国会研究服务部-人工智能:背景、选定问题和政策考虑(英文)-2021.5-47正式版_第1页
国会研究服务部-人工智能:背景、选定问题和政策考虑(英文)-2021.5-47正式版_第2页
国会研究服务部-人工智能:背景、选定问题和政策考虑(英文)-2021.5-47正式版_第3页
国会研究服务部-人工智能:背景、选定问题和政策考虑(英文)-2021.5-47正式版_第4页
国会研究服务部-人工智能:背景、选定问题和政策考虑(英文)-2021.5-47正式版_第5页
已阅读5页,还剩42页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

May19,2021

CongressionalResearchService

R46795

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

SUMMARY

R46795

May19,2021

Thefieldofartificialintelligence(AI)—atermfirstusedinthe1950s—hasgonethrough

LaurieA.Harris

AnalystinScienceand

multiplewavesofadvancementoverthesubsequentdecades.Today,AIcanbroadlybethought

TechnologyPolicy

ofascomputerizedsystemsthatworkandreactinwayscommonlythoughttorequire

intelligence,suchastheabilitytolearn,solveproblems,andachievegoalsunderuncertainand

varyingconditions.Thefieldencompassesarangeofmethodologiesandapplicationareas,

includingmachinelearning(ML),naturallanguageprocessing,androbotics.

Inthepastdecadeorso,increasedcomputingpower,theaccumulationofbigdata,andadvancesinAItechniqueshaveledtorapidgrowthinAIresearchandapplications.GiventhesedevelopmentsandtheincreasingapplicationofAItechnologiesacrosseconomicsectors,stakeholdersfromacademia,industry,andcivilsocietyhavecalledforthefederalgovernmenttobecomemoreknowledgeableaboutAItechnologiesandmoreproactiveinconsideringpublicpoliciesaroundtheiruse.

FederalactivityaddressingAIacceleratedduringthe115thand116thCongresses.PresidentDonaldTrumpissuedtwoexecutiveorders,establishingtheAmericanAIInitiative(E.O.13859)andpromotingtheuseoftrustworthyAIinthefederalgovernment(E.O.13960).Federalcommittees,workinggroups,andotherentitieshavebeenformedtocoordinateagencyactivities,helpsetpriorities,andproducenationalstrategicplansandreports,includinganupdatedNationalAIResearchandDevelopmentStrategicPlanandaPlanforFederalEngagementinDevelopingTechnicalStandardsandRelatedToolsinAI.InCongress,committeesheldnumeroushearings,andMembersintroducedawidevarietyoflegislationtoaddressfederalAIinvestmentsandtheircoordination;AI-relatedissuessuchasalgorithmicbiasandworkforceimpacts;andAItechnologiessuchasfacialrecognitionanddeepfakes.Atleastfourlawsenactedinthe116thCongressfocusedonAIorincludedAI-focusedprovisions.

TheNationalDefenseAuthorizationActforFY2021(P.L.116-283)includedprovisionsaddressingvariousdefense-andsecurity-relatedAIactivities,aswellastheexpansiveNationalArtificialIntelligenceInitiativeActof2020(DivisionE).

TheConsolidatedAppropriationsAct,2021(P.L.116-260)includedtheAIinGovernmentActof2020(DivisionU,TitleI),whichdirectedtheGeneralServicesAdministrationtocreateanAICenterofExcellencetofacilitatetheadoptionofAItechnologiesinthefederalgovernment.

TheIdentifyingOutputsofGenerativeAdversarialNetworks(IOGAN)Act(P.L.116-258)supportedresearchonGenerativeAdversarialNetworks(GANs),theprimarytechnologyusedtocreatedeepfakes.

P.L.116-94establishedafinancialprogramrelatedtoexportsinAIamongotherareas.

AIholdspotentialbenefitsandopportunities,butalsochallengesandpitfalls.Forexample,AItechnologiescanaccelerateandprovideinsightsintodataprocessing;augmenthumandecisionmaking;optimizeperformanceforcomplextasksandsystems;andimprovesafetyforpeopleindangerousoccupations.Ontheotherhand,AIsystemsmayperpetuateoramplifybias,maynotyetbefullyabletoexplaintheirdecisionmaking,andoftendependonvastdatasetsthatarenotwidelyaccessibletofacilitateresearchanddevelopment(R&D).Further,stakeholdershavequestionedtheadequacyofhumancapitalinboththepublicandprivatesectorstodevelopandworkwithAI,aswellastheadequacyofcurrentlawsandregulationsfordealingwithsocietalandethicalissuesthatmayarise.Together,suchchallengescanleadtoaninabilitytofullyassessandunderstandtheoperationsandoutputsofAIsystems—sometimesreferredtoasthe“blackbox”problem.

Becauseofthesequestionsandconcerns,somestakeholdershaveadvocatedforslowingthepaceofAIdevelopmentanduseuntilmoreresearch,policymaking,andpreparationcanoccur.OthershavecounteredthatAIwillmakelivessafer,healthier,andmoreproductive,sothefederalgovernmentshouldnotattempttoslowit,butrathershouldgivebroadsupporttoAItechnologiesandincreasefederalAIfunding.

Inresponsetothisdebate,Congresshasbegundiscussingissuessuchasthetrustworthiness,potentialbias,andethicalusesofAI;possibledisruptiveimpactstotheU.S.workforce;theadequacyofU.S.expertiseandtraininginAI;domesticandinternationaleffortstosettechnologicalstandardsandtestingbenchmarks;andthelevelofU.S.federalinvestmentsinAIresearchanddevelopmentandhowthatimpactsU.S.internationalcompetitiveness.CongressislikelytocontinuegrapplingwiththeseissuesandworkingtocraftpoliciesthatprotectAmericancitizenswhilemaximizingU.S.innovationandleadership.

CongressionalResearchService

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

Contents

Introduction

1

WhatIsAI?

1

AITerminology

3

AlgorithmsandAI

5

HistoricalContextofAI

5

WavesofAI

5

RecentGrowthintheFieldofAI

6

AIResearchandDevelopment

6

PrivateandPublicFunding

8

SelectedResearchandFocusAreas

11

ExplainableAI

11

DataAccess

12

AITrainingwithSmallandAlternativeDatasets

14

AIHardware

15

FederalActivityinAI

16

ExecutiveBranch

16

ExecutiveOrdersonAI

17

NationalScienceandTechnologyCouncilCommittees

17

SelectAIReportsandDocuments

18

FederalAgencyActivities

19

Congress

22

Legislation

23

Hearings

26

SelectedIssuesforCongressionalConsideration

27

ImplicationsfortheU.S.Workforce

28

JobDisplacementandSkillShifts

28

AIExpertWorkforce

30

InternationalCompetitionandFederalInvestmentinAIR&D

35

StandardsDevelopment

37

Ethics,Bias,Fairness,andTransparency

39

TypesofBias

41

Figures

Figure1.TotalNumberofAI-RelatedPublicationsonarXiv,

byFieldofStudy,2015-2020

8

Figure2.ExamplesofNon-ExplainableandExplainableAISystems

12

Figure3.MentionsofArtificialIntelligenceandMachineLearningintheCongressional

Record,2011-2020

23

Contacts

AuthorInformation

43

CongressionalResearchService

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

Introduction

Artificialintelligence(AI)—atermfirstusedinthe1950s—canbroadlybethoughtofascomputerizedsystemsthatworkandreactinwayscommonlythoughttorequireintelligence,suchastheabilitytolearn,solveproblems,andachievegoalsunderuncertainandvaryingconditions.1Inthepastdecade,increasesincomputingpower,theavailabilityoflarge-scaledatasets(i.e.,bigdata),andadvancesinthemethodologiesunderlyingAI,haveledtorapidgrowthinthefield.AItechnologiescurrentlyshowpromiseforimprovingthesafety,quality,andefficiencyofworkandforpromotinginnovationandeconomicgrowth.Atthesametime,theapplicationofAItocomplexproblemsolvinginreal-worldsituationsraisesconcernsabouttrustworthiness,bias,andethicsandpotentialdisruptiveeffectsontheU.S.workforce.Inaddition,numerouspolicyquestionsareatissue,includingthoseconcerningtheappropriateU.S.approachtointernationalcompetitioninAIresearchanddevelopment(R&D),technologicalstandardsetting,andthedevelopmentoftestingbenchmarks.

GiventheincreasinguseofAItechnologiesacrosseconomicsectors,stakeholdersfromacademia,industry,andcivilsocietyhavecalledforthefederalgovernmenttobecomemoreknowledgeableaboutAItechnologiesandmoreproactiveinconsideringpublicpoliciesaroundtheiruse.ToassistCongressinitsworkonAI,thisreportprovidesanoverviewofAItechnologiesandtheirdevelopment,recenttrendsinAI,federalAIactivity,andselectedissuesandpolicyconsiderations.

ThisreportdoesnotattempttoaddressallapplicationsofAI.InformationontheapplicationofAItechnologiesintransportation,nationalsecurity,andeducationcanbefoundinseparateCRSproducts.2

WhatIsAI?

Whilethereisnosingle,commonlyagreedupondefinitionofAI,theNationalInstituteofStandardsandTechnology(NIST)hasdescribedAItechnologiesandsystemsascomprising“softwareand/orhardwarethatcanlearntosolvecomplexproblems,makepredictionsorundertaketasksthatrequirehuman-likesensing(suchasvision,speech,andtouch),perception,cognition,planning,learning,communication,orphysicalaction.”3DefinitionsmayvaryaccordingtothedisciplineinwhichAIisbeingdiscussed.4AIisoftendescribedasafieldthatencompassesarangeofmethodologiesandapplicationareas,suchasmachinelearning(ML),naturallanguageprocessing(NLP),androbotics.

AdaptedfromOfficeofScienceandTechnologyPolicy,PreparingfortheFutureofArtificialIntelligence,October2016,p.6.

SeeCRSReportR44940,IssuesinAutonomousVehicleDeployment,byBillCanis;CRSInFocusIF10737,AutonomousandSemi-autonomousTrucks,byJohnFrittelli;CRSReportR45178,ArtificialIntelligenceandNationalSecurity,byKelleyM.Sayler;andCRSInFocusIF10937,ArtificialIntelligence(AI)andEducation,byJoyceJ.LuandLaurieA.Harris.

NationalInstituteofStandardsandTechnology,U.S.LeadershipinAI:APlanforFederalEngagementinDevelopingTechnicalStandardsandRelatedTools,August9,2019,pp.7-8.

See,forexample,AIdefinitionsinthecategoriesofordinarylanguage,computationaldisciplines,engineering,economicsandsocialsciences,ethicsandphilosophy,andinternationallawandpolicy,inSaraMattingly-Jordanetal.,EthicallyAlignedDesign:FirstEditionGlossary,InstituteofElectricalandElectronicsEngineers(IEEE),January2019,p.8,at/content/dam/ieee-standards/standards/web/documents/other/ead1e_glossary.pdf.

CongressionalResearchService 1

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

DefiningAIisnotmerelyanacademicexercise,particularlywhendraftinglegislation.AIresearchandapplicationsareevolvingrapidly.Thus,congressionalconsiderationofwhethertoincludeadefinitionforAIinabill,andifsohowtodefinethetermorrelatedterms,necessarilyincludeattentiontothescopeofthelegislationandthecurrentandfutureapplicabilityofthedefinition.ConsiderationsincraftingadefinitionforuseinlegislationincludewhetheritisexpansiveenoughnottohinderthefutureapplicabilityofalawasAIdevelopsandevolves,whilebeingnarrowenoughtoprovideclarityontheentitiesthelawaffects.Somestakeholders,recognizingthemanychallengesofdefiningAI,haveattemptedtodefineprinciplesthatmighthelpguidepolicymakers.ResearchsuggeststhatdifferencesindefinitionsusedtoidentifyAI-relatedresearchmaycontributetosignificantlydifferentanalysesandoutcomesregardingAIcompetition,investments,technologytransfer,andapplicationforecasts.5

TheJohnS.McCainNationalDefenseAuthorizationActforFiscalYear2019(P.L.115-232)includedthefirstdefinitionofAIinfederalstatute.6Likethoseinotherpreviouslyintroducedbills,thedefinitionincorporatedacommonlycitedframeworkoffourpossiblegoalsthatAIsystemsmaypursue:systemsthatthinklikehumans(e.g.,neuralnetworks),actlikehumans(e.g.,naturallanguageprocessing),thinkrationally(e.g.,logicsolvers),oractrationally(e.g.,intelligentsoftwareagentsembodiedinrobots).7However,AIresearchandapplicationsdonotnecessarilyfallsolelywithinanyoneofthesefourcategories.

InDecember2020,theNationalArtificialIntelligenceActof2020,enactedaspartoftheWilliamM.(Mac)ThornberryNationalDefenseAuthorizationAct(NDAA)forFiscalYear2021(P.L.

116-283),includedthefollowingdefinition:

Theterm“artificialintelligence”meansamachine-basedsystemthatcan,foragivensetofhuman-definedobjectives,makepredictions,recommendationsordecisionsinfluencingrealorvirtualenvironments.Artificialintelligencesystemsusemachineandhuman-basedinputsto—(A)perceiverealandvirtualenvironments;(B)abstractsuchperceptionsintomodelsthroughanalysisinanautomatedmanner;and(C)usemodelinferencetoformulateoptionsforinformationoraction.8

CurrentAIsystemsareconsideredtobenarrowAI,meaningthattheyaretailoredtoparticular,narrowlydefinedtasks.ExampleapplicationsofAIineverydaylifeincludeemailspamfiltering,voiceassistance(e.g.,Siri,Alexa,Cortana),financiallendingdecisions,andsearchengineresults.AItechnologiesarebeingintegratedinarangeofsectors,includingtransportation,healthcare,education,agriculture,manufacturing,anddefense.SomeAIexpertsusethetermsaugmentedintelligenceorhuman-centeredAItocapturethevariousAIapplicationsinphysicalandconnectedsystems,suchasroboticsandtheInternetofThings,9andtoemphasizetheuseofAItechnologiestoenhancehumanactivitiesratherthantoreplacethem.

MostanalystsbelievethatgeneralAI,meaningsystemsthatdemonstrateintelligentbehavioracrossarangeofcognitivetasks,isunlikelytooccurforadecadeorlonger.SomeAIresearchers

DeweyMurdick,JamesDunham,andJenniferMelot,AIDefinitionsAffectPolicymaking,CenterforSecurityandEmergingTechnology,June2020,at/wp-content/uploads/CSET-AI-Definitions-Affect-Policymaking.pdf.

P.L.115-232,Section238;10U.S.C.§2358note.

StuartRussellandPeterNorvig,ArtificialIntelligence:AModernApproach,3rded.(UpperSaddleRiver,NJ:PrenticeHall,2010),pp.1-5.

P.L.116-283(hereinafterreferredtoastheFY2021NDAA);H.R.6395,DivisionE,Section5002(3).

FormoreinformationontheInternetofThings,seeCRSInFocusIF11239,TheInternetofThings(IoT):AnOverview,byPatriciaMoloneyFigliola;andtoidentifyadditionalCRSexpertswhoworkonIoTandrelatedtopics,seeCRSReportR44225,TheInternetofThings:CRSExperts,coordinatedbyPatriciaMoloneyFigliola.

CongressionalResearchService 2

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

believethatgeneralAIcanbeachievedthroughincrementaldevelopmentandrefiningofcurrentAIandmachinelearningtools,whileothersbelieveitwillrequirethediscoveryanddevelopmentofanewbreakthroughtechnique.

JustasthereisdebateoverthedefinitionofAI,thereisdebateoverwhichtechnologiesshouldbeclassifiedasAI.Forexample,roboticprocessautomation(RPA)hasbeendefinedas“theuseofsoftwaretoautomatehighlyrepetitive,routinetasksnormallyperformedbyknowledgeworkers.”10Becauseitautomatesactivitiesperformedbyhumans,itisoftendescribedasanAItechnology.However,somearguethatRPAisnotAIbecauseitdoesnotincludealearningcomponent.OthersdiscussRPAasabasictoolthatcanbecombinedwithAItocreatecomplexprocessautomation(CPA)orintelligentprocessautomation(IPA),alongan“intelligentautomationcontinuum.”11

AITerminology

Somestakeholders,includingindustry,advocacygroups,andpolicymakers,haveraisedquestionsaboutwhetherspecificAItechnologiesandtechniquesrequiretailoredlegislation.Forexample,legislationenactedinthe116thCongressfocusedongenerativeadversarialnetworks(GANs),describedbelow,whicharethemainunderlyingAItechniqueusedingeneratingdeepfakes,12whicharemostcommonlydescribedasrealisticaudio,video,andotherforgeriescreatedusingAItechniques.13ThissectionismeanttoprovideabroadunderstandingofasubsetofcommontermsusedinthefieldofAIandhowtheyrelatetooneanother.Theseincludethesubfieldofmachinelearning(ML);MLtechniquessuchasdeeplearning,neuralnetworks,andGANs;andtrainingmethodssuchassupervised,unsupervised,andreinforcementlearning.However,justastherearevariationsinhowAIisdefined,researchersandpractitionersdescribevariousAI-relatedtermsinslightlydifferentways.Further,thefollowingtermsandtechniquesarenotmutuallyexclusive;AIsystemsmayemploymorethanone.Forexample,AlphaGo—thefirstAIprogramtobeatahumanmasterattheancientChinesegameofGo—combineddeepneuralnetworks,supervisedlearning,andreinforcementlearning.14

Machinelearning(ML),oftenreferredtoasasubfieldofAI,examineshowtobuildcomputerprogramsthatautomaticallyimprovetheirperformanceatsometaskthroughexperiencewithoutrelyingonexplicitrules-basedprogrammingtodoso.15OneofthegoalsofMListoteachalgorithmstosuccessfullyinterpretdatathathavenotpreviouslybeenencountered.MLisoneofthemostcommonAItechniquesinusetoday,andmostMLtasksarenarrowlyspecifiedtooptimize

SeeIBM,“AutomateRepetitiveTasks,”at/automation/rpa.

IBMGlobalBusinessServices,“UsingArtificialIntelligencetoOptimizetheValueofRoboticProcessAutomation,”September2017,at/downloads/cas/KDKAAK29.

TheIdentifyingOutputsofGenerativeAdversarialNetworks(IOGAN)Act(P.L.116-258).

Foradditionalinformationondeepfakes,seeCRSInFocusIF11333,DeepFakesandNationalSecurity,byKelleyM.SaylerandLaurieA.Harris.

RichardS.SuttonandAndrewG.Barto,ReinforcementLearning:AnIntroduction,2nded.(Cambridge,MA:MITPress,2018),pp.441-442.

AdaptedfromErikBrynjolfsson,TomMitchell,andDanielRock,“WhatCanMachinesLearn,andWhatDoesIt

MeanforOccupationsandtheEconomy?,”AEAPapersandProceedings,vol.108(May1,2018),pp.43-47,at/~tom/pubs/AEA2018-WhatCanMachinesLearn.pdf.MLisdefinedinP.L.116-293tomean

“anapplicationofartificialintelligencethatischaracterizedbyprovidingsystemstheabilitytoautomaticallylearnandimproveonthebasisofdataorexperience,withoutbeingexplicitlyprogrammed.”

CongressionalResearchService 3

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

specificfunctionsusingparticulardatasets.Deeplearning,neuralnetworks,andGANsrepresentafewoftheMLtechniquesfrequentlyusedtoday.

Deeplearning(DL)systemslearnfromlargeamountsofdatatosubsequentlyrecognizeandclassifyrelated,butpreviouslyunobserved,data.Forexample,neuralnetworks,oftendescribedasbeinglooselymodeledafterthehumanbrain,consistofthousandsormillionsofprocessingnodesgenerallyorganizedintolayers.Thestrengthoftheconnectionsamongnodesandlayersarerepeatedlytuned—basedoncharacteristicsofthetrainingdata—tocorrespondtothecorrectoutput.Advancesinhardware,suchasthedevelopmentofgraphicalprocessingunits(GPUs),haveallowedthesenetworkstohavemanylayers,whichiswhatputsthe“deep”indeeplearning.DLapproacheshavebeenusedinsystemsacrossmanyareasofAIresearch,fromautonomousvehiclestovoicerecognitiontechnologies.16

Generativeadversarialnetworks(GANs)consistoftwocompetingneuralnetworks—ageneratornetworkthattriestocreatefakeoutputs(suchaspictures),andadiscriminatornetworkthattriestodeterminewhethertheoutputsarerealorfake.AmajoradvantageofthisstructureisthatGANscanlearnfromlessdatathanotherdeeplearningalgorithms.17AdversarialMLsystemscanbeusedinotherways,aswell;forexample,theycantrytoimprovefairnessinfinancialservicedecisionmakingbyhavingasecondmodeltrytoguesstheprotectedclassofapplicantsbasedonmodelsbuiltbyanothermodel.18

Supervisedlearningalgorithmslearnfromatrainingsetofdatathatislabeledwiththecorrectdescription(e.g.,thecorrectlabelforthispictureis“cat”);thesystemsubsequentlylearnswhichcomponentsofthedataareusefulforclassifyingitcorrectlyandusesthatinformationtocorrectlyclassifydataithasneverencounteredbefore.Incontrast,unsupervisedlearningalgorithmssearchforunderlyingstructuresinunlabeleddata.

Reinforcementlearning(RL)referstogivingcomputerprogramstheabilitytolearnfromexperience,providingthemwithminimalinputsandhumaninterventions.19RLalgorithmslearnbytrialanderror,beingrewardedforreachingspecifiedobjectives—bothforimmediateactionsandlong-termgoals.Theemphasisonsimulatedmotivationandlearningfromdirectinteractionwiththeenvironment,withoutrequiringexplicitexamplesandmodels,areamongthecharacteristicsthatsetRLapartfromotherMLapproaches.20

LarryHardesty,“Explained:NeuralNetworks,”MassachusettsInstituteofTechnology(MIT)News,April14,2017,at/2017/explained-neural-networks-deep-learning-0414.

JamieBeckett,“What’saGenerativeAdversarialNetwork?LeadingResearcherExplains,”NVIDIA,May17,2017,at/blog/2017/05/17/generative-adversarial-network/.

SallyWard-Foxton,“ReducingBiasinAIModelsforCreditandLoanDecision,”EETimes,April30,2019,at/reducing-bias-in-ai-models-for-credit-and-loan-decisions/#.

SeanGarrish,HowSmartMachinesThink(Cambridge,MA:MITPress,2018),p.91.

AdaptedfromRichardS.SuttonandAndrewG.Barto,ReinforcementLearning:AnIntroduction,2nded.(Cambridge,MA:MITPress,2018).

CongressionalResearchService 4

ArtificialIntelligence:Background,SelectedIssues,andPolicyConsiderations

AlgorithmsandAI

AsinterestinAIcontinuestogrow,someanalystsassertthatgeneraldataanalyticsandspecializedalgorithmsareincreasinglybeingreferredto,erroneously,asAI.Itcanbechallengingtomakesuchdistinctionsclearly,giventhevariabilityindefinitionsofAIandrelatedtermsandbecausethesedistinctionshavearguablyshiftedovertime.Forexample,analgorithmisbasicallyaprocedureorsetofinstructionsdesignedtoperformaspecifictaskorsolveamathematicalproblem.SomeearlyproductsofAIresearch,suchasrule-basedexpertsystems,arealgorithmsencodedwithexpertknowledgebutlackingalearningcomponent.Somefeelthatrule-basedsystemsareasimpleformofAIbecausetheysimulateintelligence,whileothersthinkthatwithoutalearningcomponentasystemshouldnotbeconsideredAI.21Generally,however,thegoalsofAI—automatingorreplicatingintelligentbehavior—haveremainedconsistent.22

HistoricalContextofAI

TheideasunderlyingAIanditsconceptualframeworkhavebeenresearchedsinceatleastthe1940sandinitiallyformalizedinthe1950s.IdeasaboutintelligentmachineswerediscussedandpopularizedbyscientistsandauthorssuchasAlanTuringandIsaacAsimov,23andtheterm“artificialintelligence”wascoinedattheDartmouthSummerResearchProjectonArtificialIntelligence,proposedin1955andheldthefollowingyear.24

Sincethattime,thefieldofAIhasgonethroughwhathavebeentermedbysomeassummersandwinters—periodsofmuchresearchandadvancement,followedbylullsinactivityandprogress.ThereasonsfortheAIwintershaveincludedafocusontheoryoverpracticalapplications,researchproblemsbeingmoredifficultthananticipated,andlimitationsofthetechnologiesofthetime.MuchofthecurrentprogressandresearchinAI,whichbeganaround2010,hasbeenattributedtotheavailabilityofbigdata,improvedMLapproachesandalgorithms,andmorepowerfulcomputers.25

WavesofAI

TheDefenseAdvancedResearchProjectsAgency(DARPA),whichhasfundedAIR&Dsincethe1960s,hasdescribedthedevelopmentofAItechnologiesintermsofthreewaves.26Thesewavesaredescribedbythevaryingabilitiesoftechnologiesineachtoperceiverich,complex,andsubtle

Forabriefdiscussionsee,forexample,Tricentis,“AIApproachesCompared:Rule-BasedTestingvs.Learning,”at/artificial-intelligence-software-testing/ai-approaches-rule-based-testing-vs-learning/.

OfficeofScienceandTechnologyPolicy,PreparingfortheFutureofArtificialIntelligence,October2016,pp.5-6.

AlanM.Turing,“ComputingMachineryandIntelligence,”Mind,vol.49(1950),pp.433-460,at/courses/471/papers/turing.pdf;andIsaacAsimov,I,Robot(GardenCity,NY:Doubleday,1950).

SeeJ.McCarthyetal.,“AProposalfortheDartmouthSummerResearchProjectonArtificialIntelligence,”August

31,1955,at/jmc/history/dartmouth/dartmouth.html.

ExecutiveOfficeofthePresident,NationalScienceandTechnologyCouncil,CommitteeonTechnology,PreparingfortheFutureofArtificialIntelligence,October2016,pp.5-6.;foradditionalinformationonthesefactorsandashorthistoryofAI,seealsotheappendixofPeterStoneetal.,“ArtificialIntelligenceandLifein2030,”OneHundredYearStudyonArtificialIntelligence:Reportofthe2015-2016StudyPanel,Stanford

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论