白宮國家人工智能研究資源任務(wù)組發(fā)布最終報(bào)告報(bào)告（2023-01-29）

ArtificialIntelligence(AI)ischangingourcountryandourworld.Fromhowcitizensnavigatetheirdailylivestohowresearchersdrivediscoveriesinthelabtohowmanufacturersbuildproducts,AIisgivingrisetonewcapabilities.NewAIandAI-drivendiscoveriesandcapabilitiesholdthepotentialtodrivepracticalsolutionstoaddresscriticalglobalchallengessuchasfoodproduction,climatechange,poverty,andcancer.tispossibleandcannotaffordtomissoutonseizingtheopportunityforleveragingAItoservethepublicgood.However,theopportunitiestopursuecutting-edgeAIresearchandapplyAItonewdomainsandchallengesarecurrentlynotaccessiblebyallofAmerica'sincredibletalentnorharnessedbythepublicsector.Muchoftoday'sAIresearchreliesonaccesstolargevolumesofdataandadvancedcomputationalpower,whichareoftenunavailabletoresearchersbeyondthoseatwell-resourcedtechnologycompaniesanduniversities.ThisaccessdividelimitstheabilitytoleverageAItotacklethebigchallengesinoursociety.ItalsoconstrainsthediversityofresearchersinthefieldandthebreadthofideasincorporatedintoAIinnovations,contributingtoembeddedbiasesandothersystemicinequalitiesfoundinAIsystemstoday.NationalScienceFoundation(NSF),inconsultationwiththeWhiteHouseOfficeoftoestablishataskforcetocreatearoadmapforaNationalAIResearchResource(NAIRR)—asharedresearchinfrastructurethatwouldprovideAIresearchersandstudentswithsignificantlyexpandedaccesstocomputationalresources,high-qualitydata,educationaltools,andusersupport.ThisfinalreportoftheNAIRRTaskForcepresentsaroadmapandimplementationplanforanationalcyberinfrastructureaimedatovercomingtheaccessdivide,reapingthexperiencesappliedtodevelopingthefutureofAItechnologyanditsroleinoursociety.Suchanationalcyberinfrastructurealsopresentsauniqueandcriticalopportunityto"designin"thestandardsforresponsibleAIresearchpracticesandgovernanceprocessesthatupholdourprioritytodevelopandharnessthesegroundbreakingtechnologiesinamannerthatreinforcesourNation'sdemocraticvaluesandAmericans'personalfreedoms.OSTPandNSFformallylaunchedtheNAIRRTaskForceinJune2021,appointing12leadingexpertsequallyrepresentingacademia,government,andprivateorganizations.Overthecourseofitswork,theTaskForceheld11publicmeetings,engagedwith65expertsonawiderangeofaspectsrelatedtothedesignoftheNAIRR,andconsideredresponsesfromthepublictotworequestsforinformation.WeextendourgratitudetothemembersoftheTaskForcewhohavedonatedanextraordinarynumberofhoursoftheirtimetothiseffort,aswellastothemanymembersofthepublicwhohavecontributedtheirexpertiseandprovidedinputstotheTaskForce.Theresultofthelastoneandone-halfyearsofeffortisthisfinalreport.WeseetheNAIRRasafoundationalinvestmentthatwouldamplifyeffortsacrosstheFederalGovernmenttocultivateAIinnovationandadvancetrustworthyAI.Research,experimentation,andinnovationareintegraltoourprogressasaNation,anditisimperativethatweengagepeoplefromeveryzipcodeandeverybackgroundtoliveuptoAmerica'suniquepromiseofpossibilityandensureourleadershipontheworldstage.TheworkoftheNAIRRTaskForceandthisreportwillbeaninvaluableresourceasweworkcollaborativelyacrossgovernmentandacrosssectorstodrivethisimportantworkforward.SethuramanPanchanathanDirectorScienceFoundationAratiPrabhakarAssistanttothePresidentforScienceandTechnologyDirector,OfficeofScienceandTechnologyPolicyArtificialIntelligence(AI)isanengineofinnovationthatisdrivingscientificdiscoveryandeconomicgrowth.Itisincreasinglybecominganintegralcomponentofsolutionsthatstandtoimpacteverythingfromroutinedailytaskstosocietal-levelchallenges,ultimatelyservingthepublicgood.Atthesametime,therearealsoconcernsthatAIcouldhavenegativesocialandenvironmentalconsequences.TorealizethepositiveandtransformativepotentialofAI,itisimperativetoharnessallofAmerica'singenuitytoadvancethefieldinamannerthataddressessocietalchallenges,worksforallAmericans,andupholdsourdemocraticvalues.YetprogressatthecurrentfrontiersofAIisoftentiedtoaccesstolargeamountsofcomputationalpoweranddata.Suchaccesstodayistoooftenlimitedtothoseinwell-resourcedcommunityandworkforcethatreflectAmerica'srichdiversityandtheabilitytoharnessAItoadvancethepublicgood.AwidelyaccessibleAIresearchcyberinfrastructurethatbringstogethercomputationalresources,data,testbeds,algorithms,software,services,networks,andexpertise,asdescribedinthisreport,wouldhelptodemocratizetheAIresearchanddevelopment(R&D)landscapeintheUnitedStatesforthebenefitofall.ItwouldhelpcreatepathwaystobroadentherangeofresearchersinvolvedinAIandtogrowanddiversifyapproachestoandapplicationsofAIThiscyberinfrastructurecanalsohelptoopenupnewopportunitiesforprogressacrossallscientificfieldsanddisciplines,includingincriticalareassuchasAIauditing,testingandevaluation,trustworthyAI,biasmitigation,andAIsafety.Increasedaccessandadiversityofperspectivescan,inturn,leadtonewideasthatwouldnototherwisematerializeandsettheconditionsfordevelopingAIsystemsthatareinclusivebydesign.NationalArtificialIntelligenceResearchResource(NAIRR)TaskForceto"investigatethefeasibilityandadvisabilityofdeveloping"theNAIRRasanationalAIresearchcyberinfrastructure,and"toproposearoadmapdetailing[howtheNAIRR]shouldbeestablishedandsustained."TherecentCHIPSandScienceActof2022reinforcestheimportanceofdemocratizingaccesstoanationalAIresearchcyberinfrastructure,viainvestmentsthatwillacceleratedevelopmentofadvancedcomputing—fromnext-generationgraphicsprocessingunitstohigh-densitymemorychips—aswellasstepstoactivelyengagebroadanddiverseU.S.talentinfrontierscienceandengineering,includingAI.ThisfinalreportistheculminationoftheTaskForce's18-monthefforttodevelopavisionandimplementationplanforestablishingtheNAIRR.ItbuildsonthefindingsandrecommendationsoutlinedintheTaskForce'sinterimreportreleasedinMay2022,providinganimplementationplantoachievetheobjectiveoftheNAIRR:tostrengthenanddemocratizetheU.S.AIinnovationecosysteminawaythatprotectsprivacy,civilrights,andcivilliberties.vTheNAIRRshouldbeestablishedwithfourmeasurablegoalsinmind,namelyto(1)spurinnovation,(2)increasediversityoftalent,(3)improvecapacity,and(4)advancetrustworthyAI.TheNAIRRshouldmeetthesegoalsbysupportingtheneedsofresearchersandstudentsfromdiversebackgroundswhoarepursuingfoundational,use-inspired,andtranslationalAIresearch.TheseusersshouldbeU.S.-basedoraffiliatedwithU.S.organizations,toincludeacademicinstitutions,non-profitorganizations,andstartupsorsmallbusinesses.TheNAIRRshouldcompriseafederatedsetofcomputational,data,testbed,andsoftwareidersalongwithtechnicalsupportandtrainingtomeettheneedsentationandevaluationoftheNAIRRshouldbecenteredaroundthefourkeygoals,andshouldsupportthecollectionofdataforassessmentofkeyindicatorsofsystemperformanceandsuccessinprogresstowardthesegoals.TheNAIRRadministrationandgovernanceshouldfollowacooperativestewardshipmodel,wherebyasingleFederalagencyservesastheadministrativehomeforNAIRRoperationsandaSteeringCommitteecomprisingprincipalsfromFederalagencieswithequitiesinAIresearchdrivesthestrategicdirectionoftheNAIRR.AProgramManagementOfficewithintheadministrativehomeagencyshouldprovidefundingandoversightforanteeringCommittee,co-chairedbytheNationalAIInitiativeOffice(NAIIO),wouldincorporateinterestsandperspectivesfromacrossFederalagenciesinthegovernanceoftheNAIRR.Theseagenciesshouldalsodirectlysupportresourceproviderswhoseresources,infederation,wouldconstitutethroughaUserCommittee,aScienceAdvisoryBoard,aTechnologyAdvisoryBoard,andanEthicsAdvisoryBoardthatprovideadvicetotheOperatingEntity.TheNAIRRshouldprovideaccesstoafederatedmixofcomputationalanddataresources,testbeds,softwareandtestingtools,andusersupportservicesviaanintegratedportal.Computationalresourcesshouldincludeconventionalservers,computingclusters,high-performancecomputing,andcloudcomputing,andshouldsupportaccesstoedgecomputingresourcesandtestbedsforAIR&D.OpenandprotecteddatashouldbemadeavailableundertedwithcomputationalresourcesTheOperatingEntityshouldnotitselfoperatethetotalityofthecomputerhardwarethatcomposestheNAIRR;instead,computing,alongwithdata,testing,andtrainingresources,shouldbedeliveredasservicesbyiesWhenfullyimplemented,theNAIRRshouldaddressboththecapacity(abilitytosupportalargenumberofusers)andcapability(abilitytotrainresource-intensiveAImodels)needsoftheAIresearchcommunity.TheNAIRRmustbebroadlyaccessibletoarangeofusersandprovideaplatformthatcanbeusedforeducationalandcommunity-buildingactivitiesinordertolowerthebarrierstoparticipationintheAIresearchecosystemandincreasethediversityofAIresearchers.TheNAIRRaccessportalandpublicwebsiteshouldprovidecatalogsandsearchanddiscoverytoolstofacilitateaccesstodata,testbeds,andeducationalandtrainingresourcesservingarangeofexperiencelevels.viimplementationofitsgovernanceprocesses.TheNAIRRmustbeproactiveinaddressingandgovernancemechanismsfromitsoutset.TheOperatingEntityshouldworkwithitsEthicsAdvisoryBoardtodevelopcriteriaandmechanismsforevaluatingproposedresearchandresourcesforinclusionintheNAIRRfromaprivacy,civilrights,andcivillibertiesperspective.andbestpracticesrelatedtoprivacy,civilrights,andcivillibertiesinAIresearch,inaccordancewiththeBlueprintforanAIBillofRightspublishedbytheWhiteHouseOfficeofScienceandTechnologyPolicyinOctober2022.TheNAIRRshouldimplementsystemsafeguardsinaccordancewithestablishedguidelines.TheseguidelinesincludethosedevelopedbytheNationalInstituteofStandardsandTechnology(NIST)andtheFiveSafesframework:safeprojects,safepeople,safesettings,safedata,andsafeoutputs.TheOperatingEntityshoulddesigntheNAIRRcyberinfrastructuretoconsistofmultipletiers,startingwithtwoprimaryzones:anopensciencezone"NAIRR-Open"andasecurezone"NAIRR-Secure."Eachzoneshouldfederatecomputational,network,anddatathezonebutdifferentbetweenzonesreflectingthedifferentprioritiesandneedsoftheusersandresourceoperators.NAIRR-Openshouldadoptthebestpracticesdevelopedovertwodecadesiniciesandmanageaccessusingsinglesignonauthenticationandaresourceallocationenclavesadheringtoacommonsetofsecuritycontrols,andhavetheabilitytosupportsecurityrequirementsarisingfromlegallyprotecteddata.publicationofthisreport.Inphaseone,CongressshouldauthorizeandappropriatefundstoestablishtheNAIRR.TheadministrativehomeagencyandtheNAIIOshouldcoordinatethepreparethesolicitationfortheOperatingEntityandmanagetheselectionprocess.PhasedNAIRRImplementationTimelineInphasetwo,theOperatingEntityshouldestablishitsactivitiesandoverseecreationoftheNAIRRportalanduserinterface,buildinginappropriatetechnicalandpolicycontrols.ThearchitectureshouldsupportcollectionofkeyperformanceindicatorsforevaluationofNAIRRprogress.Resourceprovidersshouldbeselectedviacoordinated,multi-agencyfundingopportunitiesideallyreleasedwithinsixmonthsoftheinitialOperatingEntityaward.viiInphasethree,theNAIRRshouldachieveinitialoperationalcapabilityandtheOperatingEntityshouldalsoformalizethepolicies,processes,andinitialtechnicalresourcestobemademixofcomputationalresourceproviders,(3)anallocationandidentitysystem,and(4)adatapublicationsystem.Inphasefour,activitiesshouldtransitionfrombuildingouttheNAIRRtoestablishingsteady-stateoperations,aswellastheplannedevolutionofNAIRRresourcesinresponsetouseruptakeanddemand.totheAIR&Dcommunity.Asenvisioned,theimpactoftheNAIRRwillbesignificantandfar-reaching,enablingresearcherstotackleproblemsthatrangefromroutinetaskstoglobalchallenges.Inordertoachieveitsvisionandgoals,theTaskForceestimatesthebudgetfortheNAIRRas$2.6billionoveraninitialsix-yearperiod.Thebulkofthisinvestment($2.25billion)istofundtheresourcestobemadeaccessibleviatheNAIRR,throughappropriationstomultipleFederalagencies.TheTaskForceestimatedthisbudgetbasedonrecentcostsofadvancedcomputingresourcesaswellasdata,training,andsoftwareresources;estimatesofusagelevelstomeetthecurrentneedsoftheAIR&Dcommunity;andexpectedgrowthoftheAIR&Dcommunity.Resourceprovidersshouldbebroughtonlineeverytwoyearswithasix-yearlifetime,sothatastmentismadeeverytwoyearstoensurethattheNAIRRresourcesremainstate-of-the-art.TheOperatingEntitywillrequirebetween$55millionand$65millionperyeartosupportthecoordinationandmanagementofNAIRRactivities.Anadditional$5millionperyearisbudgetedforexternalevaluationoftheOperatingEntityandNAIRRperformance.ThevisionfortheNAIRRlaidoutinthisreportisdesignedtomeetthenationalneedforincreasedaccesstothestate-of-the-artresourcesthatfuelAIinnovation.TheroadmapforachievingthisvisionbuildsonexistingFederalinvestments;designsinprotectionsforprivacy,civilrights,andcivilliberties;andpromotesdiversityandequitableaccess.Ifsuccessful,theNationalAIResearchResourcewouldtransformtheU.S.nationalAIresearchecosystemandfacilitatetheabilitytoaddresssocietal-levelproblemsbystrengtheninganddemocratizingparticipationinfoundational,use-inspired,andtranslationalAIR&DintheUnitedStates.viiiuction TheCurrentLandscapeofAIR&D 1AnOpportunityforStrengtheningAIR&DintheUnitedStates 3TheNationalAIResearchResourceTaskForce 4StructureofThisReport 62.ANationalCyberinfrastructuretoDemocratizeandAccelerateAIR&D 7NAIRRVisionandGoals 7TheNAIRRUserBase 8NAIRRConstituents 10Government 10Academia 10Industry 11CivilSociety 113.NAIRROrganization,Management,andGovernance 12NAIRROrganizationalStructure 12SteeringCommittee 14IndividualAgencies 15ProgramManagementOffice 17OperatingEntity 18NAIRRStaffandExecutiveLeadershipTeam 20NAIRRAdvisoryBoards 21EvaluationEntity 22UserAccessandResourceAllocation 22Privacy,CivilRights,andCivilLibertiesProtections 24ScientificIntegrity 26SystemSecurityandUserAccessControls 26Open-SourcePrinciples 28EnvironmentalSustainability 284.NAIRRStructureandSpecificationsforResourceElements 30AccessPortalandUserInterface 30ComputationalResources 31CapacityandCapability 31NAIRRSoftwareResources 31DataandDatasets 32DatasetAcceptanceCriteriaandMetadataStandards 33RoleoftheOperatingEntityinIncentivizingandCuratingContributedDatasetsandOtherResources 33TheNAIRRandExistingFederalGovernmentData 34LegalCompliance 35Co-LocationofResources 36EducationalToolsandServices 36TieredTechnicalTrainingandSupport 37CurationofTrainingMaterials 37PlatformforEducationalActivities 37TechnicalIntegration 37SoftwareforIntegration 37IntegratingDataResources 38Testbeds 385.PhasedBuildoutofNAIRROrganizationandResources 40Phase1:ProgramLaunchandOperatingEntitySelection 41Phase2:OperatingEntityStartup 41InternalPlanningandOperations 41EstablishmentofInitialNAIRRResourceComponents 42Phase3:NAIRRInitialOperationalCapabilities 44InitialComputationalResources 44InitialDataResources 45InitialResearchUsingtheNAIRR 45Phase4:NAIRROngoing(Steady-State)Operations 46EvolutionofNAIRRResources 46PartnershipEngagementOperations 47UserOutreach,Engagement,andSupportOperations 47OutreachandInternationalCollaboration 47NAIRRBudget 48NAIRREvaluation(Phases1–4) 50RoadmapforImplementation 52StepstoInitiatetheNAIRRin2023:ActionsfortheU.S.Government 54ForthePresidentandExecutiveBranchDepartmentsandAgencies 54ForCongress 546.Conclusion 55AppendixA.Definitions A-1AppendixB.DetailsofNAIRRTaskForceEstablishmentandApproachtoRoadmapDevelopment B-1ChargetotheNAIRRTaskForceB-1TaskForceApproachB-2InitialPhaseB-2FinalPhaseB-3AppendixC.BrieferstotheTaskForceC-1AppendixD.PublicInputProvidedontheInterimReportinResponsetotheFederalRequestforInformationD-1AppendixE.PublicInputProvidedontheInitialFederalRequestforInformationonDesigningtheNAIRRE-1AppendixF.SubjectMatterExpertsConsultedbyTaskForceMembersF-1AppendixG.NAIRRPublicListeningSessionG-1AppendixH.NAIRRTaskForceStaffandContributorsH-2AppendixI.ExamplesofNAIRREvaluationMetricsI-1xAppendixJ.DraftLegislativeLanguageforNAIRRAuthorizationJ-1AppendixK.AbbreviationsK-1AppendixL.NotesL-1Figure1.TheCurrentFabricofU.S.ResearchCyberinfrastructure 6Figure2.NAIRRStrategicObjectiveandGoals 8Figure3.AVisionforNAIRRUsersandResourceElements 9Figure4.ProposedNAIRRGovernanceStructure 14Figure5.ProcessforSelectionandIntegrationofNAIRRResourceProviders 16Figure6.ExampleElementsofaTheoryofChangefortheNAIRR 50Figure7.NAIRRImplementationRoadmap 53Table1.NAIRRSix-YearBudgetSummary 48Table2.OperatingEntityCosts 50Table3.RolesinKPIDefinitionandFrequencyofExternalEvaluation 51TableH.1.ExamplesofMetricsthatCouldbeAssociatedwithaNAIRRTheoryofChange I-1TableH.2.ExamplesofMetricsforAssessingProgresstowardNAIRRGoals I-2xiTESSDEBLANC-KNOWLES(CO-CHAIR,BEGINNINGAUGUST2022)SeniorPolicyAdvisor,NationalAIInitiativeOffice,WhiteHouseOfficeofScienceandTechnologyPolicyMANISHPARASHAR(CO-CHAIR,BEGINNINGOCTOBER2021)OfficeDirectoroftheOfficeofAdvancedCyberinfrastructure,NationalScienceFoundationLYNNEPARKER(CO-CHAIR,JULY2021–AUGUST2022)FormerDeputyU.S.ChiefTechnologyOfficerandFoundingDirector,NationalAIInitiativeOffice,WhiteHouseOfficeofScienceandTechnologyPolicyERWINGIANCHANDANI(CO-CHAIR,JULY2021–OCTOBER2021)AssistantDirectorforTechnology,Innovation,andPartnerships,NationalScienceFoundationDANIELABRAGAFounder&CEOofDefined.aiMARKE.DEANORENETZIONICEO,AllenInstituteforAIJULIALANEProfessor,NewYorkUniversityFEI-FEILISequoiaProfessorofComputerScienceatStanfordUniversityandDenningCo-DirectoroftheStanfordInstituteforHuman-CenteredAI(HAI)ANDREWMOOREtGeneralManagerGoogleCloudAIIndustrySolutionsMICHAELL.NORMANDistinguishedProfessor,UniversityofCalifornia,SanDiegoDANSTANZIONEExecutiveDirector,TexasAdvancedComputingCenterAssociateVicePresidentforResearch,TheUniversityofTexasatAustinFREDERICKH.STREITZComputationalScientistLawrenceLivermoreNationalLaboratorySeniorAdvisor,CDCCenterforForecastingandOutbreakAnalyticsELHAMTABASSIofStaffInformationTechnologyLaboratoryNationalInstituteofStandardsandTechnology1.Introduction1TheeconomicandnationalsecurityoftheUnitedStateshaslongreliedonitsuniqueandicalinnovationTheUnitedStatesinvestsinresearchanddevelopment(R&D)acrossscienceandengineeringdisciplinestoadvanceunderstandingofnatural,built,andhumansystemsanddeveloptoolsandmethodsforsolvinglchallengesThisRDleadstodownstreamdevelopmentofapplicationsandcommercialproductsthatdriveeconomicgrowthwhilesupportingthehumanaspirationtoexplore,understand,andimprovetheconditionsofourworld.ndinnovationecosystemasitholdsthepotentialtopowerdiscovery,innovation,andeconomicgrowthacrosseveryfieldofscienceandeverysectoroftheeconomy.However,achievingthispotentialandharnessingAItotackleandworkforcedevelopment.1ItalsorequiresaccesstotheinfrastructurenecessaryforAIexperimentationandtraining.Currently,unevenaccesstotheresourcesthatfuelAIR&DandtraininghavelimitedopportunitiesforresearchersandcontributedtoalackofdiversityintheeoftalentisnotbeingleveragedforthisworkLackofdiversitymayalsocontributetothedevelopmentofbiasedorharmfulAIsystemsandnnovationpotentialandgloballeadershipConcernsrelatedtomisuseofAItrainingresourcesavailabletomoreofAmerica’sresearchersthroughanapproachgroundedinequityandsecuritycanchartapathforward.InthisfutureAmericacanresponsiblyharnessthepotentialofAIforsocietalgoodandeconomicwellbeing—whilealsostrengtheningAmericantechnologicalcompetitivenessfordecadestocome.TheCurrentLandscapeofAIR&DTheterm"ArtificialIntelligence"referstoamachine-basedsystemthatcan,foragivensetencingrealorvirtualenvironments(seeBox1).2Box1.DefinitionofArtificialIntelligence3Box1.DefinitionofArtificialIntelligence3Theterm"artificialintelligence"meansamachine-basedsystemthatcan,foragivensetofhuman-definedobjectives,makepredictions,recommendations,ordecisionsinfluencingrealorvirtualenvironments.Artificialintelligencesystemsusemachineandhuman-basedinputsto:(A)Perceiverealandvirtualenvironments.(B)Abstractsuchperceptionsintomodelsthroughanalysisinanautomatedmanner.(C)Usemodelinferencetoformulateoptionsforinformationoraction.tasksspanningdiverseareas,includingplanningandoptimization,perceptionandvision,modelingandsimulationnaturallanguageunderstanding,roboticprocessautomation,recommendation,andprediction.Thesetaskscanbeaccomplishedthroughstatisticalinferenceextractedfrom"training"[ML])orprogrammedlogical2reasoning(aswithexpertsystems).Today,thecomputationalandstoragecapacityofcomputerageandanalysisoflargequantitiesofdatahasbecomenotonlypossible,butalsoanincreasinglydominantenablerofR&D.ParalleldevelopmentofadvancedsoftwaretoolsandalgorithmshavefacilitatedrealizationofpowerfulanalyticalandpredictivemethodsbasedonAI,whicharebeingappliedbroadlyacrossfieldsofscienceandengineering.AItechnologiesandsustainedinvestmentsincyberinfrastructurehavesupportedscientificdiverseareassuchasproteinfoldingnuclearfusionandevenprogramming.Thebreakthroughsdidnothappenbychance.Theyemergedfromanecosystemtematicinvestmentsincyberinfrastructureeducationandtrainingandlargeandgrowingamountsofdataandcomputationalpower;andtherichcollaborationsbetweenacademicresearchersandtheprivatesector.ThepotentialfortheU.S.researchcommunitytocontributetotheglobalAIresearchandinnovationecosystemisgrowing.Inrecentyears,academiahasseenasignificantgrowthinAIandcomputerscienceresearchymembershavepublishedatleastundergraduatestudents