History of Artificial Intelligence 人工智能的歷史

1HistoryofArtificialIntelligenceDanaNejedlováDepartmentofInformaticsFacultyofEconomicsTechnicalUniversityofLiberec2WhatisIntelligence?Commondefinitionofartificialintelligence:AIisafieldwhichattemptstobuildintelligentmachinesandtriestounderstandintelligententities.Butwhatisintelligence?Learning,manipulatingwithfacts,butalsocreativity,consciousness,emotionandintuition.Canmachinesbeintelligent?Uptothepresentdayitisnotsurewhetheritispossibletobuildamachinethathasallaspectsofintelligence.ThiskindofresearchiscentralinthefieldofAI.3WhatIsArtificialIntelligence?Buildingmachinesthatareableofsymbolicprocessing,recognition,learning,andotherformsofinferenceSolvingproblemsthatmustuseheuristicsearchinsteadofanalyticapproachUsinginexact,missing,orpoorlydefinedinformationFindingrepresentationalformalismstocompensatethisReasoningaboutsignificantqualitativefeaturesofasituationWorkingwithsyntaxandsemanticsFindinganswersthatareneitherexactnoroptimalbutinsomesense?sufficient“Theuseoflargeamountsofdomain-specificknowledgeTheuseofmeta-levelknowledge(knowledgeaboutknowledge)toeffectmoresophisticatedcontrolofproblemsolvingstrategies4BeforetheCreationofElectronicComputersAncientandmedievalmythsTalos,Pandora,Golemartificialmen,robots,automatonsBook“GodsandRobots”(2018)Today’sAIceasestobecomprehensiblesimilarlytoancientrobots’innerworkings.Researchintheantiquitytillthe17thcenturyAristotle,RamonLlull,GottfriedWilhelmLeibnizautomationofreasoningThomasHobbes,RenéDescartesmechanisticunderstandingoflivingbeings20thcentury,1948NorbertWiener–Cybernetics:OrtheControlandCommunicationintheAnimalandtheMachine.Intelligentbehavioristheresultofthefeedbackmechanism.5TheBeginningsofElectronicComputersJohnLouisvonNeumann(1903–1957)VonNeumann’sarchitectureofacomputerConsultationsontheEDVACProject(1945)GameTheory(1944)Itcanbeappliedtotheinteractingintelligentagents.Cellularautomata(1966)Theyhavecomputationalcapacity.AlanMathisonTuring(1912–1954)TuringMachine(1936)formalizationofalgorithm,abstractionofcomputerTuringTest(1950)proposalhowtotesttheabilityofamachinetodemonstratethinkingProgrammingof“ManchesterMarkI”computer(1949)6Thebirthof“ArtificialIntelligence”JohnMcCarthyusedtheterm“ArtificialIntelligence”forthefirsttimeasthetopicoftheDartmouthconferencein1956.Venue:DartmouthCollege,Hanover,stateNewHamphshire,USAOrganizers:JohnMcCarthy,MarvinMinsky,NathanielRochester,andClaudeShannonParticipants:RaySolomonoff,OliverSelfridge,TrenchardMore,ArthurSamuel,HerbertSimon,andAllenNewellProposal:Toprovethateveryaspectoflearningoranyotherfeatureofintelligencecanbesopreciselydescribedthatamachinecanbemadetosimulateit.7ApproachestoArtificialIntelligenceGoodOld-fashionedArtificialIntelligence(GOFAI)orsymbolicartificialintelligence(JohnHaugeland,1985)Program(e.g.classifier)intheGOFAIstyleiscomposedofparts(e.g.rules),thathaveclearrelationtotherealworld.New-fangledArtificialIntelligenceThemostimportantbranchwasconnectionism–artificialneuralnetworks(McCulloch–Pitts,1943).Thecentralideaisthatalargenumberofsimplecomputationalunitscanachieveintelligentbehaviorwhennetworkedtogether.Geneticalgorithms(Holland,1975)andotherkindsofbiologicallyinspiredinformationprocessingStrongAI(JohnSearle,1980)(CurrentlycalledAGI.)Artificialintelligenceisrealintelligence.Solutionofcomplexproblems,e.g.robotics.WeakAI(CurrentlycallednarrowAI.)Artificialintelligenceisamereimitationofhumanrealintelligence.Solutionofaspecificproblemsthatdonotcoverthewholescaleofhumancapabilities,e.g.OCRorchess.8MotivationsforBiologicallyInspiredInformationProcessingDannyHillis:TheConnectionMachine(1985)MachinesprogrammedinaGOFAIstyletendtoslowdownastheyacquiremoreknowledge.Theymustsearchtheirknowledgebase.Humanshavetheoppositeproperty.Theyhavemassivelyparallelbrainarchitecture.Humanswerenotproducedbyanengineeringprocess.Theyaretheresultofevolution.MarvinMinsky:TheSocietyofMind(1986)Modelofhumanintelligencewhichisbuiltfromtheinteractionsofsimplepartscalledagentswhicharethemselvesmindless.Itwouldbedifficulttoimaginehowevolutioncouldshapeasinglesystemascomplexasmind.Evolutioncould,however,shapeindividualspecializedcognitiveunitsandformthemechanismsthatenablethemodulestointeract.MarvinMinsky:TheEmotionMachine(2006)Emotionsaredifferentwaystothinkthatourmindusestoincreaseourintelligence.9ArtificialIntelligencePhilosophyWhatisintelligenceandthinking?Turingtest(1950)AccordingtoGOFAIthinkingissymbolmanipulation,thatiswhyprogramintheGOFAIstyleisthinking.ChineseRoomProblem(JohnSearle,1980)Thinkingofhumansandcomputersisdifferent.Ishumanintelligenceinseparablefrommindandemotions?Inwhatsensecanwesaythatacomputercanunderstandnaturallanguage?WhoisresponsibleforthedecisionsmadebyAI?Whatshouldbetheethicsofpeopleofdealingwiththecreationsofartificialintelligence?treatingrobotsusingproductsofgenerativemachinelearning10HardVersusSoftComputingGoodOld-fashionedArtificialIntelligenceIF–THENRulesHeuristicsNew-fangledArtificialIntelligenceNeuralnetworksFuzzylogicProbabilisticreasoningbeliefnetworks(Bayesnetworks)geneticalgorithmschaostheorypartsoflearningtheory(machinelearning)11HeuristicsProblem-solvingmethodthatisusuallysuccessful,butcanfailisomesituationsUnclearlydefinedproblemswithmissingorambiguousdataMedicaldiagnosisVision,speechrecognitionHelpstodecideamonginfinitenumberofpossibleinterpretations.Aproblemmayhaveanexactsolution,butthecomputationalcostoffindingitmaybeprohibitive.Chess,tic-tac-toe,15or8-puzzle,scheduling,path-finding…HeuristicevaluationfunctionEvaluateseachstageofsolution.NumberofconflictsinanumberofpossibleschedulesHelpstodecideaboutthenextstepleadingtothegoal.Selectingtheschedulewithminimumnumberofconflictsforthenextsmallchangesattemptingtofindsomecorrectschedule12ExpectationsfromArtificialIntelligencePredictionsofHerbertSimonandAllenNewell(HeuristicProblemSolving,1958),thatwithintenyearsadigitalcomputerwillbetheworld'schesschampion,adigitalcomputerwilldiscoverandproveanimportantnewmathematicaltheorem,adigitalcomputerwillcomposecriticallyacclaimedmusic,mosttheoriesinpsychologywilltaketheformofcomputerprograms.AndrewNg(Chinese-Americancomputerscientistfocusingonimprovingpeople’slivesusingAI)Ifatypicalpersoncandoamentaltaskwithlessthanonesecondofthought,wecanprobablyautomateitusingAIeithernoworinthenearfuture./2016/11/what-artificial-intelligence-can-and-cant-do-right-now13TypicalAIProblem8QueensPuzzleIsthereawayofplacing8queensonthechessboardsothatnotwoqueenswouldbeabletoattackeachother?14HardProblemforAITruncatedChessboardProblemIsthereawayofplacingdominosontheboardsothateachsquareiscoveredandeachdominocoversexactlytwosquares?Peoplesolvetheproblemeasily,butitishardforthemtodescribeitformally.15LimitationsofArtificialIntelligenceDavidHilbert(1862–1943)andKurtG?del(1906–1978)G?del‘sIncompletenessTheorem(1931)Consistencyofaformalsystemcannotbeprovedwithinthesystem,becauseitcancontainstatementswithself-reference–logicalparadoxesoftheliarparadoxtype:Thisstatementisfalse.Sometaskshavenoalgorithms.HaltingproblemItisnotdecidablewhetherthealgorithmwillhaltornot.Thealgorithmsinquestioncontainagainself-reference.ComplexityTheory(NP-completeness,1971)Sometaskshavealgorithms,butthecomputationcannotbecompletedinpractice(onarealcomputer),becauseitwouldtaketoomuchtime.RogerPenrose(booksTheEmperor‘sNewMind,ShadowsoftheMind)Itmaynotbepossibletocompletelysimulatebiologicalintelligencebycomputationalapproachesasitmaybebasedon(apparentlyquantum)phenomenathatwedonotknowandarenotabletoimitate.16ThreatsofArtificialIntelligenceTechnologicalSingularityHypothesisthattheinventionofartificialsuperintelligencewouldacceleratetechnologicalprogresswhichwouldhaveunpredictableeffectonhumansocietyExistentialriskfromartificialgeneralintelligenceBook?Superintelligence“byNickBostromGoalsofAIshouldbecarefullydefinedsothatAIfulfillingthesegoalsdoesnotdestroyhumans.

ExampleSuperintelligenceisdefinedbyBostromasanartificialintellectthatissuperiortohumanintellectineveryaspectincludingcreativity,generalwisdom,andsocialskills./blog/articles/ai-quotes-from-some-of-the-worlds-top-minds//bizchina/tech/2017-07/28/content_30278816.htmPeopleviewingAIasthreatElonMusk,StephenHawking,GeoffreyHinton,JamesCameronPeopleviewingAIasbenefittomankindMarkZuckenberg,AndrewNgWeshouldbemoreconcernedwithhowhumansabusethepowerAIoffers.autonomousweaponssystemsfacialrecognitionsystemsusedformasssurveillanceofpeoplesystemsmonitoringpeople’sbehaviorontheinternetusedfortheirscoringChatbotsonsocialnetworksinfluencepoliticalviews.Neuralnetworkscanbefooledintoerroneousrecognitionbypresentingspecificallymodifiedinput(pictures,text,audio)tothem(Adversarialattacks).Peoplecanbefooledbydeepfakes.ThecopyrightownersoftheworksonwhichtheAI??learnsarenotyetcompensated.17G?del‘sIncompletenessTheoremThereareunprovablestatementsineveryaxiomaticmathematicalsystemexpressiveenoughtodefinethesetofnaturalnumbers.Exampletheorem1=2Proofofthetheorem:Ifa=b,a≠0,b≠0,thenthetwofollowingequalitiesarealsotrue:a2–b2=(a–b)?(a+b),a2–b2=a2–ab.Andthefollowingstatementscanbederivedfromthem:a2–ab=(a–b)?(a+b)a?(a–b)=(a–b)?(a+b)a=a+ba=a+aa=2a1=2Truthcanbeverifiedonlywhenknowledgebeyondthenaturalfinitenumbersarithmeticisused.18TheLogicTheorist–TheFirstArtificialIntelligenceProgramAllenNewell,J.C.ShawandHerbertSimonatCarnegieInstituteofTechnology,nowCarnegieMellonUniversity,in1955Itdidlogicproofsfromthebook“PrincipiaMathematica”(BertrandRussellandAlfredNorthWhitehead,1910).Itusedmentalprocessesofhumanexperts.cognitivescienceToimplementLogicTheoristonacomputer,thethreeresearchersdevelopedaprogramminglanguage,IPL,apredecessorofLisp.19ProgrammingLanguagesTaskslikenaturallanguageprocessing,knowledgerepresentation,ortheoremprovingneededaspeciallanguageallowingprocessingofsymbolicdata.Lisp(JohnMcCarthy,USA,1958)functionalparadigm/listprocessingProgramconsistsoffunctionsofnestedfunctions.Dataandprogramsarerepresentedthesameway:alist.(+123)isabothalistof4atomsandafunctionreturningvalue6.Programcanserveasdataforanotherprogram!Powerfulfeatureallowingflexibleandproductivecoding.Prolog(AlainColmerauer,Europe,1972)declarativeparadigm/logicprogrammingProgramconsistsoffactsandrules.Programmerdescribes(i.e.declares)aproblem.Compilerdeducesnewfactsfromthem.Programmerdoesnotwritethealgorithmforthesolution.Differentiableprogrammingnewparadigmenablingmachinelearningprogramstochangetheirstructureaccordingtodatathankstonewcompilersreferences:medium1,medium2,medium3,towardsdatascience20ProgramswithSymbolicArtificialIntelligenceTheGeneralProblemSolver(1957)Itwassolvingformalizedsymbolicproblems,e.g.mathematicalproofsandchess.TheGeometryTheoremProver(1958)Itwasprovingtheoremswiththehelpofexplicitlyrepresentedaxioms.SAINT(SymbolicAutomaticINTegrator)Integralcalculus(1961)ANALOGY(1963)ThepictureAistopictureBlikepictureCtopictureD.IQtestsareusedformeasuringtheintelligenceofpeople.ComputerscanbeprogrammedtoexcelinIQtests.Butthoseprogramswouldbestupidinreal-worldsituations.21NaturalLanguageProcessingSTUDENT(1964,1967)Itwassolvingwordproblemsinalgebra.SIR(SemanticInformationRetrieval,1968)Itwasreadingsimplesentencesandansweredquestions.ELIZA(1965)Itwassimulatingpsychologist.TLC(TeachableLanguageComprehender)(1969)Itwasreadingtextandmakingsemanticnetwork.SUR(SpeechUnderstandingResearch)(1971)5-yearplanoftheARPA(todayDARPA)agencyofaresearchincontinuousspeechrecognition22ExpertSystemsTheybelongtothesymbolicAI.Theyuseasetofrulesandheuristics.MACSYMA(MIT,1968-1982)Itwasdoingsymbolicmathcalculations.DENDRAL(SRI,1965)Itisidentifyingchemicals.MYCIN(SRI,EdwardShortliffe,1974)Itdiagnosedinfectiousblooddiseases.Thefollowingsystems:EMYCIN,PUFF,INTERNIST-CADUCEUS23CommercialExpertSystemsPROSPECTOR(SRI,1974–1983)Itisanalyzinggeologicaldataandsearchingfordepositsofminerals.XCON–eXpertCONfigurer(CMU,1978)ItwasconfiguringDEC’sVAXcomputers.TEIRESIAS(SRI,RandallDavis,1976)KnowledgeAcquisitionSystem(KAS)Itisacquiringknowledgefromhumanexperts.Itisbuildingknowledgebasesforexpertsystems.24RoboticsMarvinLeeMinsky(*1927)Freddy(UniversityofEdinburgh,1973)SHAKEY(SRI,1969)SHRDLU(MIT,TerryWinograd,1970)blocksworlds(MIT,1970)Robothastomanipulatebuildingblocksaccordingtoinstructions.computervisionnaturallanguageunderstandingplanning25TheFirstArtificialNeuralNetworksWarrenMcCullochandWalterPittsModelofartificialneuron(1943)Neuronrepresentsfunctions.DonaldOldingHebbRuleforneuralnetworktraining(1949)MarvinMinskyandDeanEdmondshavebuiltthefirstcomputerwithneuralnetwork.SNARC(1951)26OtherArtificialNeuralNetworksFrankRosenblattPerceptron(1957)asingle-layernetworkanditslearningrulecapableoflearninglinearlyseparablefunctionsBernardWidrowandMarcianTedHoffMinimizationofnetwork’srootsquareerrorDeltarule(learningruleofaneuralnetwork)ADAptiveLINEarSystemsorneuronsorADALINEs(1960)MADALINEs(1962)multi-layerversionsofADALINEs27NeuralNetworksCritiqueBook?Perceptrons“(MarvinMinskyandSeymourPapert,1969)Whensingle-layerneuralnetworksofaPerceptrontypecannotlearnXORfunction(itislinearlyinseparable),alsomulti-layernetworkscannotlearnit.Hencefundingofneuralnetworkresearchwasstoppeduntilthebeginningofthe20thcentury80’s.Butmulti-layerneuralnetworkscanlearntheXORfunction.Allthatisneededforthisistofindtherightalgorithmfortheirtraining.28NeuralNetworksResurrectionHopfieldnet(JohnHopfield,1982)Itcanlearnacoupleofpictures(patterns).Self-OrganizingMap(SOM)(TeuvoKohonen,1982)Itcandounsupervisedlearning.Backpropagation(ArthurBrysonandYu-ChiHo,1969)algorithmfortrainingofamultilayerneuralnetworkItneedsnetwork’sneuronsnottohaveasharpthreshold.Becauseitwasnotnoticed,itwasthenrediscoveredseveraltimesinthe70’sandthe80’softhe20thcenturyandpopularizedin1986.NETtalk(TerrySejnowskiandCharlesRosenberg,1986)Multi-layerneuralnetwork,thatlearnedEnglishpronunciationandcouldgeneralize.Itusedthebackpropagationalgorithm.29NeuralNetworksatPresentNeuralnetworkswiththequantityoflayersandneuronsusedsince1980shadproblemsinlearning.Since2006waysoftraininglargenetworksareaccomplishedtobefound.Thesenetworkshavehuman-levelperformance.DeeplearningneuralnetworkswithlargernumberoflayersDeepBeliefNetworkItslayerscanbetrainedseparatelybyunsupervisedlearning.multilayerneuralnetworktrainedusingbackpropagationItsvariantisconvolutionalneuralnetworkforrecognitionofimages.Themaingoalistoachievetheabilityofextractionofthefeaturesoftheobjecttoberecognizedondifferentlevelsofabstraction.Seee.g.alectureofYoshuaBengio.Pixelsaretheinputintothefirstlayerofthenetwork,edges,contours,andobjectpartsarerecognizedonsubsequentlayers,theclassoftheobjectisreturnedbytheoutputlayer.Differentlevelsofabstractionaresolvedondifferentlayersofthenetwork.Thenetworkassignsthelevelsofabstractiontoitslayersbyitsowndecision.Prerequisitestheexistenceoflargequantitiesofdataandtrainingdatasets,e.g.:labeledimagestextsandtheirtranslationsintomultiplelanguagesdistributedcomputing(e.g.cloud),morepowerfulCPUs,utilizingofGPUsCritiqueGeoffreyHintonsays"Myviewisthrowitallawayandstartagain."Peopledonotneedasmuchdataassupervisedlearningdoes.30PresentApplicationsofNeuralNetworksNaturallanguageprocessing(e.g.IBMProjectDebater)GenerationofimagesfromotherimagesorfromtextcaptionsalgorithmsofGoogleaFacebookOpenAI’sDALL·EandfreeDALL·EminiMusiccompositionwiththehelpofmachinelearninge.g.neuralnetworkCoconettrainedonthemusicofJ.S.BachReinforcementlearningSolvestasks