. In IJCAI(p. 1034-1036).
Nash, A., & Koenig, S. (2013). Any-angle pathplanning. AI Magazine, 34(4), 9.
Rabin, S., & Sturtevant, N. R. (2016). Com-bining bounding boxes and jps to prunegrid pathfinding. In AAAI.
Russell, S., & Norvig, P. (2009). Artificial in-telligence: A modern approach (3rd ed.).Upper Saddle River, NJ, USA: PrenticeHall.
Sturtevant, N. R., & Felner, A. (2018). A briefhistory and recent achievements in bidi-rectional search. In Aaai conference onartificial intelligence.
Sturtevant, N. R., Felner, A., Barer, M., Scha-effer, J., & Burch, N. (2009). Memory-based heuristics for explicit state spaces.IJCAI, 609–614.
Sturtevant, N. R., & Rabin, S. (2016). Canon-ical orderings on grids. IJCAI, 683–689.
Nathan R. Sturtevant isa Professor of ComputerScience at the Universityof Alberta doing researchin heuristic search, ar-tificial intelligence, andgames. He has imple-mented his research incommercial games and isa senior member of theAssociation for the Ad-
vancement of Artificial Intelligence.
18
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
AI Profiles: An Interview with Kristian KerstingMarion Neumann (Washington University in St. Louis; [email protected])DOI: 10.1145/3284751.3284758
Introduction
This column is the sixth in our series profilingsenior AI researchers. This month we inter-view Kristian Kersting, Professor in ComputerScience and Deputy Director of the Centre forCognitive Science at the Technical Universityof Darmstadt, Germany.
Figure 1: Kristian Kersting
Biography
After receiving his Ph.D. from the University ofFreiburg in 2006, he was with the MIT, Fraun-hofer IAIS, the University of Bonn, and the TUDortmund University. His main research in-terests are statistical relational artificial intel-ligence (AI), probabilistic deep programming,and machine learning. Kristian has publishedover 170 peer-reviewed technical papers andco-authored a book on statistical relational AI.He received the European Association for Arti-ficial Intelligence (EurAI, formerly ECCAI) Dis-sertation Award 2006 for the best AI disser-tation in Europe and two best-paper awards(ECML 2006, AIIDE 2015). He gave severaltutorials at top AI conferences, co-chaired sev-eral international workshops, and cofoundedthe international workshop series on Statisti-cal Relational AI (StarAI). He regularly serves
Copyright c© 2018 by the author(s).
on the PC (often at senior level) for several topconference and co-chaired the PC of ECMLPKDD 2013 and UAI 2017. He is the Special-ity Editor in Chief for Machine Learning and AIof Frontiers in Big Data, and is/was an actioneditor of TPAMI, JAIR, AIJ, DAMI, and MLJ.
Getting to Know Kristian Kersting
When and how did you become interestedin AI?
As a student, I was attending an AI course ofBernhard Nebel at the University of Freiburg.This was the first time I dived deep into AI.However, my interest in AI was probably trig-gered earlier. Around the age of 16, I think, Iwas reading about AI in some popular sciencemagazines. I did not get all the details, but Iwas fascinated.
What professional achievement are youmost proud of?
We were collaborating with biologists on un-derstanding better how plants react to (a)bioticstress using machine learning to analyze hy-perspectral images. We got quite encourag-ing results. The first submission to a jour-nal, however, got rejected. As you can imag-ine, I was disappointed. One of the biologistsfrom our team looked at me and said ”Kristian,do not worry, your research helped us a lot.”This made me proud. But also the joint workwith Martin Mladenov on compressing linearand quadratic programs using fractional auto-morpishms. This provides optimization flagsfor ML and AI compilers. Turning them onmakes the compilers attempt to reduce thesolver costs, making ML and AI automaticallyfaster.
What would you have chosen as yourcareer if you hadn’t gone into CS?
Physics, I guess, but back then I did not seeany other option than Computer Science.
19
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
What do you wish you had known as aPh.D. student or early researcher?
That “sleep is for post-docs,” as MichaelLittman once said.
Artificial Intelligence = Machine Learning.What’s wrong with this equation?
Machine Learning (ML) and Artificial Intelli-gence (AI) are indeed similar, but not quite thesame. AI is about problem solving, reason-ing, and learning in general. To keep it sim-ple, if you can write a very clever program thatshows intelligent behavior, it can be AI. But un-less the program is automatically learned fromdata, it is not ML. The easiest way to think oftheir relationship is to visualize them as con-centric circles with AI first and ML sitting inside(with deep learning fitting inside both), sinceML also requires to write programs, namely,of the learning process. The crucial point isthat they share the idea of using computationas the language for intelligent behavior.
As you experienced AI research andeducation in the US and in Europe, whatare the biggest differences between thetwo systems and what can we learn fromeach other?
If you present a new idea, US people will usu-ally respond with “Sounds great, let’s do it!”,while the typical German reply is “This won’twork because ...”. Here, AI is no exception.It is much more critically received in Germanythan in the US. However, this also provides re-search opportunities such as transparent, fairand explainable AI. Generally, over the pastthree decades, academia and industry havebeen converging philosophically and physi-cally much more in the US than in Germany.This facilitate the transfer of AI knowledge viawell-trained, constantly learning AI experts,who can then continuously create new ideaswithin the company/university and keep pacewith the AI development. To foster AI researchand education, the department structure andtenure-track system common in the US is ben-eficial. On the other hand, Germany is offeringaccess to free higher education to all students,regardless of their origin. AI has no borders.We have to take it from the ivory towers andmake it accessible for all.
What is the most interesting project youare currently involved with?
Deep learning has made striking advancesin enabling computers to perform tasks likerecognizing faces or objects, but it does notshow the general, flexible intelligence that letspeople solve problems without being speciallytrained to do so. Thus, it is time to boost itsIQ. Currently, we are working on deep learn-ing approaches based on sum-product net-works and other arithmetic circuits that ex-plicitly quantify uncertainty. Together withcolleagues—also from the Centre of CognitiveScience—we combining the resulting proba-bilistic deep learning with probabilistic (logical)programming languages. If successful, thiswould be a big step forward in programminglanguages, machine learning and AI.
AI is grown up - it’s time to make use of itfor good. Which real-world problem wouldyou like to see solved by AI in the future?
Due to climate change, population growth andfood security concerns the world has to seekmore innovative approaches to protecting andimproving crop yield. AI should play a majorrole here. Next to feeding a hungry world,AI should aim to help eradicate disease andpoverty.
We currently observe many promising andexciting advances in using AI ineducation, going beyond automatingPiazza answering, how should we makeuse of AI to teach AI?
AI can be seen as an expanding and evolv-ing network of ideas, scholars, papers, codesand showcases. Can machines read thisdata? We should establish the “AI Genome”,a dataset, a knowledge base, an ongoing ef-fort to learn and reason about AI problems,concepts, algorithms, and experiments. Thiswould not only help to curate and personal-ize the learning experience but also to meetthe challenges of reproducible AI research. Itwould make AI truly accessible for all.
What is your favorite AI-related movie orbook and why?
“Ex Machina” because the Turing test is shap-ing its plot. It makes me think about current
20
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
real-life systems that give the impression thatthey pass the test. However, I think AI is hardthan many people think.
Help us determine whoshould be in the AI Mat-ters spotlight!
If you have suggestionsfor who we should pro-file next, please feel freeto contact us via email [email protected].
21
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
AI Policy MattersLarry Medsker (George Washington University; [email protected])DOI: 10.1145/3284751.3284759
Abstract
AI Policy is a regular column in AI Mattersfeaturing summaries and commentary basedon postings that appear twice a month in theAI Matters blog (https://sigai.acm.org/aimatters/blog/). Selected posts are summarized in is-sues of AI Matters.
Introduction
The SIGAI Public Policy goals are to
• promote discussion of policies related to AIthrough posts in the AI Matters blog on the1st and 15th of each month,
• help identify external groups with commoninterests in AI Public Policy,
• encourage SIGAI members to partner inpolicy initiatives with these organizations,and
• disseminate public policy ideas to the SIGAImembership through articles in the newslet-ter.
I welcome everyone to make blog commentsso we can develop a rich knowledge base ofinformation and ideas representing the SIGAImembers.
WEF Report on the Future of Jobs
The World Economic Forum recently releaseda report (link) on the future of jobs . Their anal-yses refer to the Fourth Industrial Revolution(link) and their Centre for the Fourth IndustrialRevolution. The report states that
The Fourth Industrial Revolution is in-teracting with other socio-economic anddemographic factors to create a perfectstorm of business model change in all in-dustries, resulting in major disruptions tolabour markets. New categories of jobswill emerge, partly or wholly displacingothers. The skill sets required in both old
Copyright c© 2018 by the author(s).
and new occupations will change in mostindustries and transform how and wherepeople work. It may also affect female andmale workers differently and transform thedynamics of the industry gender gap. TheFuture of Jobs Report aims to unpack andprovide specific information on the relativemagnitude of these trends by industry andgeography, and on the expected time hori-zon for their impact to be felt on job func-tions, employment levels and skills.
The report concludes that by 2022 more jobscan be created than the number lost but thatvarious stakeholders, including those makingeducation policy, must make wise decisions.
Vehicle Automation: Safe Design,Scientific Advances, and Smart Policy
Updating previous columns on terminologyand popular discourse about AI, a currentexample is the impact on policy of the waywe talk about automation. “Unmanned Au-tonomous Vehicle (UAV)” is a term that jus-tifiably creates fear in the general public, buttalk about a UAV usually misses the neces-sary roles of humans and human decisionmaking. Likewise, discussions about an “au-tomated decision maker (ADM)” ignores thesocial and legal responsibilities of those whodesign, manufacture, implement, and oper-ate “autonomous” systems. The AI commu-nity has an important role to influence correctand realistic use of concepts and issues in dis-cussions of science and technology systemsthat increase automation. The concept “hybridsystem” might be more helpful here for under-standing the potential and limitations of com-bining technologies and humans into AI andAutonomous Systems (AI/AS) requiring less,but not zero, need for humans over time.
Safe Design
In addition to avoiding confusion and man-aging expectations, design approaches andanalyses of the performance of existing sys-tems with automation are crucial to developing
22
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
safe systems with which the public and policy-makers can feel comfortable. In this regard,stakeholders should read information on de-sign of systems with automation components,such as the IEEE report “Ethically Aligned De-sign” (link), which is subtitled “A Vision for Pri-oritizing Human Wellbeing with Artificial Intel-ligence and Autonomous Systems”. The re-port says about AI and Autonomous Systems(AI/AS):
We need to make sure that these tech-nologies are aligned to humans in termsof our moral values and ethical princi-ples. AI/AS have to behave in a way thatis beneficial to people beyond reachingfunctional goals and addressing technicalproblems. This will allow for an elevatedlevel of trust between humans and ourtechnology that is needed for a fruitful per-vasive use of AI/AS in our daily lives.
See also Ben Shneiderman’s excellent sum-mary and comments (link) on the report aswell as the YouTube video of his Turing In-stitute Lecture on “Algorithmic Accountability:Design for Safety”. See also his proposal for aNational Algorithms Safety Board (link).
Advances in AI/AS Science and Technology
Another perspective on the automation issueis the need to increase safety of systemsthrough advances in science and technology.In a future blog, we will present the transcriptof an interview with Dr. Harold Szu, about theneed for a next generation of AI that movescloser to brain-style computing and that incor-porates human behaviors into AI/AS systems.Dr. Szu was the founder and former president,and former governor, of the International Neu-ral Network Society. He is acknowledged foroutstanding contributions to ANN applicationsand scientific innovations.
Policy and Ethics
Over the summer 2018, increased activity incongress and state legislatures focused onunderstandings, accurate and not, of “un-manned autonomous vehicles” and what poli-cies should be in place. The following exam-ples are interesting for possible interventions,but also for the use of AI/AS terminology:
• House Energy & Commerce Committee’s
press release (link): the SELF DRIVE Act.• CNBC Commentary (link) by Reps. Bob
Latta (R-OH) and Jan Schakowsky (D-IL).• Politico, 08/03/2018.: “Trial lawyers speak
out on Senate self-driving car bill”, by Bri-anna Gurciullo with help from Lauren Gard-ner.
AV NON-STARTER: After being mumfor months, the American Associationfor Justice said publicly Thursday thatit has been pressing for the Senate’sself-driving car bill, S. 1885 (115) (def-initions on p.42 link), to stipulate thatcompanies can’t force arbitration, ourTanya Snyder reports for Pros. The triallawyers group is calling for a provisionto make sure when a person, whethera passenger or pedestrian, is injured orkilled by a driverless car, that personor their family is not forced into a se-cret arbitration proceeding, according toa statement. Senate Commerce Chair-man John Thune (R-S.D.) has said thatarbitration has been “a thorny spot” inbill negotiations.
ACM Code of Ethics
On Tuesday, July 17, ACM announced the up-dated Code of Ethics and Professional Con-duct. Please note the message from ACMHeadquarters and check the link below: “Wewould like your support in helping to reach asbroad an audience of computing professionalsas possible with this news.” The updated Codeis at link.
We encourage you to share the updated Codewith your friends and colleagues at that time.If you use social media, please take partin the conversation around computing ethicsusing the hashtags #ACMCodeOfEthics and#IReadTheCode. And if you are not doingso already, please follow the @TheOfficialACMand @ACM Ethics Twitter handles to shareand engage with posts about the Code.
News From the ACM US TechnologyPolicy Committee
The USACM has a new name Please notethe change and remember that SIGAI will con-tinue to have a close relationship with the ACM
23
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
US Technology Policy Committee (link). Hereis a reminder of the purpose and goals: “TheACM US Technology Policy Committee is aleading independent and nonpartisan voice inaddressing US public policy issues related tocomputing and information technology. TheCommittee regularly educates and informsCongress, the Administration, and the courtsabout significant developments in the comput-ing field and how those developments affectpublic policy in the United States. The Com-mittee provides guidance and expertise in var-ied areas, including algorithmic accountabil-ity, artificial intelligence, big data and analyt-ics, privacy, security, accessibility, digital gov-ernance, intellectual property, voting systems,and tech law. As the internet is global, theACM US Technology Policy Committee workswith the other ACM policy entities on publica-tions and projects related to cross-border is-sues, such as cybersecurity, encryption, cloudcomputing, the Internet of Things, and internetgovernance.”
The ACM US Technology Policy Commit-tee’s New Leadership ACM named Prof. JimHendler as the new Chair of the ACM U.S.Technology Policy Committee (formerly US-ACM) under the new ACM Technology PolicyCouncil. In addition to being a distinguishedcomputer science professor at RPI, Jim haslong been an active USACM member and hasserved as both a committee chair and as anat-large representative. This is a great choiceto guide USACM into the future within ACM’snew technology policy structure. Please addindividually to the SIGAI Public Policy congrat-ulations to Jim. Our congratulations and ap-preciation go to outgoing Chair Stuart Shapirofor his outstanding leadership of USACM.
Larry Medsker is Re-search Professor ofPhysics and founding di-rector of the Data Sciencegraduate program at theData Science graduateprogram at The GeorgeWashington University.Dr. Medsker is a formerDean of the Siena Col-
lege School of Science, and Professor inComputer Science and in Physics, wherehe was a co-founder of the Siena Institutefor Artificial Intelligence. His research andteaching continues at GW on the nature ofhumans and machines and the impacts of AIon society and policya
ahttp://humac-web.org/. Professor Medskers re-search in AI includes work on artificial neural net-works and hybrid intelligent systems. He is thePublic Policy Officer for the ACM SIGAI.
24
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
The Partnership on AIJeffrey Heer (University of Washington; [email protected])DOI: 10.1145/3284751.3284760
The Partnership on AI to Benefit People andSociety (or PAI) is a consortium of industrial,non-profit, and academic partners for promot-ing public understanding and beneficial useof Artificial Intelligence (AI) technologies. Pri-mary goals of the PAI are “to study and for-mulate best practices on AI technologies, toadvance the public’s understanding of AI, andto serve as an open platform for discussionand engagement about AI and its influenceson people and society.”
PAI membership includes representa-tives from over 70 partner organizations(https://www.partnershiponai.org/partners/), spanning nine countries (pri-marily in North America, Europe, and Asia)and including technology companies ofvarious sizes, academic institutions, and anumber of non-profits (including ACM). TheBoard of Directors represents a numberof large tech companies (Amazon, Apple,Facebook, IBM, Google/DeepMind, Microsoft)as well as non-profits (ACLU, MacArthurFoundation, OpenAI) and academics (ArizonaState University, Harvard, UC Berkeley). Thefounding co-chairs are Eric Horvitz (Microsoft)and Mustafa Suleyman (DeepMind), withday-to-day operations of the PAI run by adedicated staff in San Francisco, CA.
The PAI’s research and outreach efforts centeraround six thematic pillars:
• Safety-Critical AI
• Fair, Transparent, and Accountable AI
• AI, Labor, and the Economy
• Collaborations Between People and AI Sys-tems
• Social and Societal Influences of AI
• AI and Social Good
For more details regarding each thematic pil-lar, see https://www.partnershiponai.org/about/#our-work.
Copyright c© 2018 by the author(s).
The PAI kick-off meeting was held in Berlinin October 2017, with wide-ranging discus-sions organized around the above pillars. Aprimary of outcome of the meeting and sub-sequent membership surveys was the forma-tion of three initial working groups focused on“Safety-Critical AI”, “Fair, Transparent, and Ac-countable AI”, and “AI, Labor, and the Econ-omy”. Each of these groups are now conduct-ing a series of meetings to develop best prac-tices and resources for their respective topics.
I am a member of the Fair, Transparent, andAccountable AI (or FTA) working group, co-chaired by Ed Felten (Princeton) and VerityHarding (DeepMind). The charter of the FTAworking group is to study issues and oppor-tunities for promoting fair, accountable, trans-parent, and explainable AI, including the de-velopment of standards and best practices asa way to avoid and minimize the risk that AIsystems will undermine fairness, equality, dueprocess, civil liberties, or human rights. TheFTA group has held two in-person meetingsso far: the first in London in May 2018 andthe second at Princeton in August 2018. Theinitial meeting resulted in the formation of fourspecific projects.
The “Papers” project aims to produce a set ofthree educational primers, one each for fair-ness, transparency, and accountability. Thegoal is to provide a starting point for people tolearn about fairness in AI, how it is conceived,and what people are currently doing about it.The purpose is not just to come up with def-initions, but provide structure for how to thinkabout problems of fairness, transparency andaccountability in the context of the work thatPartner organizations are conducting.
The “Case Studies” project focuses of identi-fying multiple case studies that illustrate howFTA concerns can be addressed, as wellas develop a framework for the analysis ofsuch case studies. Meanwhile, the “GrandChallenges” project is working to create chal-lenges or endorse existing contests developedaround FTA issues in practice.
25
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
The “Diverse Voices” project seeks to under-stand different communities, how their voicescan be involved in tech policy making, andthe impact of AI deployment. The project in-tends to engage with these groups in a mean-ingful manner and ensure under-representedgroups’ voices are heard. One goal is to pro-vide guidance for outreach, as many compa-nies are interested in deploying ethical AI butare unsure how to address the problem.
Each of these projects is a work-in-progress,with additional milestones and meetingsplanned over the coming year.
Jeffrey Heer is a Pro-fessor of Computer Sci-ence & Engineering at theUniversity of Washingtonand Co-Founder of Tri-facta Inc., a provider ofinteractive tools for scal-able data transformation.
His research spans human-computer interac-tion, visualization, data science, and interac-tive machine learning. He also serves as anACM representative to the Partnership on AI.
26
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Mechanism Design for Social GoodRediet Abebe (Cornell University; [email protected])Kira Goldner (University of Washington; [email protected])DOI: 10.1145/3284751.3284761
Introduction
Across various domains—such as health, ed-ucation, and housing—improving societal wel-fare involves allocating resources, setting poli-cies, targeting interventions, and regulatingactivities. These solutions have an immenseimpact on the day-to-day lives of individuals,whether in the form of access to quality health-care, labor market outcomes, or how votes areaccounted for in a democratic society. Prob-lems that can have an outsized impact onindividuals whose opportunities have histori-cally been limited often pose conceptual andtechnical challenges, requiring insights frommany disciplines. Conversely, the lack of inter-disciplinary approach can leave these urgentneeds unaddressed and can even exacerbateunderlying socioeconomic inequalities.
To realize the opportunities in these domains,we need to correctly set objectives and reasonabout human behavior and actions. Doing sorequires a deep grounding in the field of in-terest and collaboration with domain expertswho understand the societal implications andfeasibility of proposed solutions. These in-sights can play an instrumental role in propos-ing algorithmically-informed policies. In manycases, the input data for our algorithms maybe generated by strategic and self-interestedindividuals who have a stake in the outcomeof the algorithm. To get around this issue, wecan deploy techniques from mechanism de-sign, which uses game theory to align incen-tives or analyze the strategic behavior of indi-viduals who interact with the algorithms.
The Mechanism Design for Social Good(MD4SG) research agenda is to address prob-lems for which insights from algorithms, op-timization, and mechanism design have thepotential to improve access to opportunity.These include allocating affordable housingservices, designing efficient health insurancemarkets, setting subsidies to alleviate eco-nomic inequality, and several other issues af-
Copyright c© 2018 by the author(s).
fecting many individuals’ livelihoods. Thisresearch area falls at the interface of artifi-cial intelligence, theoretical computer science,and the social sciences. Since the fall of2016, the authors of this piece have been co-organizing the Mechanism Design for SocialGood research group, workshop series, andcolloquium series Abebe and Goldner [2016,2018]. The group comprises a large net-work of researchers from various disciplines,including computer science, economics, soci-ology, operations research, and public policy.Members of the group partner with domain ex-perts in non-government organizations, thinktanks, companies, and other entities with ashared mission. The mission is to explore newfrontiers, garner interest in directions in whichalgorithmic and mechanism design insightshave been under-utilized but have the po-tential to inform innovative interventions, andhighlight exemplary work.
In this piece, we discuss three exciting re-search avenues within MD4SG. For each ofthese, we showcase ongoing work, underlinenew directions, and discuss potential for im-plementing existing work in practice.
Access to Opportunity in theDeveloping World
New technologies and data sources are fre-quently leveraged to understand, evaluate,and address societal concerns across theworld. In many developing nations, however,there is a lack of information regarding un-derlying matters—whether that be the preva-lence of diseases or accurate measurementsof economic welfare and poverty—due to theunavailability of high-quality, comprehensive,and reliable data UN [2014]. This limits the im-plementation of effective policies and interven-tions. An emerging solution, which has beensuccessfully demonstrated by the InformationCommunication Technology for Development(ICT4D) research community, has been totake advantage of high phone and Internet
27
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
penetration rates across developing nations todesign new technologies which enable collec-tion and sharing of high-quality data. Therehas also been recent work from within the AIcommunity to use new data sources to closethis information gap Abebe et al. [2018], Jeanet al. [2016]. Such AI-driven approaches sur-face new algorithmic, modeling, and mecha-nism design questions to improve the lives ofmany under-served individuals.
A prominent example is in agriculture, whichaccounts for a large portion of the econ-omy in many developing nations. Here, vi-ral disease attacks on crops is a leadingcause of food insecurity and poverty. Tradi-tional disease surveillance methods fail to pro-vide adequate information to curtail the im-pact of diseases Mwebaze and Biehl [2016],Mwebaze and Owomugisha [2016], Quinnet al. [2011]. The Cassava Adhoc Surveil-lance Project from Makerere University imple-ments crowd-sourcing surveillance using pic-tures taken by mobile phones in order to ad-dress this gap Mutembesa et al. [2018]. Thetool is set up as a game between farmers andother collaborators, and aims to collect truth-ful, high-value data (e.g., data from hard-to-reach locations). This approach underlinesinteresting challenges, such as how to opti-mally incentivize individuals to collect high-quality information and how to augment thisinformation with existing methods. Similarissues arise in other domains—e.g., in citi-zen science and in computational sustainabil-ity Xue et al. [2016a,b]. Finding solutions inthe context of the developing world may there-fore have a broader global impact.
Lack of information also leads to inefficien-cies in existing systems, presenting a possi-bility to introduce solutions that abide by exist-ing cultural and technological constraints. Forinstance, large price discrepancies and majorarbitrage opportunities present in markets foragricultural products in Uganda suggest largemarket inefficiencies Ssekibuule et al. [2013].To alleviate this, Newman et al. [2018] intro-duce Kudu—a mobile technology that func-tions over feature phones via SMS service.Kudu facilitates transactions between farmersin rural areas and buyers at markets in citiesby allowing sellers and buyers to post theirasks and offers. Kudu has been adapted byusers across Uganda and many trades have
been realized through this system.
Availability of new technologies also presentsopportunities to tackle fundamental problemsrelated to poverty. Advances in last-mile pay-ment technologies, for example, enable large-scale, secure cash transfers. GiveDirectlyleverages this and the popularity of mobilemoney across the world to create a systemwhere donors can directly transfer cash to re-cipients GD, Blattman and Niehaus [2014].GiveDirectly moves the decision about how touse aid from policy-makers to recipients, giv-ing recipients maximum flexibility. Such aidgenerates heterogeneity in outcomes—e.g.,families may use aid to start a business, payrent, cover health costs, and so on. Policy-makers used to prioritizing specific outcomesmay be uncomfortable by such a model. A re-search question then is: can we predict how agiven population will use aid? Likewise, howcan we target people for whom the interven-tions will make the largest difference? Aid hashistorically been targeted on the basis of find-ing the most deprived people. The ability tomodel heterogeneous treatment effects opensthe door for designing more nuanced mecha-nisms that fairly and efficiently allocate subsi-dies in order to maximize a desired outcome.
Problems in the developing world surfaceunique challenges at the intersection of AI,ICT4D, and development economics. So-lutions often have to be implemented inresource-constrained environments (e.g., overfeature phones or with low network connectiv-ity) Brunette et al. [2013], Patel et al. [2010].Key populations of interest (e.g., women, peo-ple living in rural parts, individuals with dis-abilities) may not be easily accessible Sultanaet al. [2018], Vashistha et al. [2015b,a]. In-dividuals may have low-literacy Sambasivanet al. [2010]. Lack of understanding of socio-cultural norms and politics, furthermore, mayinhibit proposed interventions Vashistha et al.[2018]. All of these highlight the need fora multi-stakeholder approach that leveragestechnological advances, innovative technicalsolutions, and partnerships with individualsand organizations that will be impacted by thesolutions. MD4SG fosters one such environ-ment in which insights from across these dis-ciplines inform the design of algorithms andmechanisms to improve the lives of individu-als across the world.
28
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Labor, Platforms, and Discrimination
Online platforms are ubiquitous, providing avast playground for algorithm design and arti-ficial intelligence. Every policy decision, how-ever, impacts and interacts with the platform’sstrategic users. In this section, we will focuson online labor markets and how discrimina-tion effects stem from a platform’s decisions.Past work begins to investigate some aspectsof platforms, of strategic agents, and of dis-crimination in labor markets, but there are stillmajor opportunities for work at the intersec-tion, and insights from mechanism design andAI are ripe for the job.
One central issue surrounding labor marketsis that of hiring, in which a firm takes infor-mation about a potential candidate and makesan employment decision. Firms act as classi-fiers, labeling each applicant as “hire” or “nothire” based on an applicant’s “features,” suchas educational investment or a worker’s pro-ductivity reputation. In the process of makinghiring decisions, however, the firm may poten-tially make discriminatory decisions—perhapsby using protected attributes, or by not cor-recting for differences in applications that stemfrom systemic discrimination Bertrand andMullainathan [2004], Marlowe et al. [1996].Bias in hiring decisions may arise due to im-plicit human bias or algorithmic bias, in whichalgorithms replicate human and/or historic dis-crimination that is reflected in the data onwhich they are trained Broussard [2018], Eu-banks [2018], Noble [2018], O’Neil [2016].
One recent line of work investigates hiringpolicies that achieve diversity or statistical par-ity (with respect to certain groups) among thehired workers, and how workers make theirinvestment decisions (e.g. whether to at-tend college) based on the hiring policies inplace. Coate and Loury [1993], Fryer Jr andLoury [2013], Hu and Chen [2017] study set-tings where there is some known underlyingbias or historical discrimination against certaingroups; the aim is to characterize hiring poli-cies that are optimal-subject-to-fair-hiring, andto quantify any loss in efficiency comparedto optimal-but-discriminatory policies. Theseworks explore two settings: first, when hiringdecisions must be “group-blind,” that is, theycannot take group membership into account,and second, when they are “group-aware”.
The aim is to choose hiring policies that willmitigate discrimination against protected cat-egories. Hu and Chen [2017] highlight addi-tional complexity that arises in dynamic set-tings where workers are hired based on invest-ment decisions (e.g. college GPA) in an ini-tial temporary labor market (e.g. internships)and this job creates a worker’s initial produc-tivity reputation that is then used in the per-manent labor market. Many of these findingsalso discuss “trade-offs” between group-blindand group-aware policies.
Another aspect of labor markets is that aworker may have the ability to pay to change afeature of her application in some illegitimateor unfair way in order to improve her outcomein the labor market. Hardt et al. [2016] exam-ine this problem from a robust machine learn-ing perspective. Under certain assumptions ofthe cost required to change an applicant’s rep-utation, they characterize classifiers that opti-mally compare to the original reputation (be-fore the applicant modified it).
These are only two aspects at the interplaybetween hiring and strategic agents; hiring,furthermore, is only one aspect of the labormarket. Consider today’s popular online la-bor markets, such as Mechanical Turk, Up-work, Task Rabbit, and Lyft, in which the plat-form’s goal is to match workers to employersor jobs. In these labor markets, the platform’sdecisions, even at a granular level, have alarge impact on the workers and firms. Con-sider the following platform decisions. Visi-bility: How many firms can workers see at atime? What capacity do they have to searchjob offers? Can workers see jobs and jobs seeworkers? Initiation: Which side (or both) cansubmit applications? Initiate messaging? Setcontract terms? Information: What informa-tion is displayed about parties on the oppositeside? Name? Photo? Ethnicity? Wage his-tory? Reputation?
Each of these decisions impacts theoutcome—not only the quality of the match,but also whether (and how much) discrim-ination occurs. In a recent paper, Levyand Barocas [2017] outline categories ofplatform decisions which may mitigate orperpetuate discrimination in labor markets,including the high-level categories of settingplatform discrimination policies or norms,
29
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
structuring information and interactions, andmonitoring/evaluating discriminatory conduct.
In offline labor markets, it may be challeng-ing or infeasible to collect data to understandthe nature and extent of discrimination. On-line labor markets, on the other hand, yieldrich data about employer-employee interac-tions and present the possibility of conduct-ing experiments aimed at reducing bias anddiscrimination or other desired societal objec-tives. For instance, Barach and Horton [2017]look at the impact on hiring of hiding workers’wage history. Horton and Johari [2015], Hor-ton [2018] look at the impact of trying to elicitadditional information (features) from workersor firms, and the impact of this strategically-reported information on hiring. Horton [2017]examines who the hired worker population iswhen a minimum wage is imposed on oneplatform. Each of these provide insights intolabor dynamics that may inform platform de-sign and interventions.
Online labor markets provide a rich play-ground for techniques from algorithms, AI, andmechanism design to study how each aspectof platform design impacts discriminatory ef-fects, workers’ actions, and the desired objec-tive for the platform.
Allocating Housing and HomelessnessResources
Allocation of resources—such as public hous-ing, housing vouchers, and homelessnessservices—has a long history in the economicsand computation literature. Even simple-to-state problems here have given rise to chal-lenging research questions, many of whichare still open. Increased scarcity of hous-ing resources, growing need for services, andthe use of algorithmic decision-making toolsall open up several avenues with major op-portunities for reforming policies and regula-tions. Here, we discuss some foundationalwork, new challenges, and opportunities thatemerge at the nexus of algorithm and mecha-nism design, AI, and the social sciences.
Millions of individuals across the US havebeen evicted or are at risk of experiencingeviction every year. In groundbreaking work,Desmond [2012, 2016] shows that eviction ismuch more common than was previously doc-
umented. By compiling the first ever evictionsdatabase, Desmond reveals that there is anestimate of 2.3 million evictions in 2016 aloneand argues that eviction is a cause of povertyEL [2018]. Using this database, and othersimilar datasets, we may be able to employ acombination of machine learning and statisti-cal techniques to gain a better understandingof what causes housing instability and home-lessness. We can then build on this work todesign algorithms and mechanisms that canimprove on allocation policies. For instance,Kube et al. [2018] use counter-factual predic-tions to improve homelessness service provi-sions. By doing so, they realize some gains onreducing the number of families experiencingrepeated episodes of homelessness. At thesame time, Eubanks [2018] emphasizes thatcaution must be taken when using automateddecision-making tools for allocating limited re-sources in such high-stakes scenarios. Suchtools may be used to reduce failure rates bycaseworkers, but, if not approached with care,can deepen already existing inequalities. Fur-thermore, the use of such tools alone is lim-ited; it does not address the lack of housingand homelessness resources or eliminate hu-man biases or discrimination. It is crucial totake advantage of the confluence of insightsfrom cross many disciplines in order to servethe needs of such vulnerable populations.
An issue that is growing in prominence inhousing contexts is that of information. Littleis documented about how landlords or hous-ing authorities screen applications and makedecisions. One exception is the work of Am-brose and Diop [2016], which shows that thereis increased restriction in access to rentalhousing since landlords mitigate informationasymmetry by investing in screening tenants.With the increased use and availability of dataabout individuals, it is of paramount impor-tance to understand the role of information inthe decision-making process of entities, suchas landlords or housing agencies, who haveenormous discretion in how and whether fam-ilies are housed.
Although the introduction of automated toolsintroduces acute challenges related to hous-ing, the use of algorithmic techniques datesback several decades and there are many fun-damental problems that remain unsolved. Anearly work here is that of Hylland and Zeck-
30
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
hauser [1979], which considers the “house al-location problem” of assigning each individ-ual to one item, such as a house. They in-troduce a mechanism which satisfies natu-ral efficiency and fairness notions but is notincentive-compatible. That is, individuals maybe able to improve their outcome by misreport-ing their true preferences. Since then, severalmechanisms have been proposed, includingthe popular Randomized Serial Dictatorship(RSD) mechanism, which uses a random lot-tery Abdulkadiroglu and Sonmez [1998]. Thismechanism is used as a standard mechanismin many domains, including housing. While itis incentive-compatible, it fails to satisfy thefairness criteria of Hylland and Zeckhauser[1979]. An important question is then the de-sign of incentive-compatible, fair, and efficientmechanisms for the house allocation problem.
Due to increased scarcity of resources, cur-rent allocation protocols often involves wait-ing lists and priority groups. Policy constraintsmake wait-list design a dynamic rationingproblem rather than the static assignmentproblem discussed above. Dynamic mecha-nisms present several technical and practicalchallenges; e.g., incentive-compatibility maybe infeasible in dynamic settings due to wait-ing time trade-offs for applicants. There areconsequential design decisions related to howto manage wait-lists and different metropolitanareas have different policies (e.g., setting pri-ority groups, conditions under which individu-als are removed from the waiting list, set ofchoices, and many others). Each of these poli-cies impacts the allocation dynamics, waitingtime, and quality of matches. Recent workhas studied how to design mechanisms sat-isfying various desiderata and quantify differ-ences in quality of matches across variousmechanisms Arnosti and Shi [2018], Leshno[2017], Thakral [2016], Waldinger [2017].
Conclusion
As the use of algorithmic and AI techniquesbecomes more pervasive, there is a growingappreciation of the fact that the most impactfulsolutions often fall at the interface of variousdisciplines. The Mechanism Design for SocialGood research agenda is to foster an envi-ronment in which insights from algorithms andmechanism design can, in conjunction with
the social sciences, be used to improve ac-cess to opportunity, especially for communi-ties of individuals for whom opportunities havehistorically been limited. In this piece, we havehighlighted MD4SG research avenues relatedto issues in developing nations, labor markets,and housing. For each of these, we have dis-cussed the need to work in close partnershipwith a wide range of stakeholders to set objec-tives that best address the needs of individ-uals and propose feasible solutions with de-sired societal outcomes. There are numerousother domains in which this kind of interdisci-plinary approach for designing algorithms andmechanisms can improve the lives of manyindividuals; we invite readers to learn morethrough our colloquium and workshop series.
Acknowledgments
We are indebted to members of the Mech-anism Design for Social Good researchgroup–Ellora Derenoncourt, Alon Eden, LilyHu, Manish Raghavan, Sam Taggart, DanielWaldinger, and Matt Weinberg—our co-organizer Irene Lo, and our advisors AnnaKarlin and Jon Kleinberg for their generos-ity in sharing their knowledge and supportthroughout the past two years. We addition-ally thank Mutembesa Daniel, Paul Niehaus,Fabian Okeke, and Aditya Vashistha for help-ful discussions and pointers. We are grate-ful, as ever, for the numerous researchersin the economics and computation researchcommunities for their enthusiastic support andguidance throughout the development of thegroup and workshop series.
ReferencesGivedirectly: Send money directly to the ex-
treme poor. https://givedirectly.org/.
A world that counts: Mobilising the data rev-olution for sustainable development. Tech-nical report, November 2014. Report pre-pared at the request of the United NationsSecretary-General by the Independent Ex-pert Advisory Group on Data Revolution forSustainable Development.
The Eviction Lab. https://evictionlab.org/, 2018.
Atila Abdulkadiroglu and Tayfun Sonmez.
31
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Random serial dictatorship and the corefrom random endowments in house alloca-tion problems. Econometrica, 66(3):689–701, 1998.
Rediet Abebe and Kira Goldner. MechanismDesign for Social Good. http://md4sg.com, 2016.
Rediet Abebe and Kira Goldner. A report onthe workshop on mechanism design for so-cial good. ACM SIGecom Exchanges, 16(2):2–11, 2018.
Rediet Abebe, Shawndra Hill, Jennifer Wort-man Vaughan, Peter M Small, and H An-drew Schwartz. Using search queries to un-derstand health information needs in africa.arXiv preprint arXiv:1806.05740, 2018.
Brent W Ambrose and Moussa Diop. Informa-tion asymmetry, regulations, and equilibriumoutcomes: Theory and evidence from thehousing rental market. 2016.
Nick Arnosti and Peng Shi. Design of lotteriesand waitlists for affordable housing alloca-tion. 2018.
Moshe Barach and John J Horton. How doemployers use compensation history?: Evi-dence from a field experiment. 2017.
Marianne Bertrand and Sendhil Mullainathan.Are Emily and Greg more employable thanLakisha and Jamal? a field experiment onlabor market discrimination. American eco-nomic review, 94(4):991–1013, 2004.
Christopher Blattman and Paul Niehaus.Show them the money: Why giving cashhelps alleviate poverty. Foreign Affairs, 93(3):117–126, 2014.
Meredith Broussard. Artificial Unintelligence:How Computers Misunderstand the World.MIT Press, 2018.
Waylon Brunette, Mitchell Sundt, Nicola Dell,Rohit Chaudhri, Nathan Breit, and GaetanoBorriello. Open data kit 2.0: expanding andrefining information services for developingregions. In Proceedings of the 14th Work-shop on Mobile Computing Systems andApplications, page 10. ACM, 2013.
Stephen Coate and Glenn C Loury. Willaffirmative-action policies eliminate nega-tive stereotypes? The American EconomicReview, pages 1220–1240, 1993.
Matthew Desmond. Eviction and the repro-duction of urban poverty. American Journalof Sociology, 118(1):88–133, 2012.
Matthew Desmond. Evicted: Poverty andprofit in the American city. Broadway Books,2016.
Virginia Eubanks. Automating Inequality: HowHigh-tech Tools Profile, Police, and Punishthe Poor. St. Martin’s Press, 2018.
Roland G Fryer Jr and Glenn C Loury. Valuingdiversity. Journal of Political Economy, 121(4):747–774, 2013.
Moritz Hardt, Nimrod Megiddo, Christos Pa-padimitriou, and Mary Wootters. Strate-gic classification. In Proceedings of the2016 ACM conference on innovations in the-oretical computer science, pages 111–122.ACM, 2016.
John J Horton. Price floors and employer pref-erences: Evidence from a minimum wageexperiment. 2017.
John J Horton. Buyer uncertainty about sellercapacity: Causes, consequences, and apartial solution. 2018.
John J Horton and Ramesh Johari. At whatquality and what price?: Eliciting buyer pref-erences as a market design problem. InProceedings of the Sixteenth ACM Con-ference on Economics and Computation,pages 507–507. ACM, 2015.
Lily Hu and Yiling Chen. Fairness at equi-librium in the labor market. arXiv preprintarXiv:1707.01590, 2017.
Aanund Hylland and Richard Zeckhauser. Theefficient allocation of individuals to posi-tions. Journal of Political economy, 87(2):293–314, 1979.
Neal Jean, Marshall Burke, Michael Xie,W Matthew Davis, David B Lobell, and Ste-fano Ermon. Combining satellite imageryand machine learning to predict poverty.Science, 353(6301):790–794, 2016.
Amanda R Kube, Sanmay Das, and Patrick J.Fowler. Allocating interventions based oncounterfactual predictions: A case study onhomelessness services. 2018.
Jacob Leshno. Dynamic matching in over-loaded waiting lists. 2017.
Karen Levy and Solon Barocas. Design-ing against discrimination in online markets.Berkeley Tech. LJ, 32:1183, 2017.
32
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
Cynthia M Marlowe, Sandra L Schneider, andCarnot E Nelson. Gender and attractive-ness biases in hiring decisions: Are moreexperienced managers less biased? Jour-nal of applied psychology, 81(1):11, 1996.
Daniel Mutembesa, Christopher Omongo, andErnest Mwebaze. Crowdsourcing real-timeviral disease and pest information: A caseof nation-wide cassava disease surveillancein a developing country. In HCOMP, pages117–125, 2018.
Ernest Mwebaze and Michael Biehl.Prototype-based classification for imageanalysis and its application to crop diseasediagnosis. In Advances in Self-OrganizingMaps and Learning Vector Quantization,pages 329–339. Springer, 2016.
Ernest Mwebaze and Godliver Owomugisha.Machine learning for plant disease inci-dence and severity measurements from leafimages. In Machine Learning and Ap-plications (ICMLA), 2016 15th IEEE Inter-national Conference on, pages 158–163.IEEE, 2016.
Neil Newman, Lauren Falcao Bergquist,Nicole Immorlica, Kevin Leyton-Brown,Brendan Lucier, Craig McIntosh, JohnQuinn, and Richard Ssekibuule. Design-ing and evolving an electronic agriculturalmarketplace in Uganda. In Proceedings ofthe 1st ACM SIGCAS Conference on Com-puting and Sustainable Societies, page 14.ACM, 2018.
Safiya Umoja Noble. Algorithms of Op-pression: How Search Engines ReinforceRacism. NYU Press, 2018.
Cathy O’Neil. Weapons of Math Destruc-tion: How Big Data Increases Inequality andThreatens Democracy. Broadway Books,2016.
Neil Patel, Deepti Chittamuru, Anupam Jain,Paresh Dave, and Tapan S Parikh. Avaajotalo: a field study of an interactive voice fo-rum for small farmers in rural india. In Pro-ceedings of the SIGCHI Conference on Hu-man Factors in Computing Systems, pages733–742. ACM, 2010.
John Alexander Quinn, Kevin Leyton-Brown,and Ernest Mwebaze. Modeling and moni-toring crop disease in developing countries.In AAAI, 2011.
Nithya Sambasivan, Ed Cutrell, KentaroToyama, and Bonnie Nardi. Intermediatedtechnology use in developing communities.In Proceedings of the SIGCHI Conferenceon Human Factors in Computing Systems,CHI ’10, pages 2583–2592, New York, NY,USA, 2010. ACM.
Richard Ssekibuule, John A Quinn, and KevinLeyton-Brown. A mobile market for agricul-tural trade in uganda. In Proceedings of the4th Annual Symposium on Computing forDevelopment, page 9. ACM, 2013.
Sharifa Sultana, Francois Guimbretiere,Phoebe Sengers, and Nicola Dell. Designwithin a patriarchal society: Opportuni-ties and challenges in designing for ruralwomen in bangladesh. In Proceedingsof the 2018 CHI Conference on HumanFactors in Computing Systems, CHI ’18,pages 536:1–536:13. ACM, 2018.
Neil Thakral. The public-housing allocationproblem. Technical report, Technical report,Harvard University, 2016.
Aditya Vashistha, Edward Cutrell, GaetanoBorriello, and William Thies. Sangeet swara:A community-moderated voice forum in ru-ral india. In Proceedings of the 33rd An-nual ACM Conference on Human Factors inComputing Systems, CHI ’15, pages 417–426. ACM, 2015a.
Aditya Vashistha, Edward Cutrell, Nicola Dell,and Richard Anderson. Social media plat-forms for low-income blind people in in-dia. In Proceedings of the 17th InternationalACM SIGACCESS Conference on Comput-ers and Accessibility, ASSETS ’15, pages259–272. ACM, 2015b.
Aditya Vashistha, Fabian Okeke, Richard An-derson, and Nicola Dell. ’you can always dobetter!: The impact of social proof on partic-ipant response bias. In Proceedings of the2018 CHI Conference on Human Factors inComputing Systems, page 552. ACM, 2018.
Daniel Waldinger. Targeting in-kind transfersthrough market design: A revealed prefer-ence analysis of public housing allocation.2017.
Yexiang Xue, Ian Davies, Daniel Fink, Christo-pher Wood, and Carla P. Gomes. Avi-caching: A two stage game for bias reduc-tion in citizen science. In Proceedings of
33
AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018
the 2016 International Conference on Au-tonomous Agents and Multiagent Systems,AAMAS ’16, pages 776–785, 2016a.
Yexiang Xue, Ian Davies, Daniel Fink, Christo-pher Wood, and Carla P. Gomes. Behav-ior identification in two-stage games for in-centivizing citizen science exploration. InMichel Rueher, editor, Principles and Prac-tice of Constraint Programming, pages 701–717, Cham, 2016b. Springer InternationalPublishing.
Rediet Abebe is a PhDcandidate in computerscience, advised by JonKleinberg. Her researchfocuses on algorithms,artificial intelligence, andtheir applications to socialgood. In particular, sheuses algorithmic and
computational insights to better understandsocioeconomic inequality and inform interven-tions for improving access to opportunity. Herwork is generously supported by fellowshipsand scholarships through Facebook, Google,and the Cornell Graduate School. Prior toCornell, she completed an M.S. and a B.A. inApplied Mathematics and Mathematics fromHarvard University and an M.A. in Mathemat-ics from the University of Cambridge. Shewas born and raised in Addis Ababa, Ethiopia.
Kira Goldner is a fifth-year PhD student at theUniversity of Washingtonin the Department ofComputer Science andEngineering, advisedby Anna Karlin. Herresearch focuses onproblems in mechanism
design, particularly on (1) maximizing rev-enue in settings that are motivated by practiceand (2) on mechanism design within healthinsurance and online labor markets. She is a2017-19 recipient of the Microsoft ResearchPhD Fellowship and was a 2016 recipient of aGoogle Anita Borg Scholarship. Kira receivedher B.A. in Mathematics from Oberlin Collegeand also studied at Budapest Semesters inMathematics.
34