人工智能外文翻译文献Word文件下载.docx

人工智能外文翻译文献Word文件下载.docx

《人工智能外文翻译文献Word文件下载.docx》由会员分享，可在线阅读，更多相关《人工智能外文翻译文献Word文件下载.docx（16页珍藏版）》请在冰点文库上搜索。

Intelligence》,2015,36（4）:

105-114

字数统计：

英文2887单词，16400字符；

中文5430汉字

外文文献：

ResearchPrioritiesforRobustandBeneficialArtificialIntelligence

Abstract Successinthequestforartificialintelligencehasthepotentialtobringunprecedentedbenefitstohumanity,anditisthereforeworthwhiletoinvestigatehowtomaximizethesebenefitswhileavoidingpotentialpitfalls.Thisarticlegivesnumerousexamples（whichshouldbynomeansbeconstruedasanexhaustivelist）ofsuchworthwhileresearchaimedatensuringthatAIremainsrobustandbeneficial.

Keywords:

artificialintelligence,superintelligence,robust,beneficial,safety,society

Artificialintelligence（AI）researchhasexploredavarietyofproblemsandapproachessinceitsinception,butforthelast20yearsorsohasbeenfocusedontheproblemssurroundingtheconstructionofintelligentagents–systemsthatperceiveandactinsomeenvironment.Inthiscontext,thecriterionforintelligenceisrelatedtostatisticalandeconomicnotionsofrationality–colloquially,theabilitytomakegooddecisions,plans,orinferences.Theadoptionofprobabilisticrepresentationsandstatisticallearningmethodshas led toa largedegree of integrationandcross-fertilization between AI, machine learning, statistics, control theory,neuroscience,andotherfields.Theestablishmentofsharedtheoreticalframeworks,combinedwiththeavailabilityofdataandprocessingpower,hasyieldedremarkable

successesinvariouscomponenttaskssuchasspeechrecognition,imageclassification,autonomousvehicles,machinetranslation,leggedlocomotion,andquestion-answeringsystems.

Ascapabilitiesintheseareasandotherscrossthethresholdfromlaboratoryresearchtoeconomicallyvaluabletechnologies,avirtuouscycletakesholdwherebyevensmallimprovementsinperformanceareworthlargesumsofmoney,promptinggreaterinvestmentsinresearch.ThereisnowabroadconsensusthatAIresearchisprogressingsteadily,andthatitsimpactonsocietyislikelytoincrease.Thepotentialbenefitsarehuge,sinceeverythingthatcivilizationhastoofferisaproductofhumanintelligence;

wecannotpredictwhatwemightachievewhenthisintelligenceismagnifiedbythetoolsAImayprovide,buttheeradicationofdiseaseandpovertyarenotunfathomable.BecauseofthegreatpotentialofAI,itisvaluabletoinvestigatehowtoreapitsbenefitswhileavoidingpotentialpitfalls.

Short-termResearchPrioritiesOptimizingAI’sEconomicImpact

The successesofindustrialapplicationsofAI,frommanufacturingtoinformationservices,demonstrateagrowingimpactontheeconomy,althoughthereisdisagreementabouttheexactnatureofthisimpactandonhowtodistinguishbetweentheeffectsofAIandthoseofotherinformationtechnologies.ManyeconomistsandcomputerscientistsagreethatthereisvaluableresearchtobedoneonhowtomaximizetheeconomicbenefitsofAIwhilemitigatingadverseeffects,whichcouldincludeincreasedinequalityandunemployment（Mokyr2014;

BrynjolfssonandMcAfee2014;

FreyandOsborne2013;

Glaeser2014;

Shanahan2015;

Nilsson1984;

Manyikaetal.2013）.Suchconsiderationsmotivatearangeofresearchdirections,spanningareasfromeconomicstopsychology.Belowareafewexamplesthatshouldbynomeansbeinterpretedasanexhaustivelist.

Labormarketforecasting:

Whenandinwhatordershouldweexpectvariousjobstobecomeautomated（FreyandOsborne2013）?

Howwillthisaffectthewagesoflessskilledworkers,thecreativeprofessions,anddifferentkindsofinformation

workers?

SomehavehavearguedthatAIislikelytogreatlyincreasetheoverallwealthofhumanityasawhole（BrynjolfssonandMcAfee2014）.However,increasedautomationmaypushincomedistributionfurthertowardsapowerlaw（Brynjolfsson,McAfee,andSpence2014）,andtheresultingdisparitymayfalldisproportionatelyalonglinesofrace,class,andgender;

researchanticipatingtheeconomicandsocietalimpactofsuchdisparitycouldbeuseful.

Othermarketdisruptions:

Significantpartsoftheeconomy,includingfinance,insurance,actuarial,andmanyconsumermarkets,couldbesusceptibletodisruptionthroughtheuseofAItechniquestolearn,model,andpredicthumanandmarketbehaviors.Thesemarketsmightbeidentifiedbyacombinationofhighcomplexityandhighrewardsfornavigatingthatcomplexity（Manyikaetal.2013）.

Policyformanagingadverseeffects:

Whatpoliciescouldhelpincreasinglyautomatedsocietiesflourish?

Forexample,BrynjolfssonandMcAfee（BrynjolfssonandMcAfee2014）explorevariouspoliciesforincentivizingdevelopmentoflabor-intensivesectorsandforusingAI-generatedwealthtosupportunderemployedpopulations.Whataretheprosandconsofinterventionssuchaseducationalreform,apprenticeshipprograms,labor-demandinginfrastructureprojects,andchangestominimumwagelaw,taxstructure,andthesocialsafetynet（Glaeser2014）?

Historyprovidesmanyexamplesofsubpopulationsnotneedingtoworkforeconomicsecurity,rangingfromaristocratsinantiquitytomanypresent-daycitizensofQatar.Whatsocietalstructuresandotherfactorsdeterminewhethersuchpopulationsflourish?

Unemploymentisnotthesameasleisure,andtherearedeeplinksbetweenunemploymentandunhappiness,self-doubt,andisolation（Hetschko,Knabe,andScho¨

b2014;

ClarkandOswald1994）;

understandingwhatpoliciesandnormscanbreaktheselinkscouldsignificantlyimprovethemedianqualityoflife.Empiricalandtheoreticalresearchontopicssuchasthebasicincomeproposalcouldclarifyouroptions（VanParijs1992;

Widerquistetal.2013）.

Economicmeasures:

ItispossiblethateconomicmeasuressuchasrealGDPpercapitadonotaccuratelycapturethebenefitsanddetrimentsofheavilyAI-and-automation-basedeconomies,makingthesemetricsunsuitableforpolicy

purposes（Mokyr2014）.Researchonimprovedmetricscouldbeusefulfordecision-making.

LawandEthicsResearch

ThedevelopmentofsystemsthatembodysignificantamountsofintelligenceandautonomyleadstoimportantlegalandethicalquestionswhoseanswersimpactbothproducersandconsumersofAItechnology.Thesequestionsspanlaw,publicpolicy,professionalethics,andphilosophicalethics,andwillrequireexpertisefromcomputerscientists,legalexperts,politicalscientists,andethicists.Forexample:

Liabilityandlawforautonomousvehicles:

Ifself-drivingcarscuttheroughly40,000annualUStrafficfatalitiesinhalf,thecarmakersmightgetnot20,000thank-younotes,but20,000lawsuits.Inwhatlegalframeworkcanthesafetybenefitsofautonomousvehiclessuchasdroneaircraftandself-drivingcarsbestberealized（Vladeck2014）?

ShouldlegalquestionsaboutAIbehandledbyexisting（software-andinternet-focused）‘‘cyberlaw’’,orshouldtheybetreatedseparately（Calo2014b）?

Inbothmilitaryandcommercialapplications,governmentswillneedtodecidehowbesttobringtherelevantexpertisetobear;

forexample,apanelorcommitteeofprofessionalsandacademicscouldbecreated,andCalohasproposedthecreationofaFederalRoboticsCommission（Calo2014a）.

Machineethics:

Howshouldanautonomousvehicletradeoff,say,asmallprobabilityofinjurytoahumanagainstthenear-certaintyofalargematerialcost?

Howshouldlawyers,ethicists,andpolicymakersengagethepublicontheseissues?

Shouldsuchtrade-offsbethesubjectofnationalstandards?

Autonomousweapons:

Canlethalautonomousweaponsbemadetocomplywithhumanitarianlaw（ChurchillandUlfstein2000）?

If,assomeorganizationshavesuggested,autonomousweaponsshouldbebanned（Docherty2012）,isitpossibletodevelopaprecisedefinitionofautonomyforthispurpose,andcansuchabanpracticallybeenforced?

Ifitispermissibleorlegaltouselethalautonomousweapons,howshouldtheseweaponsbeintegratedintotheexistingcommand-and-controlstructuresothatresponsibilityandliabilityremainassociatedwithspecifichumanactors?

Whattechnicalrealitiesandforecastsshouldinformthesequestions,andhow

should‘‘meaningfulhumancontrol’’overweaponsbedefined（Roff2013,2014;

Anderson,Reisner,andWaxman2014）?

Areautonomousweaponslikelytoreducepoliticalaversiontoconflict,orperhapsresultin‘‘accidental’’battlesorwars（Asaro2008）?

Wouldsuchweaponsbecomethetoolofchoiceforoppressorsorterrorists?

Finally,howcantransparencyandpublicdiscoursebestbeencouragedontheseissues?

Privacy:

HowshouldtheabilityofAIsystemstointerpretthedataobtainedfromsurveillancecameras,phonelines,emails,etc.,interactwiththerighttoprivacy?

Howwillprivacyrisksinteractwithcybersecurityandcyberwarfare（SingerandFriedman2014）?

OurabilitytotakefulladvantageofthesynergybetweenAIandbigdatawilldependinpartonourabilitytomanageandpreserveprivacy（Manyikaetal.2011;

AgrawalandSrikant2000）.

Professionalethics:

WhatroleshouldcomputerscientistsplayinthelawandethicsofAIdevelopmentanduse?

PastandcurrentprojectstoexplorethesequestionsincludetheAAAI2008–09PresidentialPanelonLong-TermAIFutures（HorvitzandSelman2009）,theEPSRCPrinciplesofRobotics（Bodenetal.2011）,andrecentlyannouncedprogramssuchasStanford’sOne-HundredYearStudyofAIandtheAAAICommitteeonAIImpactandEthicalIssues.

Long-termresearchpriorities

Afrequentlydiscussedlong-termgoalofsomeAIresearchersistodevelopsystemsthatcanlearnfromexperiencewithhuman-likebre