從深度學(xué)習(xí)到理性機(jī)器：哲學(xué)史對(duì)人工智能未來的啟示

FromDeepLearningtoRationalMachines–FinalManuscriptVersion–CameronBuckner

cjbuckner@

PAGE

10

ThisisadraftofachapterthathasbeenacceptedforpublicationbyOxfordUniversityPressintheforthcomingbook:

FromDeepLearningtoRationalMachines

WhattheHistoryofPhilosophyCanTeachUsabouttheFutureofArtificialIntelligence

byCameronJ.Bucknerdueforpublicationin2023

Chapter1

ModerateEmpiricismandMachineLearning

Insteadoftryingtoproduceaprogrammetosimulatetheadultmind,whynotrathertrytoproduceonewhichsimulatesthechild's?Ifthiswerethensubjectedtoanappropriatecourseofeducationonewouldobtaintheadultbrain.Presumablythechild-brainissomethinglikeanote-bookasonebuysitfromthestationers.Ratherlittlemechanismandlotsofblanksheets.…Ourhopeisthatthereissolittlemechanisminthechild-brainthatsomethinglikeitcanbeeasilyprogrammed.

-AlanM.Turing(1950)

1.1Playingwithfire?Naturevs.nurtureforcomputerscience

Inhumaninquiry,theintroductionofagranddichotomy—goodvs.evil,mortalvs.divine,emotionvs.reason—cantakeonthevitalimportance,aswellastheattendantdanger,ofthediscoveryoffire.Whilesuchdichotomiessupportqualitativeshiftsinthereachofourtheorizing,theyareoftenquicklytakenforgranted,perhapstooquickly,asanelementalforcegoverningtheworldandourplacewithinit.Thedistinctionbetweennatureandnurturestandsasaprimeexample.Thisoppositionhasanimatedthehumanintellectforthousandsofyears,motivatingthesystematicexplorationofcompetingstylesoftheoryinnearlyeveryacademicdiscipline.Wetendtohavestrongintuitionsastowhetherhumanknowledgeisproducedbyturninginwardtounpackourinnatementalendowmentorbyturningoutwardtointerpretthecipherofexperience,andtheenergyprovidedbytheseintuitionshaspoweredavarietyofscientificandtechnologicalinnovations.AswithotherPrometheanbargains,however,suchadvancesareboughtattheexpenseofnewandpersistentdangers.Vigorouslyrubbingtheseopposingintuitionsagainstoneanothercangeneratefrictionwithoutillumination,causingtheoriststopursuearesearchprogramlongafteritsempiricalprospectshavegrowncold,ortolosesightofthedetailsofoneanother’sviewsinahazeofmisunderstandingand

exaggeration.And,ofcourse,firesthatgrowtoolargecanburndangerouslyoutofcontrol.Lestwegetsinged,thesedistinctionsmustbecontinuouslywatchedandcarefullytended—particularlywhenapowerfulnewsourceoffuelistossedintotheflames.

Wearenowinthemiddleofjustsuchaconflagration,andthenewfuelsourcegoesbythenameof“deeplearning.”Indeed,fundingandresearchfordeeplearningiscurrentlyblazing;asof2023,everymajortechcompany’smarqueeR&Dgroupisfocusedondeeplearning,withfiercebiddingwarsfortoptalent.

MostissuesofprestigepublicationslikeScienceandNaturefeatureoneofthesegroups’latestexperiments.Thesepublicationsreportaseriesoftransformativebreakthroughsinartificialintelligence,includingsystemsthatcan:recognizecomplexobjectsinnaturalphotographsaswellorbetterthanhumans;defeathumangrandmastersinstrategygamessuchaschess,Go,shoji,orStarcraftII;createnovelpicturesandbodiesoftextthataresometimesindistinguishablefromthoseproducedbyhumans;siftthroughthefaintestradioechoestodiscovernewexoplanetsorbitingstarsthousandsoflightyearsaway;crunchmassiveamountsofdatageneratedbyparticleacceleratorstotrytofindcounterexamplestotheStandardModelinphysics;andpredicthowproteinswillfoldmoreaccuratelythanhumanmicrobiologistswhohavedevotedtheirlivestothetask.1

Inshort,deeplearning’scurrentfortunesarewhite-hot;but,aswithallsystemsofknowledgeacquisition,ourexpectationsofitscontinuedprosperityareshapedbyourviewsonthenature-nurturedichotomy.Deeplearning’scurrentstatusandfuturedevelopmentarethereforemeaningfullyinformedbyphilosophicalpositions,particularlythoseonofferinthehistoricallygroundedbutongoingdebatebetweenempiricistsandnativists.Atfirstblush,thisdebateconcernstheoriginsofhumanknowledge:empiricistsholdthatknowledgeisderivedfromsensoryexperience,whereasnativiststendtoberationalistswhoinsteadprizeourcapacitytoreason—usuallydrivenbyaninnatetheoryoftheworld’sbasicstructureand/orof

1Fordetails,see(Baldi,Sadowski,andWhiteson2014;Brownetal.2020;Chowdheryetal.2022;Jumperetal.2021;Krizhevsky,Sutskever,andHinton2012;Rameshetal.2022;ShallueandVanderburg2018;Silveretal.2017;Vinyalsetal.2019).

rationalminds—asthesourceofgenuineknowledge.2Whentreatedasanapproachtoartificialintelligence,deeplearningisalreadyidentifiedasanurture-favoring,empiriciststyleoftheory,thoughIarguethatitsachievementsvindicateamoderatelyempiricistapproachtocognitionthatismorenuancedandresourcefulthantheempiricismtypicallysurveyedinevaluationsofdeeplearning’spotential.Thismoderatelyempiricistapproach,legitimatedbyaninvestigationofthehistoricaloriginsofthephilosophicaldebateintheworkofinfluentialempiricistphilosophersandtheapplicationoftheirviewstotherelationshipbetweenmachinelearningmodelsandthemind,suggeststhattoday’sachievementsindeeplearningsubstantiallyincreasetheplausibilitythatrationalcognitioncanbeachieved—andisachieved,inhumans,manyanimals,and,ifwehopetosucceed,artificialagents—withouttheaidoftheinnatetheoriesorconceptsusuallyrecommendedbytheopposed,nature-favoring,rationalistfactionoftheorists.

Whileempiricistandnativiststheoristsfightoverthepast,present,andfutureofdeeplearningsystemsdevelopment,thecurrententhusiasmforempiricisminengineeringandbusinessthreatenstoburnoutofcontrol—thoughthisparticularstrainofempiricismsometimesdrawsoxygenfromasimplisticunderstandingoftherelationshipbetweenthesuccessesofdeeplearningsystemsandthewaythathumansandanimalsactuallysolveproblems.Researchismovingsorapidlythataninfluentialdeeplearningpublicationcanreceive20,000citationsbythetimeitisonlytwoorthreeyearsold—manyofthosewhileitisavailableonlyonapre-printarchive,meaningthatithasnotyetgonethroughthenormalprocessofpeer-reviewbyotheracademicswhocouldskepticallyassessitsclaims.Meanwhile,leadingnativistsaregoinghoarsecallingforthefirebrigade.Thesenativistsworrythatdeeplearningisbeingappliedtoawiderangeofproblemswithoutafirmunderstandingofhoworwhyitworks,andthatthesolutionsdiscoveredbydeeplearningagentsarebrittleanddonotgeneralizetonewsituationsaswellasthestrategiesdeployedbyhumansandanimals.Dependinguponwhetheryouaskempiricistsornativists,deeplearningsystemscaneitheralreadyprocessinputdatasoeffectivelythattheyareatleastslightlyconsciousandonthevergeofachievingescapevelocityintoworld-spanningsuperintelligence,ortheycandolittlemorethanbludgeonproblemswith

2Toforestallconfusion,thephilosophicalrationalismattributabletothinkerslikeDescartes,Leibniz,andSpinozaisnottobeconflatedwiththenew“rationalism”associatedwithblogslikeLessWrongorSlateStarCodex,forwhichthetraditionalphilosophicaldistinctionisorthogonal.

massiveamountsofstatisticsandlinearalgebrathatcanimitatetheoutwardappearanceofhumanintelligencebut,becausetheylacktheunderlyingstructureprovidedbythehumanmind’sinnatestartupsoftware,willnevercaptureeventhemostbasicaspectsofhumanmentality.

Althoughdeeplearningcanbeunderstoodinpurelytechnicaltermsoutsidethenature-nurturedichotomy,andhenceoutsidetheempiricist-nativistdebate,itisdifficulttoassessitsprospectsasaroutetoartificialintelligenceexceptthroughitslight,withallitsattendantprospectsandperils.Thisdebateofcoursehasancienthistoricalorigins,yetinfluentialscientistsfrequentlyinvokeitstermstoexplainandmotivatetheircurrentviews.Forinstance,inafrontNaturearticle,ateamfromGoogle’sDeepMinddivisionpitchedtheirAlphaZerosystem—whichcaneasilydefeathumangrandmastersattheChineseboardgameofGo,agamethatisinsomewaysmorecomplexthanchess—asoperatingwitha“tabularasa”orblankslatealgorithm(Silveretal.2017).ThisempiricistmetaphorenteredtheWesternlexiconviaAristotle’sDeAnima(III,429b-430a),whichcomparesthehumanmindtothewax-coveredtabletswhichtheGreekacademiesusedfornotes;thesetabletswere“blanked”byheatingthemuntilthewaxmelted,smoothingthesurfaceforre-use.Themetaphorfortheinfant’smindbecamecanonicalthroughitsrepetitionbyarangeofempiricistphilosophers,fromAristotle’sinheritorsIbnSina(Avicenna)andSt.ThomasAquinas(thelatterofwhichsummarizeditwiththePeripateticMaxim,whichstatesthat“nihilestinintellectuquodnonsitpriusinsensu”or“nothinginthemindwhichisnotfirstinthesenses”—DeVeritate2.3.19),totheEarlyModernempiricistsJohnLockeandDavidHume,withwhomtheviewistodaymostcommonlyassociated.3

Deeplearningenthusiastsarenottheonlyonestosummonthehistoryofphilosophyinthiscontext.

Contemporarynativistshavealsobeeneagertoalignthecurrentdebatewithhistoricalpositions.InhiscritiqueoftheAlphaZeropaper,forexample,thenativistpsychologistGaryMarcusassociatesSilveretal.’sblankslatelanguagewiththeviewsofLocke,whowrotethat“allideascomefromsensationorreflection”(E

3Otherphilosophicaltraditionsalsohaveviewswhichappearrecognizablyempiricistbythestandardsofthisdebate;forexample,someoftheYogācāraBuddhistphilosopherslikeDharmakīrtiareidentifiedasempiricistbyinterpreters(Powers1994;Tillemans2021)andsomehaveevenwonderedwhetherWesternempiricistslikeHumewereinfluencedbyexposuretoBuddhistphilosophy(Gopnik2009).Othercommentators,however,viewsuchtrans-culturallinkageswithskepticism(Conze1963;Montalvo1999).Atanyrate,averyinterestingbooksimilartothisonecouldbewrittenbydrawinguponthefacultypsychologyinthesealternativetraditionstointerpretandguidethedevelopmentofdeeplearning.IamgratefultoAmitChaturvadifordrawingmyattentiontothesepotentialparallels.

II.ii.2).MarcuscouldjustaswellhavelinkedittoHume,whodeclaredthat“alloursimpleideasintheirfirstappearancearederiv’dfromsimple[sensory]impressions”(commonlyreferredtoashis“CopyPrinciple”–T/4).Hume,however,ismorefrequentlythetargetofJudeaPearl.Oneofthemostinfluentiallivingcomputerscientistsandafrequentdeeplearningcritic,Pearlhasrecentlyworriedthatdeeplearningtheoriststakeasself-evidenta“radicalempiricism”accordingtowhichallknowledge“canbeanalyzedbyexaminingpatternsofconditionalprobabilitiesinthedata”(2021).4

Thehistoryofphilosophycertainlyspeakstodeeplearning’sachievements,butnotintermsassimpleastheseinterlocutorssuggest.Wheretheyseeastarkdichotomy,LockeandHumedeveloptheirkeystonemantrasintoanelaborateempiricisttheoryofhumancognitionthatismorenuancedandflexible.Infact,mostresearchindeeplearningismotivatedbyasetofassumptionsmoreconsistentwiththesephilosophers’lessradicaltakeonempiricism,andoneofthemaintasksofthisbookistoarticulateexactlywhichversionofempiricismismostsupportedbyrecentdevelopments.Identifyingandclarifyingthemoderatelyempiricistapproachtoooftenlostintheflashpointdebatescanunlockuntappedexplanatorypower,bothforunderstandingdeeplearning’scurrentmethodsandforchartingtheoptimalcoursetofuturebreakthroughs.Thechallengeisthataswithpoliticalslogans,eventheseeminglysimplestatementsoftheempiricistcreedcanmeandifferentthingstodifferentconstituencies.Byputtingintheinterpretiveworktounderstandthemcharitably,wecanavoidtalking-pastanddirectevaluativeeffortstowardsfruitfulfutureresearch.

Unsurprisingly,eventhemostprominentnativistsandempiriciststodayinterprettheaforementionedsloganstoimplyquitedifferentthings.Nativist-leaningtheoriststendtoassociateblankslateswiththelastgreatempiricistinferno,thebehavioristmovementinAmericanpsychology,whichreachedtheheightofitspowerandthenquicklydwindledtoembersinthemiddleofthelastcentury.Suchtheoriststypicallyconnecttheempiricistblankslatewithradicallyinsufficientexplanationsforhumanlearning.StevenPinkerarticulatesthisperspectiveclearlyinhisbookTheBlankSlate.AccordingtoPinker,today’sempiricistshaverevivedthe

4Ingeneral,Pearlislessconcernedherewiththedebateovernativismandanti-nativisminpsychologythantheseothercritics,andmoreengagedinthebattlebetweenskepticalHumeanandrealistapproachestocausationinmetaphysicsandphilosophyofscience.

doomedmissionofthebehaviorists,who“throughmostofthe20thcentury…triedtoexplainallofhumanbehaviorbyappealingtoacoupleofsimplemechanismsofassociationandconditioning”(Pinker2003).5Lakeetal.alsocalledoutthe“strongempiricismofmodernconnectionistmodels”whichtheyidentifywiththe“oversimplifiedbehaviorism”thatwas“repudiated”bythecognitiverevolutioninthelatterhalfofthe20thcentury(2017,p.4).ThisreportedabrogationoccurredwhenNoamChomskysmotheredbehaviorismunderawaveofhis“Cartesianlinguistics,”whichexplicitlyinvokedtherationalistnativismofFrenchphilosopherRenéDescartes(Chomsky1966)toinspirehisargumentsforanintricatesetofinnategrammaticalrulestoexplainhumanlinguisticability.6Marcusevenformalizesthisbehavioristinterpretationofempiricismbydefiningcognitionasafunctionrangingoverfourvariables:

cognition=f(a,r,k,e),

wherea=algorithms,r=representationalformats,k=innateknowledge,ande=experience.Marcus’construaloftheempiricistapproach—which,asmentionedabove,MarcusattributestoLocke—“wouldsetkandrtozero,setatosomeextremelyminimalvalue,(e.g.,anoperationforadjustingweightsrelativetoreinforcementsignals),andleavetheresttoexperience”(Marcus2018).7

Onthispoint,nativistspracticesomethingoftheradicalsimplificationtheycritique,byassumingthatforthemindtobe“blank”atbirth,itmustbeginwithvirtuallynoinnatestructureatall.ThemorecharitablenativistphilosophersLaurenceandMargolis(2015)haverecentlyworriedthatsummarizingcurrentdebatesincognitivescienceasthequestionofwhetherthemindhasanyinnatestructurewhatsoeverhastheunfortunateconsequencethat“therearen’treallyanyempiricists.”8Inreality,acompletelystructurelessmind,likeaninertmineralslab,wouldnotlearnanythingbybeingsubjectedtoanyamountofstimulus.Thisseems

5SeealsoChilders,Hvorecky,andMeyer(2021),whoalsolinkdeeplearningtobehaviorism;Idefendaverydifferentapproachtolinkingdeeplearningtothehistoryoftheempiricist-rationalistdebate.

6Whilealsoendorsingarichpackageof“startupsoftware”forthemind(whichintheirfavoredBayesianmodelsistypicallyprogrammedmanuallyinsymbolicform,includingmanuallyspecifiedrepresentationalprimitivesandpriorprobabilityestimations)whichtheythinkshouldincludecomponentsofCoreKnowledge,Lakeetal.(2017)areofficiallyagnosticastowhetherthatsoftwareisinnateorlearnedveryearlyinchildhood.

7Whatdoes“innate”meanhere?Anentiresubareaofphilosophyofsciencehasburgeonedaroundthequestionofhowbesttodefineinnateness(Ariew1996;GriffithsandMachery2008;Khalidi2001,2016,2016;MallonandWeinberg2006;MameliandBateson2006;NorthcottandPiccinini2018;Samuels2004,2007).Forpresentpurposes,wecanproceedwithaminimalistnotionthatimpliesatleast“notlearned”(Ritchie2021).

8Theempiricist-leaningdevelopmentalpsychologistLindaSmithhasalsocriticizedthisframinginherarticle,“Avoidingassociationswhenit’sbehaviorismyoureallyhate”(Smith2000).

tobesomethingthatnearlyallinfluentialempiricistshaveacknowledged.Backinthetwilightofbehaviorism’sreign,theempiricistphilosopherWillardvanOrmanQuineobservedthateventhemostradicalbehaviorists,likeJohnWatsonandB.F.Skinner,were“knowinglyandcheerfullyupto[their]neckininnatemechanisms”(quotedinLaurenceandMargolis2015;Quine1969:95–96):theymustassumearicharrayofbiologicalneeds,sensorymechanisms,attentionalbiases,andreflexivebehaviorswhichcouldbeassociatedwithoneanotherbeforeeventhesimplestformsofassociativelearningcouldbegin.Theitemsonthislistsuitorganismstotheirevolutionarynicheswithoutappealtoinnateknowledgestructures,illustratingwhyamoredetailedexaminationofempiricist-brandedtheorizinginbothphilosophyandcomputerscienceisrequired.Amoresystematicexaminationofthehistoryofempiricisttheorizingquicklyrevealsappealstoinnatefactorsmoreexpansivethanthislist.Thus,whileMarcus’formalizedmodelofempiricismissharperthantheempiricistmantrasinitsimplications,itisalsolessuseful,particularlyifweaimforacharitableevaluationofdeeplearning’sprospects.

Theprecedingillustrationoftheempiricist-nativistdichotomy,asitinformsthedevelopmentofdeeplearningsystems,offersaparadigmaticexampleofthenature-nurturedichotomy’senduringinfluenceonhumanthought.Bothdistinctionsaretoooftenresolvedintostarkbinaries,whereasthedebateisbetterrepresentedintermsofsubtlecontinuumsanddifferencesamongststylesofexplanation.Althoughthepersistenceoftheoppositionbetweennatureandnurturesuggestsanunsolvablephilosophicalriddleattheheartofknowledgeacquisition,itcan,withcare,beofusetous.Thesameistrueoftheempiricist-nativistdichotomy.Wheninterpretedwithmoreattentiontothehistoryofphilosophyanditsprecisecontextofapplication,itcanencouragemoreusefulandprincipleddebatesbetweendistinctresearchmethodologies.

Infact,incaseswherescientistshavetakenpainstounderstandthedebate’shistory,itcanbeseentohavefosterednotablescientificdiscoveriesofthelastcentury,suchasAlbertEinstein’stheoryofspecialrelativityortheveryinventionofthedigitalcomputerandartificialneuralnetworksoverwhichtoday’sdebatesrage.ThephilosopherofscienceJohnNortonarguesthatEinstein’stheoryofspecialrelativitywasinspiredbyhisparticipationinareadinggrouponHume’sTreatisearound1902-1903withthemathematicianConradHabichtandphilosopherMauriceSolovine,fromwhichEinsteinobtainedadeepregardforHume’s

empiricism.Inautobiographicalnotesfrom1946,Einsteinwritesofhisdiscoveryoftherelativityofsimultaneity(toaninertialframeofreference)whichundergirdsspecialrelativitythat“thiscentralpointwasdecisivelyfurthered,inmycase,bythereadingofDavidHume’sandErnstMach’sphilosophicalwritings”(quotedinNorton2010).WhilerationalistphilosopherslikeImmanuelKantthoughtthatabsolutesimultaneitywasnecessarilyentailedbyouraprioriconceptionofspacetime,Einsteinreasonedthatifeventhesebedrockconceptswerelearnedfromexperience,thentheremightbeexceptionstotheminextremeconditions,suchaswhenobjectstravelatvelocitiesapproachingthespeedoflight.

Equallymomentousachievementscanbeattributedtoscientistslisteningtothenativistmuse;theneuroscientistGraceLindsayrecountshowtheearlyneuralnetworkpioneersMcCullochandPitts(1943)idolizedtherationalistphilosopherGottfriedLeibniz,whotheorizedthatthemindoperatesoveraninnatelogicalcalculusfromwhichalltruepropositionscouldbemechanicallydeduced(Lindsay2021Ch.3).

McCullochandPitts’ideathatthesecomplexlogicalandmathematicaloperationscouldbecomputedbylargenumbersofsimplecomponentsorganizedintherightkindofnetworkarrangementservedasdirectinspirationforbothJohnvonNeumann(1993)andFrankRosenblatt(1958),whoseworkscanbeseentohaveproducedboththeopposingresearchtraditionsresponsibleforthedigitalmicroprocessorarchitectureanddeepneuralnetworks(DNNs),respectively.

Here,Iarguethatthecurrentincarnationofthenativist-empiricistdebateinartificialintelligencepresentsuswithasimilargoldenopportunity,inwhichwemightattemptoneoftherarestfeatsofintellectualalchemy:theconversionofatimelessphilosophicalriddleintoatestableempiricalquestion.For,ifwecouldapplythedistinctiontothedeeplearningdebatewithoutconfusionorcaricature,thenwecouldsimplybuildsomeartificialagentsaccordingtonativistprinciples,andotherartificialagentsaccordingtoempiricistprinciples,andseewhichonesareultimatelythemostsuccessfulorhuman-like.Specifically,wecanmanuallyprogramthenativistsystemswithinnateabstractknowledge,andendowempiricistsystemswithgeneralcapacitiestolearnabstractknowledgefromsensoryexperience,andcomparetheperformanceofthesystemsonarangeofimportanttasks.Crucially,however,theempiricistsinthiscompetitionmustbeallowedmorerawmaterialsthanMarcus’formalspecificationallows,ifweaimtoholdafairandinformativecompetition.

Ifwecouldaccomplishthisconversion,philosophersandcomputerscientistswouldbothreaptherewards.Onthephilosophyside,empiricistshavefrequentlybeenaccusedofappealingtomagicatcriticalpointsintheirtheoriesofrationalcognition.LockeandHume,forexampleoftenassertedthatthemindperformssomeoperationwhichallowsittoextractsomeparticularbitofabstractknowledgefromexperiencebut—giventhescantunderstandingofthebrain’soperationsavailableatthetime—theycouldnotexplainhow.Carefullyexaminingthedetailsofrecentdeeplearningachievementsmightredeemsomeofthelargestsuchpromissorynotes,byshowinghowphysicalsystemsbuiltaccordingtoempiricistprinciplescanactuallyperformtheseoperations.Indexingthephilosophicaldebatetothesesystemscanfurtherimproveitsclarity;wherephilosophicalslogansarevagueandsubjecttointerpretation,computationalmodelsareprecise,withalltheirassumptionsexposedforphilosophicalscrutinyandempiricalvalidation.Wheresuccessful,theplausibilityoftheempiricistapproachtorationalcognitionsubstantiallyincreasesasaresult.Ofthebenefitstocomputerscience,philosophershavethoughtlongandhardaboutthechallengeofprovidingacompleteapproachtothehumanmindthatisconsistentwithempiricistconstraints,includinghowthemind’svariouscomponentsmightinteractandscaleuptothehighestformsofabstractknowledgeandrationalcognition.

Deeplearningisonlynowreachingfortheseheightsinitsmodelingambitions(e.g.GoyalandBengio2020),andsotheremaystillyetbeinsightstominefromthehistoryofempiricistphilosophythatcanbetransmutedintothenextengineeringinnovations.

Totheseends,Iheremountaninterdisciplinaryinvestigationintotheprospectsandimplicationsofrecentachievementsindeeplearning,combininginsightsfrombothcomputerscienceandphilosophy.Doingsocanbothanimatecurrentengineeringresearchwiththewarmthandwisdomofaclassicphilosophicaldebate,whilstsimultaneouslyrenderingthetermsofthatdebateclearerthantheyhaveyetbeeninitslonganddistinguishedhistory.Nevertheless,Iknowthatsuchaninterdisciplinaryprojectisbesetwithitsowndistinctiverisk.RichardEvans—aninterdisciplinaryresearcheratDeepMindwhohassoughttocreatemorepowerfuldeeplearningsystemsbyaugmentingthemwithlogicalmaximsthatheextractsfromKant’sCritiqueofPureReason(includingKant’saforementionedmaximofsimultaneity)—hasissuedasalutarywarningforprojectsembarkingundersuchambitions:

Thisisaninterdisciplinaryprojectandassuchisinever-presentdangeroffallingbetweentwostools,

neitherphilosophicallyfaithfultoKant’sintentionsnorcontributingmeaningfullytoAIresearch.Kanthimselfprovides:‘thewarningnottocarryonatthesametimetwojobswhichareverydistinctinthewaytheyaretobehandled,foreachofwhichaspecialtalentisperhapsrequired,andthecombinationofwhichinonepersonproducesonlybunglers.’[AK4:388]Thedangerwithaninterdisciplinaryproject,partAIandpartphilosophy,isthatbothpotentialaudiencesareunsatis?ed.(Evans2020)

WemusttakeEvans’s(andKant’s)warningtoheart.Yet,wemustalsoacknowledgethat,inpartbecausedeeplearningisimplicatedinthenature-nurturedistinction,philosophersareparticularlysuitedtoundertaketheproject.Whateverourotherbungles,wehaveexperiencetendingtothisparticularfire.Toproceed,however,wemustdiscardstoolsaltogether.Wewillbebetterabletogaugethecurrentandfutureachievementsofdeeplearningbyinsteadbuildingamoreaccommodatingbench,withampleroomforaspectrumofdistinctivebackgroundsandexpertise.Giventheintensityofthecurrentdiscussionamongsttheoristsgrapplingwithdeeplearning’spotential,themostproductivewayforwardinvolvesloweringthedebate’stemperatureuntilthesmokeclears,andinvitingtheoristsfromavarietyofbackgroundswithdistinctiveexpertiseandastakeindeeplearning’simplicationstopatientlyworkthroughthedetailstogether.

Howtosimmerthingsdown:FromFormsandslatestostylesoflearning

Thankstorigorousinvestigationinseveraldisciplines,todayweknowthatnearlyallknowledgeoriginatesfromacombinationofbothinnateandexperientialfactors.Bothradicalnativismandradicalempiricismare,inshort,false.Despitethis,morenuancedexpositionsofthedistinctionbetweenempiricismandnativismremaintheexceptionandhavebeenalmostentirelyabsentfromdiscussionsoverdeeplearning.9Withoutabetterwaytounderstandthesubstanceofthedistinction,therecognitionofthisecumenicaloutcomecarries,onbothsides,thethreatofobliteration.ThismaybewhyLockeandHume’se