ai matters, volume 5, issue 3 5(3) 2019 ai...

AI MATTERS, VOLUME 5, ISSUE 3 5(3) 2019

AI MattersAnnotated Table of Contents

Welcome to AI Matters 5(3)Amy McGovern, co-editor & IolandaLeite, co-editor

Full article: http://doi.acm.org/10.1145/3362077.3362078

Welcome and summary

ACM SIGAI Activity ReportSven Koenig, Sanmay Das, Rose-mary Paradis, John Dickerson, YolandaGil, Katherine Guo, Benjamin Kuipers,Iolanda Leite, Hang Ma, Nicholas Mat-tei, Amy McGovern, Larry Medsker,Todd Neller, Marion Neumann, PlamenPetrov, Michael Rovatsos & David Stork


ACM SIGAI Activity Report

Help Communities Solve Real-WorldProblems with AI – Become a Tech-novation Mentor!Tara Chklovski


Mentors needed to help families, schools andcommunities to learn, play, and create with AI

EventsMichael Rovatsos


Upcoming AI events

AI Education Matters: Building aFake News DetectorMichael Guerzhoy, Lisa Zhang &Georgy Noarov


An AI assignment for building a Fake News Detec-tor

AI Education Matters: A First Intro-duction to Modeling and Learning us-ing the Data Science WorkflowMarion Neumann


Modeling and Learning using the Data ScienceWorkflow

AI Policy MattersLarry Medsker


Policy issues relevant to SIGAI

Advancing Non-Convex and Con-strained Learning: Challenges andOpportunitiesTianbao Yang


Latest research trends in AI

Considerations for AI Fairness forPeople with DisabilitiesShari Trewin, Sara Basson, MichaelMuller, Stacy Branham, Jutta Trevi-ranus, Daniel Gruen, Daniel Hebert, Na-talia Lyckowski & Erich Manser


Contributed article

The Intersection of Ethics and AIAnnie Zhou


Winning essay from the 2018 ACM SIGAI Student

1


Essay Contest

Artificial Intelligence: The SocietalResponsibility to Inform, Educate,and RegulateAlexander D. Hilton


Winning essay from the 2018 ACM SIGAI StudentEssay Contest

The Necessary Roadblock to Artifi-cial General Intelligence: CorrigibilityYat Long Lo, Chung Yu Woo & Ka LokNg


Winning essay from the 2018 ACM SIGAI StudentEssay Contest

AI Fun MattersAdi Botea


AI generated Crosswords

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: 10.1145/3362077

Join SIGAI

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, Co-Editor, U. OklahomaIolanda Leite, Co-Editor, KTHSanmay 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]

2


Contents Legend

Book Announcement

Ph.D. Dissertation Briefing

AI Education

Event Report

Hot Topics

Humor

AI Impact

AI News

Opinion

Paper Précis

Spotlight

Video or Image

Details at http://sigai.acm.org/aimatters

3

AI MATTERS, VOLUME 5, ISSUE 3 SEPTEMBER 2019

Welcome to AI Matters 5(3)Amy McGovern, co-editor (University of Oklahoma; [email protected])Iolanda Leite, co-editor (Royal Institute of Technology (KTH); [email protected])DOI: 10.1145/3362077.3362078

Issue overview

Welcome to the third issue of the fifth volumeof the AI Matters Newsletter. With this issue,we want to welcome our new SIGAI Execu-tive Committee. Elections were completed thisSpring, and we have a new leadership teamin place. Sanmay Das of Washington Uni-versity in St. Louis (former vice-chair) is thenew chair, Nicholas Mattei of Tulane Univer-sity the new vice-chair, and John Dickerson ofthe University of Maryland the new secretary-treasurer. Nicholas and John were formerlyactive as the appointed AI and Society and La-bor Market officers respectively. Sven Koenigwill transition into the role of past-chair, andcontinue to serve on the EC in that role.

The new officers want to express their sincerethanks to Sven Koenig and to Rosemary Par-adis (the former secretary/treasurer) for thewealth of novel initiatives they spearheaded inthe last three years and the untiring energythey brought to their roles. SIGAI is deeplyindebted to them!

We would like to mention that there has been alot of activity in the space of significant awardsin AI. The inaugural SIGAI Industry Award forExcellence in Artificial Intelligence (AI) waspresented at IJCAI 2019. The award went tothe Real World Reinforcement Learning Teamfrom Microsoft, for identification and develop-ment of cutting-edge research on contextual-bandit learning that led to new decision sup-port tools that were broadly integrated into abroad range of Microsoft products. John Lang-ford and Tyler Clintworth received the awardon behalf of the Microsoft team and presenteda talk on the work at IJCAI. For more on thisaward, please see https://sigai.acm.org/awards/industry award.html

We also congratulate Marijn Heule, MattiJärvisalo, Florian Lonsing, Martina Seidl andArmin Biere who have been awarded the 2019IJCAI-JAIR prize for their 2015 paper “Clause

Copyright c© 2019 by the author(s).

Elimination for SAT and QSAT” (https://jair.org/index.php/jair/article/view/10942). This pa-per describes fundamental and practical re-sults on a range of clause elimination pro-cedures as preprocessing and simplificationtechniques for SAT and QBF solvers. Since itspublication, the techniques described thereinhave been demonstrated to have profound im-pact on the efficiency of state-of-the-art SATand QBF solvers. The work is elegant and ex-tends beautifully some well-established theo-retical concepts. In addition, the paper givesnew emphasis and impulse to pre- and in-processing techniques - an emphasis that res-onates beyond the two key problems, SAT andQBF, covered by the authors.

We would also like to note that SIGAI andAAAI will be jointly presenting a new annualaward for the best doctoral dissertation in AI.The award will be presented at AAAI, andnominations for the inaugural award are dueby November 15, 2019. Please see http://sigai.acm.org/awards/nominations.html for details andinformation on how to submit a nomination!

This issue is full of great new articles and sto-ries for you! We open with the annual report ofSIGAI. We then bring you a story from a newway to teach kids and families about AI: Tech-novation Familes’ AI challenge, which bringsAI into the home by educating parents andchildren about AI and providing an opportunityfor them to prototype AI solutions to real-worldproblems. They are seeking new mentors forthis year’s challenges!

In our regular articles, Michael Rovatsos re-ports on upcoming AI events and we have twosubmissions for AI Education. First, MichaelGuerzhoy talks about building a fake news de-tector. Second, Marion Neumann talks aboutbringing AI and ML to a younger audience,much like CS for all, instead of focusing on se-niors and graduate students. In the policy col-umn, Larry Medsker summarizes recent poli-cies covering face recognition (how much datashould we record and share?), upcoming AI

4


regulation, and more. Our final regular columnis our AI crosswords from Adi Botea. Enjoy!

We have a new regular column where weinvite researchers to present latest researchtrends in AI. In the inaugural article of this col-umn, Tianbao Yang describes challenges andopportunities of non-convex and constrainedlearning.

In our contributed articles, Shari Trein et al.describe some of the opportunities and risksacross four emerging AI application areas:employment, education, public safety, andhealthcare, identified in a workshop with par-ticipants experiencing a range of disabilities.Finally, this issue features the second set ofwinning essays from the 2018 ACM SIGAIStudent Essay Contest. In addition to hav-ing their essay appear in AI Matters, the con-test winners received either monetary prizesor one-on-one Skype sessions with leading AIresearchers.

Special Issue: AI For Social GoodRecognizing the potential of AI in solv-ing some of the most pressing challengesfacing our society, we are excited to an-nounce that the next Newsletter of AI Mat-ters will be a special issue on the themeof “AI for Social Good.” We solicit arti-cles that discuss how AI applications and/orinnovations have resulted in a meaning-ful impact on a societally relevant prob-lem, including problems in the domains ofhealth, agriculture, environmental sustain-ability, ecological forecasting, urban plan-ning, climate science, education, socialwelfare and justice, ethics and privacy, andassistive technology for people with dis-abilities. We also encourage submissionson emerging problems where AI advanceshave the potential to influence a transfor-mative change, and perspective articlesthat highlight the challenges faced by cur-rent standards of AI to have a societal im-pact and opportunities for future researchin this area. More details to be coming soonon http://sigai.acm.org/aimatters. Pleaseget in touch with us if you have any ques-tions!

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 co-editor 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.

Iolanda Leite is co-editorof AI Matters. She is anAssistant Professor at theSchool of Electrical En-gineering and ComputerScience at the KTH RoyalInstitute of Technology inSweden. Her research in-terests are in the areas of

Human-Robot Interaction and Artificial Intelli-gence. She aims to develop autonomous so-cially intelligent robots that can assist peopleover long periods of time.

5


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 (elected; 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)Iolanda Leite (appointed; ACM SIGAI Newsletter Editor-in-Chief)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)Marion Neumann (appointed; ACM SIGAI Diversity Officer)Plamen Petrov (appointed; ACM SIGAI Industry Liaison Officer)Michael Rovatsos (appointed; ACM SIGAI Conference Coordination Officer)David Stork (appointed; ACM SIGAI Award Officer)DOI: 10.1145/3362077.3362079

Abstract

We are happy to present the annual activityreport of the ACM Special Interest Group onAI (ACM SIGAI), covering the period from July2018 to June 2019.

The scope of ACM SIGAI consists of thestudy of intelligence and its realization incomputer systems (see also its web-site 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 androbotics.

Members come from academia, industry andgovernment agencies worldwide. ACM SIGAIrecently added two new ACM SIGAI chap-ters, namely one professional chapter in La-guna Nigel (USA) and one student chapter atthe SRM Institute of Science & Technology inChennai.Copyright c© 2019 by the author(s).

ACM SIGAI also added two new officers thisyear to be able to serve its membership better,namely Iolanda Leite from the Royal Instituteof Technology (Sweden) as second newslet-ter editor-in-chief and Marion Neumann fromWashington University in St. Louis (USA) asdiversity officer, thus increasing diversity inthe ACM SIGAI leadership committee by in-creasing both the number of international of-ficers and the number of female officers andalso furthering the internationalization of theACM SIGAI newsletter. Marion will be cover-ing diversity also as part of the ACM SIGAInewsletter. ACM SIGAI started officers meet-ings at major AI conferences already in 2018and continued the new practice in 2019 (sofar at the AAAI Conference), in addition to all-officers teleconferences and a monthly ACMSIGAI leadership teleconference.

Meetings

ACM SIGAI decided to participate on a trialbasis in ACM’s voluntary carbon-offset pro-gram for conferences. Introduction of thisscheme will give conference participants theoption of making voluntary contributions to off-set the carbon footprint of their trips to confer-

6


ences when they register online. ACM SIGAIplans to test this scheme at upcoming edi-tions of the AAAI/ACM AI, Ethics and Soci-ety (AIES) and ACM Intelligent User Interfaces(IUI) conferences in cooperation with AAAIand SIGCHI, respectively, and hopes that itwill enable the ACM SIGAI and wider ACMmembership to contribute to the environmen-tal sustainability of our communities.

ACM SIGAI continues to support AIES, whichit co-founded in 2017 to fill a scientific void. AsAI is becoming more pervasive in our lives, itsimpact on society is more significant, raisingethical concerns and challenges regarding is-sues such as value alignment, safety and se-curity, data handling and bias, regulations, ac-countability, transparency, privacy and work-force displacement. Only a multi-disciplinaryand multi-stakeholder effort can find the bestways to address these concerns, by includ-ing experts from various disciplines, such asethics, philosophy, economics, sociology, psy-chology, law, history and politics. AIES wasco-located with AAAI 2019 in Honolulu and willagain be co-located with AAAI 2020 in NewYork City.

ACM SIGAI sponsored the following confer-ences in addition to AIES 2019:

• WI 2018• ASE 2018• IVA 2018• HRI 2019• IUI 2019

and it will sponsor the following conferencescoming up in 2019 and 2020:

• IVA 2019• K-CAP 2019• ASE 2019• WI 2019• ASE 2020• HRI 2020• IUI 2020

ACM SIGAI approved the following in-cooperation and sponsorship requests fromevents covering a wide thematic and geo-graphical range across the international AIcommunity:

• iWOAR 2018• ICPRAM 2018• IEA/AIE 2019• FW 2018• BIOSTEC 2019• RecSys 2019• FW 2019• ICAART 2019• AAMAS 2019• iWOAR 2019• KMIKS 2019• ICPRAM 2019• FDG 2019• ICAIL 2019• IC3K 2019• IJCCI 2019• IEA/AIE ’20• AAMAS 2020ACM SIGAI also organizes – jointly with theAssociation for the Advancement of AI (AAAI)– the annual joint job fair at the AAAI con-ference, where attendees can find out aboutjob and internship opportunities from repre-sentatives from industry, universities and otherorganizations. The AAAI/ACM SIGAI jobfair was held at AAAI 2019 in Honolulu, co-organized by the ACM SIGAI labor market of-ficer. Twenty-six employers formally attended,while a handful of exhibitors who did not for-mally sign up also took part. Hundreds ofCVs and resumes were collected before, dur-ing and after the job fair from students, post-doctoral researchers and other job seekersvia the job fair web-site; these were sharedwith interested employers. This year, the or-ganizers also purchased a dedicated domain(aaaijobfair.com) to allow present and fu-ture firms and participants to view previous it-erations of the job fair. The ACM SIGAI la-bor market officer believes that we can use in-sights from AI to create an even better func-tioning job market and works actively towarddesigning the job market of the future. Towardthat end, he has begun to gather requirementswith a large number of chairs of top computerscience departments in the USA as well as inIsrael and Europe and is working to formulatea model that will be translated into a larger jobfair (in terms of participating employers as wellas applicants) in the near future.

7


ACM SIGAI also co-sponsors – jointly withAAAI – the annual joint doctoral consortium atthe AAAI conference, which provides an op-portunity for Ph.D. students to discuss theirresearch interests and career objectives withthe other participants and a group of estab-lished AI researchers who act as their men-tors. The AAAI/ACM SIGAI doctoral consor-tium was held at AAAI 2019 in Honolulu.

Awards

ACM SIGAI sponsors the ACM SIGAI Au-tonomous Agents Research Award, an annualaward for excellence in research in the area ofautonomous agents. The recipient is invited togive a talk at the International Conference onAutonomous Agents and Multiagent Systems(AAMAS). The 2019 ACM SIGAI AutonomousAgents Research Award was presented at AA-MAS 2019 in Montreal to Carles Sierra, thevice-director of the AI Research Institute of theSpanish National Research Council, for sem-inal contributions to research on negotiationand argumentation, computational trust andreputation and artificial social systems.

ACM SIGAI also sponsors the ACM SIGAIIndustry Award for Excellence in AI, a newannual award which is given annually to anindividual or team in industry who createda fielded AI application in recent years thatdemonstrates the power of AI techniques viaa combination of the following features: nov-elty of application area, novelty and techni-cal excellence of the approach, importanceof AI techniques for the approach and ac-tual and predicted societal impact of the ap-plication. The inaugural ACM SIGAI IndustryAward for Excellence in AI will be presented atthe International Joint Conference on AI (IJ-CAI) 2019 in Macau to the Real World Rein-forcement Learning Team from Microsoft forthe identification and development of cutting-edge research on contextual-bandit learning,the manifest cooperation between researchand development efforts, the applicability ofthe decision support throughout the broadrange of Microsoft products and the quality ofthe final systems.

ACM SIGAI also recently created – jointly withAAAI – the joint AAAI/ACM SIGAI DoctoralDissertation Award to recognize and encour-age superior research and writing by doctoral

candidates in AI. This new annual award willbe presented at the AAAI Conference on AIin the form of a certificate and is accompa-nied by the option to present the dissertationat the AAAI conference as well as to submit asix-page summary to both the AAAI proceed-ings and the ACM SIGAI newsletter. The nom-ination deadline for the inaugural AAAI/ACMSIGAI Doctoral Dissertation Award will be an-nounced later this year and is expected to bein late Fall 2019.

Public Policy Activities

ACM SIGAI promotes the discussion of poli-cies related to AI through posts in the AI Mat-ters blog, helps to identify external groupswith common interests in AI public policy,encourages ACM SIGAI members to part-ner in policy initiatives with these organiza-tions, disseminates public policy ideas to theACM SIGAI membership through articles inthe ACM SIGAI newsletter and ensures thatevery technologist is educated, trained andempowered to prioritize ethical considerationsin the design and development of autonomousand intelligent systems. ACM SIGAI partici-pates in the ACM US Technology Policy Com-mittee (ACM USTPC), formerly USACM, viathe ACM SIGAI public policy officer and in avariety of other policy efforts, including thoseof other societies (such as the IEEE Global Ini-tiative on Ethics of Autonomous and IntelligentSystems). ACM USTPC addresses US publicpolicy issues related to computing and infor-mation technology and regularly educates andinforms US Congress, the US Administrationand the US courts about significant develop-ments in the computing field and how thosedevelopments affect public policy. For exam-ple, the ACM SIGAI public policy officer joinedthe comments of ACM USTPC on the draftof the “20-Year Roadmap for AI Research inthe US” of the Computing Community Con-sortium. He also studies how organizationscollect and analyze data and whether thesepractices are consistent with recommenda-tions by the USTPC working group on algo-rithmic accountability, transparency and bias.He represented ACM SIGAI via his talks “Fu-ture of Work, AI Education, and Public Policy”at EAAI 2019 and “Transparency, Accessibil-ity, and Ethics in AI” at Dalhousie Universityin 2019. He was also the moderator of the

8


panel “Are We Ready for AI” at the AnnualConsumer Assembly of the Consumer Feder-ation of America in 2019.

Educational Activities

ACM SIGAI held a second ACM SIGAI Stu-dent Essay Contest focused on AI ethics (or-ganized by the ACM SIGAI AI and society offi-cer), after the success of the first such compe-tition in 2017. Students could win cash prizesof US$500 or Skype conversations with se-nior AI researchers from academia or industry(including the director of Microsoft ResearchLabs and the director of research at Google) iftheir essays provided good answers to one orboth of the following topic areas (or any otherquestion in this space that they considered im-portant):

• What requirements, if any, should be imposedon AI systems and technology when interactingwith humans who may or may not know that theyare interacting with a machine? For example,should they be required to disclose their identi-ties? If so, how?

• What requirements, if any, should be imposedon AI systems and technology when making de-cisions that directly affect humans? For exam-ple, should they be required to make transparentdecisions? If so, how?

This year, ACM SIGAI received 18 submis-sions, of which eight were selected for publica-tion and prizes. The winning essays are listedbelow in alphabetical order by author. ACMSIGAI intents to hold a third ACM SIGAI stu-dent Essay Contest later this year.

• Janelle Berscheid and Francois Roewer-Despres – Beyond Transparency: A ProposedFramework for Accountability in Decision-Making AI Systems

• Gage Garcia – AI Education: A Requirement fora Strong Democracy

• Alexander Hilton – AI: The Societal Responsibil-ity to Inform, Educate, and Regulate

• Michelle Seng Ah Lee – Context-ConsciousFairness in Using Machine Learning to MakeDecisions

• Yat Long Lo, Chung Yu Woo and Ka Lok Ng –The Necessary Roadblock to Artificial GeneralIntelligence: Corrigibility

• Grace McFassel – Embedding Ethics: Design ofFair AI Systems

• Matthew Sun and Marissa Gerchick – TheScales of (Algorithmic) Justice: Tradeoffs andRemedies

• Annie Zhou – The Intersection of Ethics and AIACM SIGAI supported the “Birds of theFeather” undergraduate research challengeorganized by the ACM SIGAI education officerat the Symposium on Educational Advancesin AI (EAAI) 2019. Six research and liberalarts institutions participated with seven papersand one poster presentation that passed peerreview. ACM SIGAI contributed US$500 ofaward funding for the best papers. The ACMSIGAI education officer intends to announcea Gin Rummy undergraduate research chal-lenge at EAAI 2020.

ACM SIGAI also started discussions with theACM Special Interest Group on Computer Sci-ence Education (ACM SIGCSE) on a collabo-ration for disseminating pointers to resourcesfor AI educators and creating incentives for theproduction and dissemination of assignmentson AI ethics.

Member Communication

ACM SIGAI communicates with its membersvia email announcements, the ACM SIGAInewsletter “AI Matters,” the AI Matters blogand webinars:

ACM SIGAI maintains a more than 4,000email address long mailing list for AI-relatedannouncements to its members and friends.

ACM SIGAI publishes four issues of itsnewsletter AI Matters per year. TheACM SIGAI newsletter is distributed via theACM SIGAI mailing list but also openlyavailable on the ACM SIGAI web-site (atsigai.acm.org/aimatters/). AI Mattersfeatures articles of general interest to the AIcommunity and added not only an additionaleditor-in-chief but also additional column edi-tors in the past year. Recent columns, led bythese and other column editors, have includedAI Interviews (organized by the ACM SIGAIdiversity officer), AI Amusements, AI Educa-tion (written or organized by the ACM SIGAIeducation officer), AI Policy Issues (writtenby the ACM SIGAI public policy officer) andAI Events (written by the ACM SIGAI confer-ence coordination officer). The editors-in-chiefhave recently added an AI crossword puzzle

9


(thanks to Adi Botea from IBM’s Ireland Re-search Laboratory) and are about to add a col-umn on current research trends in AI, writtenby recent grantees of research funds (such asNSF CAREER or European Research Councilawards). AI Matters has also started to pub-lish the winning student essays of the secondACM SIGAI Student Essay Contest.

ACM SIGAI also maintains an AI Matters blog(at sigai.acm.org/aimatters/blog/)as a forum for important announcements andnews. For example, the ACM SIGAI publicpolicy officer posts new information every twoweeks in the blog to survey and report oncurrent AI policy issues and raise awarenessabout the activities of other organizations thatshare interests with ACM SIGAI.

After a hiatus due to the illness of oneof the organizers, ACM SIGAI recently re-started the ACM SIGAI webinars with awebinar on “Advances in Socio-BehavioralComputing” and several more in prepara-tion. The webinars are streamed live but thevideos can still be watched on demand atlearning.acm.org/webinar/.

ACM SIGAI is also a founding member of AIHub (at aihub.org), a new non-profit siblingto Robohub (at robohub.org) dedicated toconnecting the AI communities of the world bybringing together experts in AI research, start-ups, business and education from across theglobe. Content-area specialists will curate allincoming AI news articles to make sure thatreporting is truthful, fair and balanced, andin-house editors will ensure that all contentmeets the highest editorial standards for lan-guage and clarity. AI Hub is expected to comeonline in Summer or Fall 2019. ACM SIGAIwill provide content to AI Hub and, conversely,AI Hub will provide AI news to the ACM SIGAImembers.

Financial Member Support

ACM SIGAI so far had concentrated its finan-cial support on travel scholarships to ACMSIGAI student members to allow them to at-tend conferences if they are otherwise missingthe financial resources to do so. The amountsof the scholarships vary but are generally inthe range of US$1,000 to US$10,000 per con-ference, depending on the conference size.

The ACM SIGAI conference coordination of-ficer recently started to test a new open stu-dent award travel scheme. Beyond provid-ing a ringfenced allocation to specific confer-ences, he created a process by which anyACM SIGAI student member who intends toattend an ACM (and, in exceptional cases,even a non-ACM) event can apply for travelsupport through the ACM SIGAI web-site. Inthe first few months since the inception of thescheme, students have already been offeredfinancial support of about US$8,000 in total.

ACM SIGAI also recently created the AI Activ-ities Fund, a new initiative to empower ACMSIGAI members and friends to organize ac-tivities with a strong outreach component toeither students, researchers or practitionersnot working on AI technologies or to the pub-lic in general. The purpose of the inauguralcall for funding proposals was to help ACMSIGAI members and friends to promote a bet-ter understanding of current AI technologies,including their strengths and limitations as wellas their promise for the future. Examples offundable activities included (but were not lim-ited to) AI technology exhibits or exhibitions,holding meetings with panels on AI technology(including on AI ethics) with expert speakers,creating podcasts or short films on AI tech-nologies that are accessible to the public andholding AI programming competitions. ACMSIGAI was looking for evidence that the infor-mation presented by the activities would be ofhigh quality, accurate, unbiased (for example,not influenced by company interests) and atthe right level for the intended audience. Theinaugural call for proposals supported the fol-lowing initiatives: a workshop on “AI for All us-ing the R Programming Language” organizedby the Indian Institute of Technology in Goa,the “Bee Network of AI” organized by the Uni-versidad Mayor in Chile and “Co-Opting AI:Public Conversations about Design, Inequalityand Technology” organized by New York Uni-versity.

Additional Member Services

ACM SIGAI also supports its members inadditional ways. For example, it nominatesthem for awards or supports their nominations.ACM SIGAI is proud of the fact that many AIresearchers in the past year received ACM

10


honors, such as becoming ACM senior mem-bers, distinguished members and fellows aswell as receiving other awards. Three AI re-searchers received the A.M. Turing Award in2018.

ACM SIGAI also actively supports the Re-search Highlight Track of the Communicationsof the ACM (CACM) by nominating publica-tions of recent, significant and exciting AI re-search results that are of interest to the com-puter science research community in generalto the Research Highlight Track. This way,ACM SIGAI helps to make important AI re-search results visible to many computer sci-entists.

Additional ACM SIGAI membership benefitsinclude reduced registration fees at many ofthe co-sponsored and in-cooperation confer-ences and access to the proceedings of manyof these conferences in the ACM Digital Li-brary.

Planning for the Future

ACM SIGAI held elections for a new chair, vicechair and secretary/treasurer in Spring 2019.Sanmay Das (the current ACM SIGAI vicechair) was elected ACM SIGAI chair, NicholasMattei (the current ACM SIGAI AI and soci-ety officer) was elected ACM SIGAI vice-chair,and John Dickerson (the current ACM SIGAIlabor market officer) was elected ACM SIGAIsecretary/treasurer. Sven Koenig (the currentACM SIGAI chair) will transition to his new roleas ACM SIGAI past chair. We are looking for-ward to the new leadership committee shap-ing the future of ACM SIGAI. In general, ACMSIGAI intends to reach out to more AI groupsworldwide that could benefit from ACM sup-port, such as providing financial support, mak-ing the proceedings widely accessible in theACM Digital Library and providing speakersvia the ACM Distinguished Speakers program.ACM SIGAI also intends to reach out more toother disciplines that share an interest in AI,for example, in terms of research methodolo-gies or applications.

11


Help Communities Solve Real-World Problems with AI – Becomea Technovation Mentor!Tara Chklovski (Founder and CEO, Technovation; [email protected])DOI: 10.1145/3362077.3362080

Abstract

Join other AI professionals as a mentor inTechnovation’s AI program for families. Shareyour expertise with adults and children whoare curious about artificial intelligence andhow it can be used to address real issues intheir communities. Help people all around theworld learn to not only use AI, but to createsolutions with AI that improve their lives andtheir communities.

Technovation Families

Technovation, a global technology educationnonprofit, is seeking mentors for its sec-ond season of Technovation Families, an AI-focused program for families with childrenages 8-13. Join peers working in ComputerScience fields and be part of the world’slargest AI mentoring program. Support fami-lies around the world as they learn about AIand develop AI-based prototypes addressingproblems they identify in their communities.

Started in 2018, Technovation Families’ AIchallenge brings families, schools, communi-ties, and mentors together to learn, play, andcreate with AI. They apply what they learn tosolve a real-world problem in their commu-nity as part of a global competition. Mentorsguide learners of all ages through Neural Net-works, data, self-driving car algorithms, andmachine learning and training models to rec-ognize images, text, and emotions through anIBM-Watson based platform Machine Learn-ing for Kids. Local educators and CS expertsoffer encouragement and guidance through-out as families learn about – and use – AI toolsfor the first time. Mentors especially are a criti-cal touch-point for families who are developingtheir confidence as problem solvers and inven-tors, and who have big ideas for applying AI tocommunity problems, but lack technical knowl-edge, experience, and confidence in their abil-ities. Mentors are able to volunteer remotely,


thereby strengthening local capacity in areasthat may not have access to technology pro-fessionals or universities.

In the first year of Technovation Families’ AIcompetition, 7,500 people across 70 chap-ters in 13 countries participated, developing200 AI-based solutions to problems in theircommunities. These solutions ranged fromimage-recognition software that scans chil-dren’s drawings for signs of bullying and awearable swimming cap for kids to detect earlysigns of drowning, to a tool to detect and re-move invasive algae from a local lake (Figure1 and Figure 2).

Figure 1: Jeff Dean (Head of Google Brain) lis-tening to a father and son team describing theirimage-recognition prototype that emits ultrasonicfrequencies when it sees a dog.

Figure 2: Six coaches from Bolivia, Palestine,Spain, United States, Pakistan and Kazakhstanwho coached their communities to create winningAI-based inventions.

12


Technovation partners with industry leadersincluding Google, NVIDIA, Intel, General Mo-tors, and the Patrick J. McGovern Foundation,to bridge the AI knowledge and confidencegap for children and adults around the world(Figure 3).

Figure 3: Mother and sons from Bolivia explainingto a judge how their Raspberry-Pi powered, imagerecognition system sucks up invasive weeds fromLake Titicaca.

The Technovation Families program is built ona community-based model that involves par-ents and caregivers so that the adults (in ad-dition to the children) can reignite their curios-ity and develop as lifelong learners. After par-ticipating in the program, more than 91% ofthe parents surveyed believed their child de-veloped a sustained interest and growing in-terest in AI and ∼85% of parents wanted tocontinue investing effort into improving theirlocal communities (Chklovski, Jung, Fofang, &Gonzales, 2019).

Mentors benefit too. Through Technovation’sprograms, mentors’ communications and pre-sentation skills improve, as do their profes-sional relationships with colleagues and lead-ers. And, their work with Technovation partic-ipants stretches and grows their creativity, or-ganizational, and project management skills.

Recently, the second season of TechnovationFamilies launched debuting an expanded cur-riculum developed in partnership with a com-mittee of AI researchers, and industry profes-sionals. The updated curriculum includes ad-ditional information about AI, good and baddatasets, machine learning and ethical inno-vation (Figure 4 and Figure 5). Through 10fun, hands-on lessons, families learn founda-tional AI concepts, identify a meaningful prob-

lem to solve in their local community, and buildan AI agent to solve it. Sign-up today to helpfamilies make the world a better place with AI!

Figure 4: First 5-weeks of project-based Techno-vation Families AI curriculum that introduces learn-ers to AI, Machine Learning, and building Imageand Text Recognition Systems.

Figure 5: Last 5-weeks of the Technovation Fami-lies AI curriculum that helps participants apply theirlearning to create AI-based prototypes that ad-dress problems in their community.

ReferencesChklovski, T., Jung, R., Fofang, J. B., & Gon-

zales, P. (2019). Implementing a 15-week ai-education program with under-resourced families across 13 global com-munities. In International joint confer-ence on artificial intelligence, macau.

13


Tara Chklovsk TaraChklovski is CEO andfounder of global techeducation nonprofit Tech-novation. A frequentadvocate for STEM edu-cation, she’s presented atthe White House STEMInclusion Summit, SXSWEDU, UNESCO’s MobileLearning Week, and ledthe education track atthe 2019 UN AI for Good

Global Summit.

14


EventsMichael Rovatsos (University of Edinburgh; [email protected])DOI: 10.1145/3362077.3362081

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 areare organizing an event and would be inter-ested in cooperating with SIGAI. For moreinformation about conference support visitsigai.acm.org/activities/requesting sponsor-ship.html.

2nd International Conference onArtificial Intelligence & Virtual Reality(AIVR 2019)San Diego, CA, December 9-11, 2019http://ieee-aivr.orgThe AIVR conference, now in its second run,is a unique event, addressing researchers andindustries from all areas of AI as well as Vir-tual, Augmented, and Mixed Reality. It pro-vides an international forum for the exchangebetween those fields to present advances inthe state of the art, identify emerging re-search topics, and together define the futureof these exciting research domains. We inviteresearchers from VR, as well as AugmentedReality (AR) and Mixed Reality (MR) to par-ticipate and submit their work to the program.Likewise, any work on AI that has a relation toany of these fields or potential for the usage inany of them is welcome.

9th International Conference onPattern Recognition Applications andMethods (ICPRAM ’20)Setúbal, Portugal, February 22-24, 2020http://www.icpram.org/The International Conference on PatternRecognition Applications and Methods wouldlike to become a major point of contact be-tween researchers, engineers and practition-ers on the areas of Pattern Recognition,both from theoretical and application perspec-tives. Contributions describing applicationsof Pattern Recognition techniques to real-world problems, interdisciplinary research, ex-


perimental and/or theoretical studies yieldingnew insights that advance Pattern Recognitionmethods are especially encouraged.Submission deadline: October 4, 2019

13th International Joint Conference onBiomedical Engineering Systems andTechnologies (BIOSTEC 2020)Valetta, Malta, February 24-26, 2020http://www.biostec.org/The purpose of BIOSTEC is to bring togetherresearchers and practitioners, including en-gineers, biologists, health professionals andinformatics/computer scientists, interested inboth theoretical advances and applications ofinformation systems, artificial intelligence, sig-nal processing, electronics and other engi-neering tools in knowledge areas related to bi-ology and medicine. BIOSTEC is composedof five co-located conferences, each special-ized in a different knowledge area.Submission deadline: October 4, 2019

12th International Conference onAgents and Artificial Intelligence(ICAART 2020)Valetta, Malta, February 24-26, 2020http://www.icaart.org/The purpose of ICAART is to bring togetherresearchers, engineers and practitioners inter-ested in the theory and applications in the ar-eas of Agents and Artificial Intelligence. Twosimultaneous related tracks will be held, cov-ering both applications and current researchwork. One track focuses on Agents, Multi-Agent Systems and Software Platforms, Dis-tributed Problem Solving and Distributed AI ingeneral. The other track focuses mainly onArtificial Intelligence, Knowledge Representa-tion, Planning, Learning, Scheduling, Percep-tion Reactive AI Systems, and EvolutionaryComputing and other topics related to Intelli-gent Systems and Computational Intelligence.Submission deadline: October 4, 2019

15


15th ACM/IEEE InternationalConference on Human-RobotInteraction (HRI 2020)Cambridge, UK, March 23-26, 2020http://humanrobotinteraction.org/2020/HRI 2020 is the 15th annual conferencefor basic and applied human-robot interac-tion research. Researchers from across theworld present their best work to HRI to ex-change ideas about the theory, technology,data, and science furthering the state-of-the-art in human-robot interaction. Each year, theHRI Conference highlights a particular areathrough a theme. The theme of HRI 2020 is“Real World Human-Robot Interaction”. TheHRI conference is a highly selective annualinternational conference that aims to show-case the very best interdisciplinary and mul-tidisciplinary research in human-robot inter-action with roots in and broad participationfrom communities that include but are not lim-ited to robotics, artificial intelligence, human-computer interaction, human factors, design,and social and behavioral sciences. Submis-sion deadline: October 1, 2019

22nd International Conference onEnterprise Information Systems (ICEIS2020)Prague, Czech Republic, May 5-7, 2020http://www.iceis.org/The purpose of ICEIS is to bring together re-searchers, engineers and practitioners inter-ested in the advances and business appli-cations of information systems. Six simul-taneous tracks will be held, covering differ-ent aspects of Enterprise Information SystemsApplications, including Enterprise DatabaseTechnology, Systems Integration, Artificial In-telligence, Decision Support Systems, Infor-mation Systems Analysis and Specification,Internet Computing, Electronic Commerce,Human Factors and Enterprise Architecture.Submission deadline: December 13, 2019

19th International Conference onAutonomous Agents and Multi-AgentSystems (AAMAS 2020)Auckland, New Zealand, May 9-13, 2020https://aamas2020.conference.auckland.ac.nz/AAMAS is the leading scientific conference

for research in autonomous agents andmulti-agent systems. The AAMAS conferenceseries was initiated in 2002 as the mergingof three respected scientific meetings: theInternational Conference on Multi-AgentSystems (ICMAS), the International Work-shop on Agent Theories, Architectures, andLanguages (ATAL), and the InternationalConference on Autonomous Agents (AA). Theaim of the joint conference is to provide asingle, high-profile, internationally-respectedarchival forum for scientific research in thetheory and practice of autonomous agentsand multi-agent systems. AAMAS 2020 isthe 19th edition of the AAMAS conference,and the first time AAMAS will be held in NewZealand. The conference solicits papersaddressing original research on autonomousagents and their interaction, including agentsthat interact with humans. In addition to themain track, there will be two special tracks:Blue Sky Ideas and JAAMAS.Submission deadline: November 15, 2020

33rd International Conference onIndustrial, Engineering and OtherApplications (IEA/AIE ’20)Kitakyushu, Japan, July 21-24, 2020https://jsasaki3.wixsite.com/ieaaie2020IEA/AIE 2020 continues the tradition of em-phasizing on applications of applied intelligentsystems to solve real-life problems in all ar-eas including engineering, science, industry,automation & robotics, business & finance,medicine and biomedicine, bioinformatics, cy-berspace, and human-machine interactions.Submission deadline: December 15, 2019

35th IEEE/ACM InternationalConference on Automated SoftwareEngineering (ASE 2020)Melbourne, Australia, September 21-25, 2020https://www.deakin.edu.au/ase2020The 35th IEEE/ACM International Conferenceon Automated Software Engineering (ASE2020) will be held in Melbourne, Australia fromSeptember 21 to 25, 2020. The conferenceis the premier research forum for automatedsoftware engineering. Each year, it bringstogether researchers and practitioners fromacademia and industry to discuss founda-tions, techniques, and tools for automating

16


the analysis, design, implementation, testing,and maintenance of large software systems.

Michael Rovatsos isthe Conference Coordi-nation Officer for ACMSIGAI, and a facultymember at the Univer-sity of Edinburgh. Hisresearch in multiagentsystems and human-friendly AI. Contact him [email protected].

17


AI Education Matters: Building a Fake News DetectorMichael Guerzhoy (Princeton University, University of Toronto, and the Li Ka Shing Knowl-edge Institute, St. Michael’s Hospital; [email protected])Lisa Zhang (University of Toronto Mississauga; [email protected])Georgy Noarov (Princeton University; [email protected])DOI: 10.1145/3362077.3362082

Introduction

Fake news is a salient societal issue, the sub-ject of much recent academic research, and,as of 2019, a ubiquitous catchphrase.

In this article, we explore using the task ofdetecting fake news to teach supervised ma-chine learning and data science, as demon-strated in our Model AI Assignment1(Neller etal., 2019). We ask students to build a seriesof increasingly complex classifiers that cate-gorize news headlines into “fake” and “real”and to analyze the classifiers they have built.Students think about the data, the validity ofthe problem posed to them, and the assump-tions behind the models they use. Studentscan compete in a class-wide competition tobuild the best fake news detector.

To help instructors incorporate fake news de-tection into their course, we briefly review re-cent research on fake news and the task offake news detection. We then describe theassignment design, and reflect on the in-classfake news detection competition we ran.

Fake News Research

Fake news is an old issue (Mansky, 2018),but the role it may have played in the 2016US Presidential Election has sparked renewedinterest in the phenomenon (Lazer et al.,2018), (Allcott & Gentzkow, 2017). Re-search on fake news is focused on under-standing its audience and societal impact,how it spreads on social media, and who itsconsumers are (Grinberg, Joseph, Friedland,Swire-Thompson, & Lazer, 2019), (Nelson &Taneja, 2018).

Fake news can be detected based on tex-tual features and social network propagation

Copyright c© 2019 by the author(s).1http://modelai.gettysburg.edu/

2019/fakenews/

Figure 1: Visualizing P (fake|keyword) for a naiveBayes model trained on our training set. Largertext corresponds to larger conditional probabilities.

patterns (Shu, Sliva, Wang, Tang, & Liu,2017). High-quality datasets of fake andreal news are scarce. Several medium-scaledatasets have recently been collected, withfake news either obtained from the web (of-ten with the help of fact-checking resourcessuch as PolitiFact.com) or written to orderby Amazon Mechanical Turk workers (Wang,2017), (Pérez-Rosas, Kleinberg, Lefevre, &Mihalcea, 2018).

The definition of the concept of “fake news”itself has proven elusive. See (Tandoc, Lim, &Ling, 2018) for an overview of the definitionsrecently used in literature.

Teaching Supervised Learning viaFake News

In our assignment, the task is to classifynews headlines as “real” or “fake”. Studentsbuild and compare several standard classi-fiers: naive Bayes, logistic regression, anda decision tree. All three classifiers use the

18


presence/absence of keywords as their fea-ture set. The detection of fake news headlinesusing naive Bayes is directly analogous to theclassic spam filtering task. See (Russell &Norvig, 2009) for an exposition and (Sahami,Dumais, Heckerman, & Horvitz, 1998) for thepaper that introduced the idea.

Our pedagogical approach emphasizes hav-ing students analyze the models they build. Inparticular, we ask students to obtain keywordswhose presence or absence most strongly in-dicates that a headline is “fake” or “real”.

To find the most important keywords for clas-sifying a headline as “real” using naive Bayes,students need to decide whether they shoulduse P (real|keyword) or P (keyword|real). Wehope they gain a deeper understanding ofnaive Bayes in the process. We use PyTorchto implement logistic regression, and suggest(as would be natural for our students) that stu-dents use multinomial logistic regression with2 outputs when predicting “fake”/“real”. Thisresults in 2k + 2 coefficients for a vocabularyof k keywords. Identifying the most impor-tant keywords based on these 2k+2 numbersnudges students towards understanding thedetails of the model. We also ask students toderive the logistic regression coefficients thatcorrespond to the naive Bayes classifier theyfit. As a final step, students fit a decision treeto the data and again identify the most impor-tant features according to the model.

As they fit a series of increasingly complexclassifiers, students observe overfitting first-hand: training performance increases withclassifier complexity, while validation perfor-mance decreases. Beating naive Bayes turnsout to be quite difficult (though doable). Stu-dents attempt to do that in the competitionphase.

Teaching Data Science via Fake News

When using the assignment in a data sci-ence rather than a machine learning course,we place more emphasis on statistical model-ing and careful examination of the data. Weask students to inspect the dataset in orderto analyze it qualitatively and discuss its lim-itations. Students are also asked to checkwhether the dataset conforms to the naiveBayes assumption (it does not; to figure out

why, students need to think about how humanlanguage works).

Another part of the assignment involves pro-ducing new data via the naive Bayes genera-tive model. The goal is for students to gain adeeper understanding of generative models.

Datasets

The dataset students use in the principal partof the assignment was compiled by combiningdata from several sources. It consists of 1298“fake news” headlines and 1968 “real news”headlines, all containing the word “Trump”.“Fake” headlines are challenging to collect; asstudents see, most headlines labeled as suchcould be argued to be merely tendentious orhyperbolic rather than fake.

We have curated a smaller private test set ofheadlines that we have verified to be eitherreal or fake2. That test set is used in our fakenews detection competition and is available toinstructors upon request.

Fake News Detection Contest

For interested students, we ran an optionalfake news detection competition. The authorsof the best-performing entries would earn asmall amount of points towards their coursegrade. Participating students could follow anyapproach they liked. We encouraged aug-menting the given training set with more data,engineering useful features, training classi-fiers of the students’ own choice, and usingensemble methods. Gratifyingly, some con-testants were able to engineer useful featuresand use modern text classification algorithmsto beat the naive Bayes baseline.

The source code for many modern text classi-fication systems is widely available and some-times comes with pre-trained weights. Stu-dents would often adapt, train, or fine-tune thesystems for their submissions.3

2While we could not fact-check the headlines tojournalistic standards, we made sure that the truthor falseness of the headlines was not in seriousdispute.

3Training deep learning systems is often re-source intensive. We refer students to servicessuch as AWS, Microsoft Azure, and Google CloudPlatform, where they are eligible for free credits.

19


Conclusion

Through building a fake news detector inclass, we are able to teach some of the foun-dational methods of supervised learning in acompelling and coherent manner. The datasetwe collected can be used in a class that em-phasizes rigorous thinking about data scienceproblems. We share our experience of run-ning an in-class fake news detection competi-tion.

References

Allcott, H., & Gentzkow, M. (2017). Socialmedia and fake news in the 2016 elec-tion. Journal of Economic Perspectives,31(2), 211–36.

Grinberg, N., Joseph, K., Friedland, L., Swire-Thompson, B., & Lazer, D. (2019). Fakenews on twitter during the 2016 us pres-idential election. Science, 363(6425),374–378.

Lazer, D. M., Baum, M. A., Benkler, Y., Berin-sky, A. J., Greenhill, K. M., Menczer, F.,. . . others (2018). The science of fakenews. Science, 359(6380), 1094–1096.

Mansky, J. (2018). The age-old problem of“fake news”. Smithsonian Magazine.https://www.smithsonianmag.com/history/age-old-problem-fake-news-180968945/.

Neller, T. W., Sooriamurthi, R., Guerzhoy, M.,Zhang, L., Talaga, P., Archibald, C., . . .others (2019). Model AI Assignments2019. In Proceedings of the AAAI Con-ference on Artificial Intelligence (Vol. 33,pp. 9751–9753).

Nelson, J. L., & Taneja, H. (2018). The small,disloyal fake news audience: The role ofaudience availability in fake news con-sumption. New Media & Society , 20(10),3720-3737.

Pérez-Rosas, V., Kleinberg, B., Lefevre, A., &Mihalcea, R. (2018). Automatic detec-tion of fake news. In Proceedings of the27th International Conference on Com-putational Linguistics (pp. 3391–3401).

Russell, S. J., & Norvig, P. (2009). Artificial In-telligence: A Modern Approach. PearsonEducation Limited.

Sahami, M., Dumais, S., Heckerman, D., &Horvitz, E. (1998). A bayesian approachto filtering junk e-mail. In Learning for

Text Categorization: Papers from the1998 workshop (Vol. 62, pp. 98–105).

Shu, K., Sliva, A., Wang, S., Tang, J., & Liu,H. (2017). Fake news detection on so-cial media: A data mining perspective.SIGKDD Explorations Newsletter , 19(1),22–36.

Tandoc, E. C., Lim, Z. W., & Ling, R. (2018).Defining “fake news”. Digital Journalism,6(2), 137-153.

Wang, W. Y. (2017). “Liar, liar pants on fire”:A new benchmark dataset for fake newsdetection. In Proceedings of the 55th An-nual Meeting of the Association for Com-putational Linguistics.

Michael Guerzhoy is aLecturer at Princeton Uni-versity, an Assistant Pro-fessor (Status Only) atthe University of Toronto,and a Scientist at the LiKa Shing Knowledge In-stitute, St. Michael’s Hos-pital. His professional in-terests are in computer

science and data science education and in ap-plications of machine learning to healthcare.

Lisa Zhang is an Assis-tant Professor, TeachingStream (CLTA) at theUniversity of TorontoMississauga. Her currentresearch interests are inthe intersection of com-puter science educationand machine learning.

Georgy Noarov is a stu-dent at Princeton Uni-versity concentrating inmathematics and pursu-ing a certificate in statis-tics and machine learn-ing. His research areasinclude algorithmic gametheory and combinatorialoptimization.

20


AI Education Matters: A First Introduction to Modeling and Learn-ing using the Data Science WorkflowMarion Neumann (Washington University in St. Louis; [email protected])DOI: 10.1145/3362077.3362083

Introduction

Traditionally artificial intelligence (AI) and ma-chine learning (ML) courses are taught at thesenior and graduate level in higher-educationcomputer science curricula following the mas-tery learning strategy, cf. Figure 1. This makessense, since most AI and ML models andthe theory behind them require a substan-tial understanding of probability and statis-tics, as well as advanced calculus and ma-trix algebra. To understand Logistic Regres-sion as a probabilistic classifier performingmaximum-likelihood or maximum-a-posterioriestimation, for example, students need to un-derstand joint and conditional probability dis-tributions. In order to derive the back propa-gation algorithm to train Neural Networks stu-dents need to understand partial derivativesand inner and outer tensor products. Theseare just two of many examples where sub-stantial mathematical background – typicallytaught at the junior level in a computer sciencemajor program – is required. With AI and MLalgorithms being used more widely by enter-prises across domains, as well as, in applica-tions and services we use in our daily lives, itmakes sense to raise awareness about whatAI is, what it can and cannot do, and how itis used to solve problems to a broader audi-ence. Very much in the same spirit as the“CS for all” idea (https://www.csforall.org), we have to extend our curricula to in-clude introductory courses to AI and ML onthe early undergraduate level (or even in high-school) to expose students to the ideas andworking principles of AI technology. One wayto achieve this is to introduce the principlesof working with data, modeling, and learningthrough the data science workflow.

Exposure First

Following the exposure – interest – mas-tery paradigm as illustrated in Figure 2,


Figure 1: Mastery learning paradigm.

we propose to gently introduce AI/ML con-cepts focusing on example applications ratherthan computational problems by incorporatingcourse modules into introductory CS coursesor design an entire course early on the cur-riculum. The goal of such intro-level mod-ules or courses is to expose students to AI/MLproblems and introduce basic techniques tosolve them without relying on the computa-tional and mathematical prerequisite knowl-edge. More concretely, the module or coursemay be designed as combined lecture andlab sessions, where a new topic is introducedin a lecture unit followed by a lab session,where students get to know a problem in thecontext of an application, explore a solutionmethod, and tackle a potentially open-endedquestion about evaluation procedures, bene-fits and challenges of the approach, or impli-cations and ethical considerations when us-ing such methods in the real world in a groupdiscussion. Lab sessions should be designedcarefully focusing on the understanding of thedata, the problem, and the results instead ofmodel implementation. We will introduce twosuch lab assignments implemented in Python

Figure 2: Exposure-Interest-Mastery paradigm.

21


Figure 3: Data science workflow using sentiment analysis as an example application.

and Jupyter notebooks in the next section.

The main aim of our assignments is to en-gage the students’ interest to acquire the pre-requisite knowledge in order to move forwardand gain a deeper understating of specific AIand ML techniques. Since we propose theseunits for a course that is taught very earlyin the CS curriculum, we face the challengethat students do not have a lot of program-ming experience nor a deep understanding ofdata structures and algorithms. Therefore, wedeveloped the lab assignments using Jupyternotebooks which nicely combine illustrative in-structions and executable starter code.

After having worked though the data scienceworkflow using illustrative applications that areboth easy to understand and relevant in thereal-world, our hope is that students developthe motivation to study traditional prerequisiteclasses for AI and ML courses like probabilityand statistics, matrix/linear algebra, and algo-rithm analysis perceiving them useful to mas-ter AI/ML instead of a nuisance.

Two Model AI Assignments

Introduction to Python for Data Science

We provide an interactive guided lab tointroduce Python for data science (DS),1which can also be used for any coursethat introduces modeling and learning us-ing Python, such as introduction to AI orML courses. We provide two Jupyter note-books, one introducing the basics of Pythonand the other the DS workflow using the

1http://modelai.gettysburg.edu/2019/intro2py/

Iris dataset (https://archive.ics.uci.edu/ml/datasets/Iris). We interac-tively introduce the use of expressions, vari-ables, strings, printing, lists, dictionaries, con-trol flow, and functions in Python to studentsthat are already familiar with a programminglanguage from an introductory CS course.The second lab aims at motivating students toacquire skills such as using statistics to modeland analyze data, knowing how to design anduse algorithms to store, process, and visual-ize data, while not forgetting the importanceof domain expertise. We begin by establish-ing the example problem to be studied basedon the Iris dataset. The next step is to acquireand process the data, where students practicehow to load data and process strings into nu-meric arrays using numpy. Then, we explaindifferent plotting methods such as box plots,histograms, and scatter plots for data explo-ration leveraging matplotlib. Finally, wesplit the data into training and test set, builda model, use it for predictions, and evaluatethe results using sklearn. The main learn-ing objectives are to get to know and practicePython in the context of a realistic data sci-ence and machine learning application.

Introducing the Data Science Workflowusing Sentiment Analysis

The second interactive lab guides studentsthrough a basic data science workflow by ex-ploring sentiment analysis.2 The data scienceworkflow along with the example sentimentanalysis application is depicted in Figure 3.The lab assignment focuses on introducing

2http://modelai.gettysburg.edu/2019/intro2ds/

22


the machinery using a given dataset of moviereviews. We further provide a follow-up home-work assignment reiterating some of the stepsand highlighting data acquisition and explo-ration with Twitter data. After introducing senti-ment analysis, we explain a simple rule-basedapproach to predict the sentiment of textual re-views using three handcrafted examples. Thisintroduction shows simple means to prepro-cess text data and exemplifies the use of listsof positive and negative expressions to com-pute a sentiment score. Then students willimplement the approach to predict the senti-ment of movie reviews and evaluate the re-sults. The lab concludes with a discussionof the limitations of the rule-based approachand a quick introduction to sentiment clas-sification via machine learning. The home-work assignment reiterates over the processof building and analyzing a sentiment predictorwith the focus on collecting and preprocessingtheir own dataset scraped from Twitter usingthe python-twitter API. The main learn-ing objective of this activity is getting to knowthe inference problem and walking through theentire data science workflow to tackle it. Sincethe module only requires minimal program-ming background it is an ideal precursor to in-troducing machine learning in an AI, ML, orDS course. It may also be used in a introduc-tion to Python course as a module focusing onusing libraries and APIs.

Our Experiences

We incorporated both lab assignments intoour “Introduction to Data Science” course forsophomore students at Washington Universityin St. Louis. One of the challenges we facedwas that our students had different levels ofPython experience, from no experience at all(51%) over some experience (36%) to quiteproficient (13%). This led to a large variancein the times needed to complete the labs. Todeal with this issue we propose to add someoptional challenge problems to the assign-ment that are not required for the homework orwill be introduced later in the course. Anotherchallenge was that some students preferredto work in groups where others did the labson their own. However, both strategies canresult in slower or faster pace given the stu-dents’ working style, group composition, andamount of group discussion. Unfortunately,

there is no unified way to tackle this issue,however, we believe that students should beencouraged to work in teams for the lab as-signments, whereas homework assignmentsshould be worked on individually. This wayboth teamwork and communication skills aswell as knowledge retention are facilitated.

Both labs were perceived as useful by our stu-dents. 97% answered Yes to the question“Did you like the lab.” for the introduction toPython lab and 81% for the sentiment analysislab. The most common reasons stated by stu-dents that didn’t like the second lab were thatthey where overwhelmed by unfamiliar codeand that it was too long. From the students’answers to our quiz and exam questions wecan also confirm that they understand basicPython processes to handle data, implementand apply simple learning models, and visual-ize and interpret their results.

Pedagogical Resources

In addition to Jupyter notebooks constitutingthe lab and homework assignments, we de-veloped lecture materials in form of slidesand worksheets for each module. The firstlecture covers an introduction to data sci-ence and machine learning, and the secondone introduces sentiment analysis, text pro-cessing, and classification respectively. Theslides are interactive with gaps to be filled inby the instructor during the lectures and theworksheets contain in-class activities for stu-dents to engage with the presented materi-als. Those resources are available from theauthors upon request.

Useful textbooks that specifically focus on in-troducing data science topics and techniquesare:

• Python Data Science Handbook Vander-Plas (2016) introduces essential tools andlibraries such as Jupyter notebooks, numpy,pandas, scikit-learn, and matplotlib for work-ing with data.

• Data Science from Scratch Grus (2019) fo-cuses on implementing learning algorithmsand data processing routines from scratch.

• Data Science for Business Provost andFawcett (2013) showcases interesting real-world use cases and emphasizes data-

23


analytic thinking while not being too techni-cal.

The first two books focus on implementationsin Python, whereas the third one details con-cepts and techniques without code examples.

ReferencesGrus, J. (2019). Data science from scratch:

first principles with python. O’Reilly Me-dia.

Provost, F., & Fawcett, T. (2013). Datascience for business: What you needto know about data mining and data-analytic thinking. ” O’Reilly Media, Inc.”.

VanderPlas, J. (2016). Python data sciencehandbook: essential tools for workingwith data. ” O’Reilly Media, Inc.”.

Marion Neumann isa Senior Lecturer atWashington University inSt. Louis and the SIGAIdiversity officer. Sheteaches Machine Learn-ing, Cloud Computing,Analysis of NetworkedData, and Introductionto Data Science. Her

research interests include graph-based ma-chine learning and analyzing networked dataas well as measuring and analyzing studentemotions in large computing courses usingsentiment analysis.

24


AI Policy MattersLarry Medsker (The George Washington University; [email protected])DOI: 10.1145/3362077.3362084

Abstract

AI Policy Matters is a regular column in AIMatters featuring summaries and commen-tary based on postings that appear twicea month in the AI Matters blog (https://sigai.acm.org/aimatters/blog/). Wewelcome everyone to make blog commentsso we can develop a rich knowledge base ofinformation and ideas representing the SIGAImembers.

About Face

Face recognition (FR) research has madegreat progress in recent years and has beenprominent in the news. In public policy, manyare calling for a reversal of the trajectory for FRsystems and products. In the hands of peo-ple of good will, using products designed forsafety and training systems with appropriatedata, FR benefits society and individuals. TheVerge reports the use in China of unique facialmarkings of pandas to identify individual an-imals. FR research includes work to mitigatenegative outcomes, as with the Adobe and UCBerkeley work on Detecting Facial Manipula-tions in Adobe Photoshop for automatic de-tection of facial images that have been manip-ulated by splicing, cloning, and removing ob-jects.

Intentional and unintentional application ofsystems that are not designed and trained forethical use are a threat to society. Screeningfor terrorists could be good, but FR lie andfraud detection systems sometimes do notwork properly. The safety of FR is currentlyan important issue for policymakers, but regu-lations could have negative consequences forAI researchers. As with many contemporaryissues, conflicts arise because of conflictingpolicies in different countries. Recent and cur-rent legislation is attempting to restrict FR useand possibly inhibit FR research; for example,

• San Francisco, CA, Somerville, MA, andCopyright c© 2019 by the author(s).

Oakland, CA, are the first three cities to limituse of FR to identify people.

• In “Facial recognition may be banned frompublic housing thanks to proposed law”CNET reports that a bill will be introduced toaddress the issue that “landlords across thecountry continue to install smart home tech-nology and tenants worry about uncheckedsurveillance.”

• A call for a more comprehensive ban onFR has been launched by the digital rightsgroup Fight for the Future, seeking a com-plete Federal ban on government use of fa-cial recognition surveillance.

Beyond legislation against FR research andbanning certain products, work is in progressto enable safe and ethical use of FR. A moregeneral example that could be applied to FRis the MITRE work The Ethical Frameworkfor the Use of Consumer-Generated Data inHealth Care, which “establishes ethical val-ues, principles, and guidelines.”

AI Regulation

With AI in the news so much over the pastyear, the public awareness of potential prob-lems arising from the proliferation of AI sys-tems and products has led to increasing callsfor regulation. The popular media, and eventechnical media, do contain misinformationand misplaced fears, but plenty of legitimateissues exist even if their relative importance issometimes misunderstood. Policymakers, re-searchers, and developers need to be in dia-log about the true needs for and potential dan-gers of regulation. From our policy perspec-tive, the significant risks from AI systems in-clude misuse and faulty unsafe designs thatcan create bias, non-transparency of use, andloss of privacy. Some AI systems are knownto discriminate against minorities, unintention-ally and not.

An important discussion we should be havingis if governments, international organizations,

25


and big corporations, which have already re-leased dozens of non-binding guidelines forthe responsible development and use of AI,are the best entities for writing and enforc-ing regulations. Non-binding principles will notmake some companies developing and apply-ing AI products accountable. An importantpoint in this regard is to hold companies re-sponsible for the product design process itself,not just for testing products after they are inuse.

Introduction of new government regulations isa long process and subject to pressure fromlobbyists, and the current US administrationis generally inclined against regulations any-way. We should discuss alternatives like clear-inghouses and consumer groups endorsing AIproducts designed for safety and ethical use.If well publicized, the endorsements of re-spected non-partisan groups including profes-sional societies might be more effective andtimely than government regulations.

The European Union has released its EthicsGuidelines for Trustworthy AI, and a seconddocument with recommendations on how toboost investment in Europe’s AI industry isto be published. In May, 2019, the Organi-zation for Economic Cooperation and Devel-opment (OECD) issued their first set of in-ternational OECD Principles on Artificial In-telligence, which are embraced by the UnitedState and leading AI companies.

The AI Race

China, the European Union, and the UnitedStates have been in the news about strate-gic plans and policies on the future of AI.The U.S. National Artificial Intelligence Re-search and Development Strategic Plan, wasreleased in June, 2019, as an update of thereport by the Select Committee on Artificial In-telligence of The National Science and Tech-nology Council. The Computing CommunityConsortium (CCC) recently released the AIRoadmap Website.

Now, the Center for Data Innovation has is-sued a Report comparing the current stand-ings of China, the European Union, and theUnited States. Here is a summary of their pol-icy recommendations: “Many nations are rac-ing to achieve a global innovation advantage in

artificial intelligence (AI) because they under-stand that AI is a foundational technology thatcan boost competitiveness, increase produc-tivity, protect national security, and help solvesocietal challenges. This report comparesChina, the European Union, and the UnitedStates in terms of their relative standing inthe AI economy by examining six categories ofmetrics: talent, research, development, adop-tion, data, and hardware. It finds that de-spite the bold AI initiatives in China, the UnitedStates still leads in absolute terms. Chinacomes in second, and the European Unionlags further behind. This order could changein coming years as China appears to be mak-ing more rapid progress than either the UnitedStates or the European Union. Nonetheless,when controlling for the size of the labor forcein the three regions, the current U.S. lead be-comes even larger, while China drops to thirdplace, behind the European Union. This re-port also offers a range of policy recommen-dations to help each nation or region improveits AI capabilities.”

US and G20 AI Policy

The G20 AI Ministers from the Group of20 major economies conducted meetings ontrade and the digital economy. They pro-duced guiding principles for using artificial in-telligence based on principles adopted earlierby the 36-member OECD and an additional sixcountries. The G20 guidelines call for usersand developers of AI to be fair and account-able, with transparent decision-making pro-cesses and to respect the rule of law and val-ues including privacy, equality, diversity andinternationally recognized labor rights. Mean-while, the principles also urge governmentsto ensure a fair transition for workers throughtraining programs and access to new job op-portunities.

Bipartisan Legislators On Deepfake Videos

Senators introduced legislation intended tolessen the threat posed by “deepfake“ videos,which use AI technologies to manipulate orig-inal videos and produce misleading informa-tion. With this legislation, the Department ofHomeland Security would conduct an annualstudy of deepfakes and related content andrequire the department to assess the AI tech-nologies used to create deepfakes. This could

26


lead to changes in regulations or to new regu-lations impacting the use of AI.

Hearing on Societal and Ethical Impacts

The House Science, Space and TechnologyCommittee held a hearing on June 26th aboutthe societal and ethical implications of artificialintelligence, now available on video. The Na-tional Artificial Intelligence Research and De-velopment Strategic Plan, released in June, isan update of the report by the Select Com-mittee on Artificial Intelligence of The NationalScience and Technology Council.

On February 11, 2019, the President signedExecutive Order 13859: Maintaining Ameri-can Leadership in Artificial Intelligence. Ac-cording to Michael Kratsios, Deputy Assis-tant to the President for Technology Policy,this order “launched the American AI Initia-tive, which is a concerted effort to promoteand protect AI technology and innovation inthe United States. The Initiative implementsa whole-of-government strategy in collabora-tion and engagement with the private sector,academia, the public, and like-minded inter-national partners. Among other actions, keydirectives in the Initiative call for Federal agen-cies to prioritize AI research and developmentinvestments, enhance access to high-qualitycyberinfrastructure and data, ensure that theNation leads in the development of technicalstandards for AI, and provide education andtraining opportunities to prepare the Americanworkforce for the new era of AI.”

The first seven strategies continue from the2016 Plan, reflecting the reaffirmation of theimportance of these strategies by multiple re-spondents from the public and government,with no calls to remove any of the strategies.The eighth strategy is new and focuses onthe increasing importance of effective partner-ships between the Federal Government andacademia, industry, other non-Federal enti-ties, and international allies to generate tech-nological breakthroughs in AI and to rapidlytransition those breakthroughs into capabili-ties.

Strategy 8: Expand Public–Private Partner-ships to Accelerate Advances in AI is new inthe June, 2019, plan and reflects the grow-ing importance of public-private partnerships

enabling AI research and expanding public-private partnerships to accelerate advances inAI. A goal is to promote opportunities for sus-tained investment in AI research and develop-ment and transitions into practical capabilities,in collaboration with academia, industry, inter-national partners, and other non-Federal enti-ties.

Continued points from the seven Strategies inthe previous Executive Order in February in-clude

• support for the development of instructionalmaterials and teacher professional develop-ment in computer science at all levels, withemphasis at the K–12 levels,

• consideration of AI as a priority area withinexisting Federal fellowship and service pro-grams,

• development of AI techniques for humanaugmentation,

• emphasis on achieving trust: AI system de-signers need to create accurate, reliablesystems with informative, user-friendly inter-faces.

The National Science and Technology Coun-cil (NSTC) is functioning again. NSTC isthe principal means by which the ExecutiveBranch coordinates science and technologypolicy across the diverse entities that make upthe Federal research and development enter-prise. A primary objective of the NSTC is toensure that science and technology policy de-cisions and programs are consistent with thePresident’s stated goals. The NSTC preparesresearch and development strategies that arecoordinated across Federal agencies aimedat accomplishing multiple national goals. Thework of the NSTC is organized under commit-tees that oversee subcommittees and workinggroups focused on different aspects of scienceand technology. More information is available.

The Office of Science and Technology Pol-icy (OSTP) was established by the NationalScience and Technology Policy, Organization,and Priorities Act of 1976 to provide the Pres-ident and others within the Executive Officeof the President with advice on the scientific,engineering, and technological aspects of theeconomy, national security, homeland secu-rity, health, foreign relations, the environment,

27


and the technological recovery and use of re-sources, among other topics. OSTP leads in-teragency science and technology policy co-ordination efforts, assists the Office of Man-agement and Budget with an annual reviewand analysis of Federal research and devel-opment in budgets, and serves as a source ofscientific and technological analysis and judg-ment for the President with respect to majorpolicies, plans, and programs of the FederalGovernment. More information is available.

Groups that advise and assist the NSTC on AIinclude

• The Select Committee on Artificial Intelli-gence addresses Federal AI research anddevelopment activities, including those re-lated to autonomous systems, biometricidentification, computer vision, human com-puter interactions, machine learning, natu-ral language processing, and robotics. Thecommittee supports policy on technical, na-tional AI workforce issues

• The Subcommittee on Machine Learningand Artificial Intelligence monitors the stateof the art in machine learning (ML) and arti-ficial intelligence within the Federal Govern-ment, in the private sector, and internation-ally

• The Artificial Intelligence Research and De-velopment Interagency Working Group co-ordinates Federal research and develop-ment in AI and supports and coordinates ac-tivities tasked by the Select Committee onAI and the NSTC Subcommittee on MachineLearning and Artificial Intelligence.

More information available.

Please join our discussions at the SIGAI Pol-icy Blog.

Larry Medsker is Re-search Professor ofPhysics and was foundingdirector of the Data Sci-ence graduate programat The George Wash-ington University. He isa faculty member in theGW Human-TechnologyCollaboration Lab and

Ph.D. program. His research in AI includeswork on artificial neural networks, hybridintelligent systems, and the impacts of AI onsociety and policy. He is the Public PolicyOfficer for the ACM SIGAI.

28


Advancing Non-Convex and Constrained Learning: Challengesand OpportunitiesTianbao Yang (The University of Iowa; [email protected])DOI: 10.1145/3362077.3362085

Introduction

As data gets more complex and applica-tions of machine learning (ML) algorithms fordecision-making broaden and diversify, tra-ditional ML methods by minimizing an un-constrained or simply constrained convex ob-jective are becoming increasingly unsatisfac-tory. To address this new challenge, recentML research has sparked a paradigm shiftin learning predictive models into non-convexlearning and heavily constrained learning.Non-Convex Learning (NCL) refers to a fam-ily of learning methods that involve optimiz-ing non-convex objectives. Heavily Con-strained Learning (HCL) refers to a family oflearning methods that involve constraints thatare much more complicated than a simplenorm constraint (e.g., data-dependent func-tional constraints, non-convex constraints), asin conventional learning. This paradigm shifthas already created many promising out-comes: (i) non-convex deep learning hasbrought breakthroughs for learning represen-tations from large-scale structured data (e.g.,images, speech) (LeCun, Bengio, & Hinton,2015; Krizhevsky, Sutskever, & Hinton, 2012;Amodei et al., 2016; Deng & Liu, 2018);(ii) non-convex regularizers (e.g., for enforc-ing sparsity or low-rank) could be more effec-tive than their convex counterparts for learn-ing high-dimensional structured models (C.-H. Zhang & Zhang, 2012; J. Fan & Li, 2001;C.-H. Zhang, 2010; T. Zhang, 2010); (iii)constrained learning is being used to learnpredictive models that satisfy various con-straints to respect social norms (e.g., fair-ness) (B. E. Woodworth, Gunasekar, Ohan-nessian, & Srebro, 2017; Hardt, Price, Srebro,et al., 2016; Zafar, Valera, Gomez Rodriguez,& Gummadi, 2017; A. Agarwal, Beygelzimer,Dudı́k, Langford, & Wallach, 2018), to improvethe interpretability (Gupta et al., 2016; Canini,Cotter, Gupta, Fard, & Pfeifer, 2016; You,Ding, Canini, Pfeifer, & Gupta, 2017), to en-hance the robustness (Globerson & Roweis,Copyright c© 2019 by the author(s).

2006a; Sra, Nowozin, & Wright, 2011; T. Yang,Mahdavi, Jin, Zhang, & Zhou, 2012), etc. Inspite of great promises brought by these newlearning paradigms, they also bring emergingchallenges to the design of computationally ef-ficient algorithms for big data and the analysisof their statistical properties.

Non-Convex Learning

In this section, we describe some recent ad-vances in non-convex learning with mention-ing some of our recent related results. Wewill also describe their limitations and pointout future directions. This article will focus onstudies that are concerned with algorithm de-sign and analysis for solving NCL and HCLproblems instead of papers that are purelyapplication-driven. It is notable that the ref-erences are not exhaustive due to a large vol-ume of related works.

Non-Convex Minimization and Deep Learn-ing. Deep learning can be formulated as thefollowing non-convex minimization problem:

minw∈Rd F (w) := Ez[f(w; z)], (1)

where z denotes a random data, and w de-notes the parameters of the neural networkto be learned, and f(w; z) denotes the lossfunction. Due to the success of deep learn-ing in many areas, this problem has attractedmuch attention from the community of math-ematical programming and machine learning.Research has been conducted in the followingdirections.

• Convergence to stationary points. Forgeneral non-convex problems, it is NP-hardto find a global minimizer (Hillar & Lim,2013). Hence, many studies have focusedon finding stationary points of (1) (Nesterov& Polyak, 2006; N. Agarwal, Allen Zhu,Bullins, Hazan, & Ma, 2017; Carmon, Duchi,Hinder, & Sidford, 2016; P. Xu, Roosta-Khorasani, & Mahoney, 2017; Cartis, Gould,& Toint, 2011b, 2011a; Royer & Wright,

29


2017; M. Liu & Yang, 2017b, 2017a; Allen-Zhu, 2017; Kohler & Lucchi, 2017; Reddi etal., 2017). Typically, two types of station-ary points are considered, namely first-orderstationary point and second-order station-ary point. A point w∗ is called a first-orderstationary point if it satisfies ∇F (w∗) = 0. Apoint w∗ is called a second-order stationaryif it satisfies ∇F (w∗) = 0 and ∇2F (w∗) � 0.These studies concentrate on the complex-ity analysis of first or second-order methods.Many first-order meth

ai matters, volume 5, issue 3 5(3) 2019 ai...

Documents