Vol 1 Meet AI - GitHub Pages

Vol 1Second Edition AITwitter: @MachineLearnist Instagram: @themachinelearnist Falaah Arif Khan, et A.I. This is a very strange time in its life.Late last year, ‘The Portrait of Edmond Belamy’, created by the Paris-basedartistic group Oblivious and generated entirely by a deep learning algorithm, soldfor 432,500 at an art auction.The media took the eventto mark the emergence ofArtificial Intelligence asa new Artistic MediumThe inevitable aiapocalypse fearmongeringfollowed:“Artists hadbetter beware:The machines areon their way, andthey’re comingfor your jobs”Internal Use - ConfidentialFacebook: l 2020,By Falaah Arif KhanWhy only artists?If Machines nowpossess thecreativity togenerate originalart, they surelymust havemastered and willsoon replacehumans in bluecollar jobs too,Right?Not quite. Lets dig a little deeper (pun-tastic!)

The GAN, The Myth, The LegendCirca 2014I’m the Generator(G). I learn thedata distributionand use it togeneratesynthetic imagesThe model that was used to generate thatpainting (and all the fabricated celebrity gossipand fake news videos that circulate on socialmedia) is known as a Generative AdversarialNetwork (or GAN). GANs are adept atgenerating synthetic samples by learning theunderlying representation of thedataGoodfellow et al unleashed this two headeddemon on the world in their 2014 paper and thealgorithm has since gained widespread researchinterest and universal scientific acclaim“The GAN and its variations thatare now being proposed is themost interesting idea in the lastdecade in Machine Learning!”-Yann LeCun2018 ACM Turing award co-recipientI’m the Discriminator(D). I predict if theimage was sampledfrom the training setor was generatedsynthetically by GThe training algorithm isa minimax two-playergame that convergeswhen G is able to fool Dby generating samplesthat look real.Fact!Although they weren’t called GANsand weren’t used to generatesynthetic images, the origins ofalgorithms that use adversarialgames between systems togenerate novel outputs dates backto over 2 decades ago!Circa 1990Jürgen Schmidhuber Proposedthe notion of “ArtificialCuriosity” in his seminal 1990paper and later demonstratedits application in reinforcementlearning settings where agentsemploy “curious exploration”to minimize the errors in theirmodel of the worldConfidential InternalFalaahArif Khan, https://falaaharifkhan.github.io/researchUse - ConfidentialFor all the cookswith their fingers inthe GAN pie overthe years, there arestill major problemssuch as ModelInstability and ModeCollapse yet to betackled. Algorithmicshortcomings arefurther exacerbatedby the use ofunsuitable data andill posed problems

Take this GAN that gave surprising and unpredictableresults when used for generating cat picturesAnd so the model taught itselfthat letters must be a part ofwhat constitutes a cat The Data that the model was trainedon contained images of cat memesfrom the internet Okay, so thealgorithms havelimitations. Surely,the scientificcommunity is workingtowards fixing them. And it began generating images ofcats embedded with characters ofsome strange fictitious language!Well, remember Tay; Microsoft’s “AIwith zero chill”, that was releasedinto the twitter-verse for all of16 hours beforeit had to beshut down?In the meantime, suchbloopers are harmlessand make for somegood ol’ innocoushumour, right?Tay was designed to learn by interacting with humans, but it was unable to distinguish goodbehavior to emulate from the deliberate trolling that is commonplace on the internet. Thisquickly turned the chatty and friendly AI into a bigoted and fascist PR nightmare.Or Youth Laboratories’beauty pageant judgedpurely by AI ThatProved to beoverwhelminglyracist in itsselection ofwinners andConsistentlydiscriminatedAgainstPeople ofcolor.Use - Confidential InternalFalaahArif Khan, https://falaaharifkhan.github.io/researchDecades of building better hardware,experimenting with deep modelarchitectures and working on imagerecognition and this is what itculminates to?Why domodels‘misbehave’?

Falaah Arif Khan, https://falaaharifkhan.github.io/researchMath, computer science and everything nice,these were the ingredients chosen,To learn patterns in a corpus of data But, we forgot to account for an intrinsic ingredientin the mix: SOCIAL BIASAnd so, we ended up teaching algorithms the prejudices that underpinour society and then deploying them back into that very settingwe inadvertently inflated these biases by assuming that algorithms are inherentlyneutral and can be relied upon to define policies and actionsInternal Use - Confidential

There exists a group of brave explorers who are voyaging across uncharteredscientific territories, where computer science meets the humanities and social sciences So,how arewe solvingthisproblem? to understandhow to ntand -henceand have sofar onlyyieldedimportantinsights onwhat we’vebeen doingwrong.Projects suchas “Gendershades”(buolamwini,gebru) broughtto light theracial andgender-baseddiscriminationby computervision modelsand otherhastily adoptedautomateddecision-makingsoftwareThe work isyet to moveout of theacademicsphere andto yieldactionableitems.On the other end, there are those whowill do whatever it takes to appear to besolving these problemsStrategy 2: Math WashingWHAT: camouflaging bias beneath the facade of‘neutral’ mathHow it’s done: shroud a product in excessive‘mathi-ness’ to the point that the general publicstops questioning the motives of its decisions.When things go wrong, blame the algorithm andabsolve yourself of any culpabilityStrategy 1: Bureaucracyfor: a given product X and anappointed group of people Y,If:Y can speak no evil about XY can see no evil in XY can hear no evil from XThen: Y can affirm that there existsno evil in X“Touch ID doesn't store anyimages of your fingerprint,and instead relies only on amathematicalrepresentation.”Apple, 2017What ends up happening is bigoted, biased algorithms being blindlystamped with an ethics certificate and pushed to production wherethey have real effects on social decision making and cause the samehavoc all over again Use -ArifConfidential InternalFalaahKhan, https://falaaharifkhan.github.io/research

Well, Because the media coverage is not an accurate depiction of ground reality. It’s more like aprojection of science fiction scenarios onto genuine scientific advancements. resembling eager parents,blinded by love (for funding from big tech) for their child, media outlets create a huge ruckus andherald each incremental output from major research labs as paradigm-changing and revolutionaryReLU stands for RectifiedLinear Unit. It’s anactivation function widelyused in Deep Learning!Lulu just said herfirst word!!! Shesaid ‘relu’ !!!OMG! She’s a genius!She’s going to grow upto solve ArtificialGeneral Intelligence!Remember theSTORY aboutschatbots thatinvented theirown secretlanguage? Media Coverage Public ExpectationFalaah Arif Khan, https://falaaharifkhan.github.io/researchThe viral news articles all spoke of how the chatbots had gone off on their own and invented theirown jargon to negotiate more effectively than was possible in English. The team then was forcedto shut down the program over concerns of what their creations could yield if allowed to continueIn reality, the bots were left to learn from each other, rather from humans and this led to themlearning each other’s mistakes, without knowing that they were mistakes in the first place.Think of it as babies conversing with each other. Yes, theyspeak a language quite unlike English and sure, bothparties seem to understand each other, but is thelanguage being spoken a higher form of conversation ormerely what started out resembling language, steadilydevolving into incoherent noises?Scientific RealityInternal Use - ConfidentialIn its coverage ofadvancements inAI, the mediasaves all itspessimism for theimplications ofthe work;regularly . Itseldom confersthe samecynicism towardsthe merit of thework

As the media rushes to report on scientific advancements, it mistakenly conflatesopinion with fact and pumps a hyperbolic narrative, dripping with conjecture.The unsuspecting public sees these futuristic scenarios as depictive of the currentlandscape of technological advancement and this breeds misinformation. Soon, thegeneral opinion is cemented on the exaggerated capability of current AI systems.With public opinion soaring, the industry finds a suitableopportunity to quote high valuations and leech investormoney. And so, practitioners that were dominating othertechnologies turn into AI experts overnight and theybegin to peddle the same overstated narratives,propagating even more speculationto further heighten public interestWhat emerges is an incentive schemethat favors impactful-looking resultsover rigorous scientific enquiry. Andthis contorts the benchmarks thatnewcomers strive towardsSo, in this battle against aburgeoning snowball of shoddyscholarship, fueled by interestand ignorance, the trueexperts find themselveson the back footThink of ML Scholarship as a game ofJenga The false prophets, with their lackadaisicalmethodology and dearth of scientific rigor go onto publish volumes of work that overstateresults, contort excessive mathematical languagefor technical rigor, and anthropomorphize simpletasks as the encoding of human abilities.they lay a rather unstable foundation for thefuture of ML Scholarship, upon which generationsof researchers inevitably begin to base their workAs novicesenter the field,they begin byturningtowards thework that isalreadypublished,because surelyit must bewell qualifiedand reviewed.They have noreason tochallengeexisting claimsand end uplearning thesame practicesof hyperboleandspeculationand eventuallybecomeincentivized tofollow thesameprotocols fortheir shot inthe spotlightThe cycleperpetuates andthose who try tofix the systemfrom the insidefind themselvesdreadfullyoutnumberedFor eachconference thatputs the effortinto meticulouslyreviewing andreproducingsubmissions beforemaking acceptancedecisions,there are tenothers thatoperate on a quidpro quo, set upto mutuallybenefit eachother’s publicationrecordsAnd so we find ourselves in thisprecarious gameThe true experts thatare trying to bringabout systemic changeare extremelyinfluential only withinthe confines of academiaThe ones whose voices do reach thegeneral public are the ones paintingdoomsday scenarios and proclaimingthe existential threat that AI poses.Ironically, these are the samepeople pumping money andeffort into filling our onlineplatforms, our roads and evenour homes with manifestationsof the very same algorithms.how many ROUNDS of thisdo we have left?Internal Use - ConfidentialBut to them, there’s nothing tofix, because, well,its nothing personal,its just business Falaah Arif Khan, https://falaaharifkhan.github.io/research