AI前沿筆記：來自數(shù)百個應(yīng)用案例的見解

NOTESFROMTHEAIFRONTIERINSIGHTSFROMHUNDREDSOFUSECASES

DISCUSSIONPAPER

APRIL2018

MichaelChui|SanFranciscoJamesManyika|SanFranciscoMehdiMiremadi|ChicagoNicolausHenke|London

RitaChung|SiliconValleyPieterNel|NewYorkSankalpMalhotra|NewYork

Sinceitsfoundingin1990,theMcKinseyGlobalInstitute(MGI)hassoughttodevelopadeeperunderstandingoftheevolvingglobaleconomy.AsthebusinessandeconomicsresearcharmofMcKinsey&Company,MGIaimstoprovideleadersinthecommercial,public,andsocialsectorswiththefactsandinsightsonwhichtobasemanagementandpolicydecisions.

MGIresearchcombinesthedisciplinesofeconomicsandmanagement,employingtheanalyticaltoolsofeconomicswiththeinsightsofbusinessleaders.Our“micro-to-macro”methodologyexaminesmicroeconomicindustrytrendstobetterunderstandthebroadmacroeconomicforcesaffectingbusinessstrategyandpublicpolicy.MGI’sin-depthreportshavecoveredmorethan20countriesand30industries.Currentresearchfocusesonsixthemes:productivityandgrowth,naturalresources,labormarkets,theevolutionofglobalfinancialmarkets,theeconomicimpactoftechnologyandinnovation,andurbanization.Recentreportshaveassessedthedigitaleconomy,theimpactofAIandautomationonemployment,incomeinequality,theproductivitypuzzle,theeconomicbenefitsoftacklinggenderinequality,aneweraofglobalcompetition,Chineseinnovation,anddigitalandfinancialglobalization.

MGIisledbythreeMcKinsey&Companyseniorpartners:JacquesBughin,JonathanWoetzel,andJamesManyika,whoalsoservesasthechairmanofMGI.MichaelChui,SusanLund,AnuMadgavkar,JanMischke,SreeRamaswamy,andJaanaRemesareMGIpartners,andMekalaKrishnanandJeongminSeongareMGIseniorfellows.

ProjectteamsareledbytheMGIpartnersandagroupofseniorfellows,andincludeconsultantsfromMcKinseyofficesaroundtheworld.TheseteamsdrawonMcKinsey’sglobalnetworkofpartnersandindustryandmanagementexperts.AdviceandinputtoMGIresearchareprovidedbytheMGICouncil,membersofwhicharealsoinvolvedinMGI’sresearch.MGICouncilmembersaredrawnfromaroundtheworldandfromvarioussectorsandincludeAndrésCadena,SandrineDevillard,RichardDobbs,TarekElmasry,KatyGeorge,RajatGupta,EricHazan,EricLabaye,AchaLeke,ScottNyquist,

GaryPinkus,SvenSmit,OliverTonby,andEckartWindhagen.Inaddition,leadingeconomists,includingNobellaureates,actasresearchadviserstoMGIresearch.

ThepartnersofMcKinseyfundMGI’sresearch;itisnotcommissionedbyanybusiness,government,orotherinstitution.ForfurtherinformationaboutMGIandtodownloadreports,pleasevisit

/mgi

.

MCKINSEYANALYTICS

McKinseyAnalyticshelpsclientsachievebetterperformancethroughdata.Weworktogetherwithclientstobuildanalytics-drivenorganizations,helpingthemdevelopthestrategies,operations,andcapabilitiestoderiverapidandsustainedimpactfromanalytics.Overthepastfiveyears,wehaveworkedwithmorethan2,000clientsacrosseveryindustryandbusinessfunction.McKinseyAnalyticsisledgloballybyNicolausHenkeandNoshirKaka,togetherwithanexecutivecommitteecomprisedof40McKinseyseniorpartnersrepresentingallregionsandpractices.Today,McKinseyAnalytics

bringstogethermorethan1,900advancedanalyticsandAIexpertsandspansmorethan125domains(industry-andfunction-specificteamswithpeople,data,andtoolsfocusedonuniqueapplicationsofanalytics).McKinseyAnalyticsincludesseveralacquiredcompaniessuchasQuantumBlack,aleadingadvancedanalyticsfirmthatMcKinseyacquiredin2015.

Learnmoreat

/business-functions/mckinsey-analytics/our-insights.

Copyright?McKinsey&Company2018

INBRIEF

NOTESFROMTHEAIFRONTIER:

INSIGHTSFROMHUNDREDSOFUSECASES

Forthisdiscussionpaper,partofourongoingresearchintoevolvingtechnologiesandtheireffectonbusiness,economies,andsociety,wemappedtraditionalanalyticsandnewer“deeplearning”techniquesandtheproblemstheycansolvetomorethan400specific

usecasesincompaniesandorganizations.DrawingonMGIresearchandtheappliedexperiencewithartificialintelligence(AI)ofMcKinseyAnalytics,weassessboththepracticalapplicationsandtheeconomicpotentialofadvancedAItechniquesacrossindustriesandbusinessfunctions.WecontinuetostudytheseAItechniquesandadditionalusecases.Fornow,hereareourkeyfindings:

AI,whichforthepurposesofthispaperwecharacterizeas“deeplearning”techniquesusingartificialneuralnetworks,canbeusedtosolveavarietyofproblems.Techniquesthataddressclassification,estimation,andclusteringproblemsarecurrentlythemostwidelyapplicableintheusecaseswehaveidentified,reflectingtheproblemswhosesolutionsdrivevalueacrosstherangeofsectors.

ThegreatestpotentialforAIwehavefoundistocreatevalueinusecasesinwhichmoreestablishedanalyticaltechniquessuchasregressionandclassificationtechniques

canalreadybeused,butwhereneuralnetworktechniquescouldprovidehigherperformanceorgenerateadditionalinsightsandapplications.Thisistruefor69percentoftheAIusecasesidentifiedinourstudy.Inonly16percentofusecasesdidwefinda“greenfield”AIsolutionthatwasapplicablewhereotheranalyticsmethodswouldnotbeeffective.

BecauseofthewideapplicabilityofAIacrosstheeconomy,thetypesofusecaseswiththegreatestvaluepotentialvarybysector.Thesevariationsprimarilyresultfromtherelativeimportanceofdifferentdriversofvaluewithineachsector.Theyarealsoaffectedbytheavailabilityofdata,itssuitabilityforavailabletechniques,andtheapplicabilityofvarioustechniquesandalgorithmicsolutions.Inconsumer-facingindustriessuchasretail,forexample,marketingandsalesistheareawiththemostvalue.Inindustriessuchasadvancedmanufacturing,inwhichoperationalperformancedrivescorporateperformance,thegreatestpotentialisinsupplychain,logistics,andmanufacturing.

Thedeeplearningtechniquesonwhichwefocused—feedforwardneuralnetworks,recurrentneuralnetworks,andconvolutionalneuralnetworks—accountforabout

40percentoftheannualvaluepotentiallycreatedbyallanalyticstechniques.Thesethreetechniquestogethercanpotentiallyenablethecreationofbetween$3.5trillionand

$5.8trillioninvalueannually.Withinindustries,thatistheequivalentof1to9percentof2016revenue.

Capturingthepotentialimpactofthesetechniquesrequiressolvingmultipleproblems.Technicallimitationsincludetheneedforalargevolumeandvarietyofoftenlabeledtrainingdata,althoughcontinuedadvancesarealreadyhelpingaddressthese.Tougherperhapsmaybethereadinessandcapabilitychallengesforsomeorganizations.Societalconcernandregulation,forexampleaboutprivacyanduseofpersonaldata,canalsoconstrainAIuseinbanking,insurance,healthcare,andpharmaceuticalandmedicalproducts,aswellasinthepublicandsocialsectors,iftheseissuesarenotproperlyaddressed.

Thescaleofthepotentialeconomicandsocietalimpactcreatesanimperativeforalltheparticipants—AIinnovators,AI-usingcompaniesandpolicy-makers—toensureavibrantAIenvironmentthatcaneffectivelyandsafelycapturetheeconomicandsocietalbenefits.

McKinseyGlobalInstitute

NotesfromtheAIfrontier:Insightsfromhundredsofusecases

PAGE

11

PAGE

2

McKinseyGlobalInstitute

1.MappingAItechniquestoproblemtypes

What’sinside

Introduction

Page1

MappingAItechniques

toproblemtypes

Page2

Insightsfrom

usecases

Page7

Sizingthepotential

valueofAI

Page17

Theroadtoimpact

andvalue

Page26

Acknowledgments

Page31

INTRODUCTION

Artificialintelligence(AI)standsoutasatransformationaltechnologyofourdigitalage.Questionsaboutwhatitis,whatitcanalreadydo—andwhatithasthepotentialtobecome—cutacrosstechnology,psychology,politics,economics,sciencefiction,law,andethics.AIisthesubjectofcountlessdiscussionsandarticles,fromtreatisesabouttechnicaladvancestotabloidheadlinesaboutitseffects.Evenasthedebatecontinues,thetechnologiesunderpinningAIcontinuetomoveforward,enablingapplicationsfromfacialrecognitioninsmartphonestoconsumerappsthatuseAIalgorithmstodetectdiabetesandhypertensionwithincreasingaccuracy.1Indeed,whilemuchofthepublicdiscussionofAIfocusesonsciencefiction-likeAIrealizationsuchasrobots,thenumberofless-noticedpracticalapplicationsforAIthroughouttheeconomyisgrowingapaceandpermeatingourlives.

ThisdiscussionpaperseekstocontributetothebodyofknowledgeaboutAIbymappingAItechniquestothe

typesofproblemstheycanhelpsolveandthenmappingtheseproblemtypestomorethan400practicalusecasesandapplicationsinbusinessesacross19industries,fromaerospaceanddefensetotravelandthepublicsector,and

ninebusinessfunctionsrangingfrommarketingandsalesandsupply-chainmanagementtoproductdevelopmentandhumanresources.2Drawingonawidevarietyofpublic

andproprietarydatasources,includingtheexperiencesofourMcKinsey&Companycolleagues,wealsoassessthepotentialeconomicvalueofthelatestgenerationsofAItechnologies.TheAItechniqueswefocusonaredeeplearningtechniquesbasedonartificialneuralnetworks,whichweseeasgeneratingasmuchas40percentofthetotalpotentialvaluethatallanalyticstechniquescouldprovide.

Ourfindingshighlightthesubstantialpotentialofapplyingdeeplearningtechniquestousecasesacrosstheeconomy;thesetechniquescanprovideanincrementalliftbeyondthatfrommoretraditionalanalyticstechniques.Weidentifytheindustriesandbusinessfunctionsinwhichthereisvaluetobecaptured,andweestimatehowlargethatvaluecouldbeglobally.Forallthepotential,muchworkneedstobedonetoovercomearangeoflimitationsandobstaclestoAIapplication.Weconcludewithabriefdiscussionofthese

obstaclesandoffutureopportunitiesasthetechnologiescontinuetheiradvance.Ultimately,thevalueofAIisnottobefoundinthemodelsthemselves,butinorganizations’abilitiestoharnessthem.Businessleaderswillneedtoprioritizeandmakecarefulchoicesabouthow,when,andwheretodeploythem.

Thispaperispartofourcontinuingresearchintoanalytics,automation,andAItechnologies,andtheireffectonbusiness,theeconomy,andsociety.3ItisnotintendedtoserveasacomprehensiveguidetodeployingAI;forexample,weidentifybutdonotelaborateonissuesofdatastrategy,dataengineering,governance,orchangemanagementandculture

1GeoffreyH.Tisonetal.,“Cardiovascularriskstratificationusingoff-the-shelfwearablesandamulti-maskdeeplearningalgorithm,”Circulation,volume136,supplement1,November14,2017.

2Wedonotidentifythecompaniesbynameorcountry,forreasonsofclientconfidentiality.

3PreviousMcKinseyGlobalInstitutereportsontheseissuesincludeTheageofanalytics:Competinginadata-drivenworld,December2016;Afuturethatworks:Automation,employmentandproductivity,January2017;andArtificialintelligence:Thenextdigitalfrontier?June2017.Seealistofourrelatedresearchattheendofthispaper.

thatarevitalforcompaniesseekingtocapturevaluefromAIandanalytics.4Theusecasesweexaminedarenotexhaustive;indeed,wecontinuetoidentifyandexamineothers,andwemayupdateourfindingsinduecourse.Nonetheless,webelievethatthisresearchcanmakeausefulcontributiontoourunderstandingofwhatAIcanandcan’t(yet)do,andhowmuchvaluecouldbederivedfromitsuse.Itisimportanttohighlightthat,evenasweseeeconomicpotentialintheuseofAItechniques,theuseofdatamustalwaystakeintoaccountconcernsincludingdatasecurity,privacy,andpotentialissuesofbias,issueswehaveaddressedelsewhere.5

MAPPINGAITECHNIQUESTOPROBLEMTYPES

Asartificialintelligencetechnologiesadvance,sodoesthedefinitionofwhichtechniquesconstituteAI(seeBox1,“Deeplearning’soriginsandpioneers”).6Forthepurposesofthispaper,weuseAIasshorthandspecificallytorefertodeeplearningtechniquesthatuseartificialneuralnetworks.Inthissection,wedefinearangeofAIandadvancedanalyticstechniquesaswellaskeyproblemtypestowhichthesetechniquescanbeapplied.

NEURALNETWORKSANDOTHERMACHINELEARNINGTECHNIQUES

Welookedatthevaluepotentialofarangeofanalyticstechniques.Thefocusofourresearchwasonmethodsusingartificialneuralnetworksfordeeplearning,whichwecollectivelyrefertoasAIinthispaper,understandingthatindifferenttimesandcontexts,othertechniquescanandhavebeenincludedinAI.Wealsoexaminedothermachinelearningtechniquesandtraditionalanalyticstechniques(Exhibit1).WefocusedonspecificpotentialapplicationsofAIinbusinessandthepublicsector(sometimesdescribed

as“artificialnarrowAI”)ratherthanthelonger-termpossibilityofan“artificialgeneralintelligence”thatcouldpotentiallyperformanyintellectualtaskahumanbeingiscapableof.

Exhibit1

Artificialintelligence,machinelearning,andotheranalyticstechniquesthatweexaminedforthisresearch

Techniques ConsideredAIforourresearch

LikelihoodtobeusedinAIapplications

LessMore

Advancedtechniques

Deeplearningneuralnetworks(e.g.,feedforwardneuralnetworks,CNNs,RNNs,GANs)

Instancebased(e.g.,KNN)

Dimensionalityreduction(e.g.,PCA,tSNE) Ensemblelearning(e.g.,random

forest,gradientboosting)

Decisiontreelearning

MonteCarlo Linearclassifiers(e.g.,Fisher’smethods lineardiscriminant,SVM)

Statisticalinference(e.g.,Bayesianinference,ANOVA)

Clustering(e.g.,k-means,treebased,dbscan)

Markovprocess RegressionAnalysis(e.g.,(e.g.,Markovchain) linear,logistic,lasso)

Descriptivestatistics(e.g.,confidenceinterval)

Traditionaltechniques

NaiveBayesclassifier

Reinforcementlearning

Transferlearning

SOURCE:McKinseyGlobalInstituteanalysis

4SeeJacquesBughin,BrianMcCarthy,andMichaelChui,“Asurveyof3,000executivesrevealshowbusinessessucceedwithAI,”HarvardBusinessReview,August28,2017.

5MichaelChui,JamesManyika,andMehdiMiremadi,“WhatAIcanandcan’tdo(yet)foryourbusiness,”

McKinseyQuarterly,January2018.

6ForadetailedlookatAItechniques,seeAnexecutive’sguidetoAI,McKinseyAnalytics,January2018.

/business-functions/mckinsey-analytics/our-insights/an-executives-guide-to-ai

Box1:Deeplearning’soriginsandpioneers

Itistooearlytowriteafullhistoryofdeeplearning—andsomeofthedetailsarecontested—butwecanalreadytraceanadmittedlyincompleteoutlineofitsoriginsandidentifysomeofthepioneers.TheyincludeWarrenMcCullochandWalterPitts,whoasearlyas1943proposed

anartificialneuron,acomputationalmodelofthe“nervenet”inthebrain.1BernardWidrowandTedHoffatStanfordUniversity,developedaneuralnetworkapplicationbyreducingnoiseinphonelinesinthelate1950s.2Aroundthesametime,FrankRosenblatt,anAmericanpsychologist,introducedtheideaofadevicecalledthePerceptron,whichmimickedthe

neuralstructureofthebrainandshowedanabilitytolearn.3MIT’sMarvinMinskyandSeymourPapertthenputadamperonthisresearchintheir1969book“Perceptrons”,byshowingmathematicallythatthePerceptroncouldonlyperformverybasictasks.4Theirbookalsodiscussedthedifficultyoftrainingmulti-layerneuralnetworks.In1986,GeoffreyHintonattheUniversityofToronto,alongwithcolleaguesDavidRumelhartandRonaldWilliams,solvedthistrainingproblemwiththepublicationofanowfamousbackpropagationtrainingalgorithm—althoughsomepractitionerspointtoaFinnishmathematician,SeppoLinnainmaa,ashavinginventedbackpropagationalreadyinthe1960s.5YannLeCunatNYUpioneeredtheuse

ofneuralnetworksonimagerecognitiontasksandhis1998paperdefinedtheconceptofconvolutionalneuralnetworks,whichmimicthehumanvisualcortex.6Inparallel,JohnHopfieldpopularizedthe“Hopfield”networkwhichwasthefirstrecurrentneuralnetwork.7ThiswassubsequentlyexpandeduponbyJurgenSchmidhuberandSeppHochreiterin1997with

theintroductionofthelongshort-termmemory(LSTM),greatlyimprovingtheefficiencyandpracticalityofrecurrentneuralnetworks.8Hintonandtwoofhisstudentsin2012highlightedthepowerofdeeplearningwhentheyobtainedsignificantresultsinthewell-knownImageNetcompetition,basedonadatasetcollatedbyFei-FeiLiandothers.9Atthesametime,JeffreyDeanandAndrewNgweredoingbreakthroughworkonlargescaleimagerecognitionatGoogleBrain.10Deeplearningalsoenhancedtheexistingfieldofreinforcementlearning,ledbyresearcherssuchasRichardSutton,leadingtothegame-playingsuccessesofsystemsdevelopedbyDeepMind.11In2014,IanGoodfellowpublishedhispaperongenerativeadversarialnetworks,whichalongwithreinforcementlearninghasbecomethefocusofmuchoftherecentresearchinthefield.12ContinuingadvancesinAIcapabilitieshaveledtoStanford

University’sOneHundredYearStudyonArtificialIntelligence,foundedbyEricHorvitz,buildingonthelong-standingresearchheandhiscolleagueshaveledatMicrosoftResearch.Wehavebenefitedfromtheinputandguidanceofmanyofthesepioneersinourresearchoverthepastfewyears.

1WarrenMcCullochandWalterPitts,“Alogicalcalculusoftheideasimmanentinnervousactivity,”BulletinofMathematicalBiophysics,volume5,1943.

2AndrewGoldstein,“BernardWidroworalhistory,”IEEEGlobalHistoryNetwork,1997.

3FrankRosenblatt,“ThePerceptron:Aprobabilisticmodelforinformationstorageandorganizationinthebrain,”

Psychologicalreview,volume65,number6,1958.

4MarvinMinskyandSeymourA.Papert,Perceptrons:Anintroductiontocomputationalgeometry,MITPress,January1969.

5DavidE.Rumelhart,GeoffreyE.Hinton,andRonaldJ.Williams,“Learningrepresentationsbyback-propagatingerrors,”Nature,volume323,October1986;foradiscussionofLinnainmaa’sroleseeJuergenSchmidhuber,Whoinventedbackpropagation?,Blogpost

http://people.idsia.ch/~juergen/who-invented-backpropagation.html,

2014.

6YannLeCun,PatrickHaffner,LeonBotton,andYoshuaBengio,Objectrecognitionwithgradient-basedlearning,ProceedingsoftheIEEE,November1998.

7JohnHopfield,Neuralnetworkdsandphysicalsystemswithemergentcollectivecomputationalabilities,PNAS,April1982.

8SeppHochreiterandJuergenSchmidhuber,“Longshort-termmemory,”NeuralComputation,volume9,number8,December1997.

9AlexKrizhevsky,IlyaSutskever,andGeoffreyE.Hinton,ImageNetclassificationwithdeepconvolutionalneuralnetworks,NIPS12proceedingsofthe25thInternationalConferenceonNeuralInformationProcessingSystems,December2012.

10JeffreyDeanetal.,Largescaledistributeddeepnetworks,NIPS2012.

11RichardS.SuttonandAndrewG.Barto,Reinforcementlearning:Anintroduction,MITPress,1998.

12IanJ.Goodfellow,Generativeadversarialnetworks,ArXiv,June2014.

Neuralnetworksareasubsetofmachinelearningtechniques.Essentially,theyareAIsystemsbasedonsimulatingconnected“neuralunits,”looselymodelingthewaythatneuronsinteractinthebrain.Computationalmodelsinspiredbyneuralconnectionshavebeenstudiedsincethe1940sandhavereturnedtoprominenceascomputerprocessingpowerhasincreasedandlargetrainingdatasetshavebeenusedtosuccessfullyanalyzeinputdatasuchasimages,video,andspeech.AIpractitionersrefertothesetechniquesas“deeplearning,”sinceneuralnetworkshavemany(“deep”)layersofsimulatedinterconnectedneurons.Beforedeeplearning,neuralnetworksoftenhadonlythreetofive

layersanddozensofneurons;deeplearningnetworkscanhaveseventotenormorelayers,withsimulatedneuronsnumberingintothemillions.

Inthispaper,weanalyzedtheapplicationsandvalueofthreeneuralnetworktechniques:

Feedforwardneuralnetworks.Oneofthemostcommontypesofartificialneuralnetwork.Inthisarchitecture,informationmovesinonlyonedirection,forward,fromtheinputlayer,throughthe“hidden”layers,totheoutputlayer.Therearenoloopsinthenetwork.Thefirstsingle-neuronnetworkwasproposedin1958byAIpioneerFrankRosenblatt.Whiletheideaisnotnew,advancesincomputingpower,trainingalgorithms,andavailabledataledtohigherlevelsofperformancethanpreviouslypossible.

Recurrentneuralnetworks(RNNs).Artificialneuralnetworkswhoseconnectionsbetweenneuronsincludeloops,well-suitedforprocessingsequencesofinputs,whichmakesthemhighlyeffectiveinawiderangeofapplications,fromhandwriting,totexts,tospeechrecognition.InNovember2016,OxfordUniversityresearchersreportedthatasystembasedonrecurrentneuralnetworks(andconvolutionalneuralnetworks)hadachieved95percentaccuracyinreadinglips,outperformingexperiencedhumanlipreaders,whotestedat52percentaccuracy.

Convolutionalneuralnetworks(CNNs).Artificialneuralnetworksinwhichtheconnectionsbetweenneurallayersareinspiredbytheorganizationoftheanimalvisualcortex,theportionofthebrainthatprocessesimages,wellsuitedforvisualperceptiontasks.

Weestimatedthepotentialofthosethreedeepneuralnetworktechniquestocreatevalue,aswellasothermachinelearningtechniquessuchastree-basedensemblelearning,classifiers,andclustering,andtraditionalanalyticssuchasdimensionalityreductionandregression.

Forourusecases,wealsoconsideredtwoothertechniques—generativeadversarialnetworks(GANs)andreinforcementlearning—butdidnotincludetheminourpotentialvalueassessmentofAI,sincetheyremainnascenttechniquesthatarenotyetwidelyappliedinbusinesscontexts.However,aswenoteintheconcludingsectionofthispaper,theymayhaveconsiderablerelevanceinthefuture.

Generativeadversarialnetworks(GANs).Theseusuallyusetwoneuralnetworkscontestingeachotherinazero-sumgameframework(thus“adversarial”).GANscanlearntomimicvariousdistributionsofdata(forexampletext,speech,andimages)andarethereforevaluableingeneratingtestdatasetswhenthesearenotreadilyavailable.

Reinforcementlearning.Thisisasubfieldofmachinelearninginwhichsystemsaretrainedbyreceivingvirtual“rewards”or“punishments,”essentiallylearningbytrialanderror.GoogleDeepMindhasusedreinforcementlearningtodevelopsystemsthatcanplaygames,includingvideogamesandboardgamessuchasGo,betterthanhumanchampions.

PROBLEMTYPESANDTHEANALYTICTECHNIQUESTHATCANBEAPPLIEDTOSOLVETHEM

Inabusinesssetting,thoseanalytictechniquescanbeappliedtosolvereal-lifeproblems.Forthisresearch,wecreatedataxonomyofhigh-levelproblemtypes,characterizedbytheinputs,outputs,andpurposeofeach.Acorrespondingsetofanalytictechniquescanbeappliedtosolvetheseproblems.Theseproblemtypesinclude:

Classification.Basedonasetoftrainingdata,categorizenewinputsasbelongingtooneofasetofcategories.Anexampleofclassificationisidentifyingwhetheranimagecontainsaspecifictypeofobject,suchasatruckoracar,oraproductofacceptablequalitycomingfromamanufacturingline.

Continuousestimation.Basedonasetoftrainingdata,estimatethenextnumericvalueinasequence.Thistypeofproblemissometimesdescribedas“prediction,”particularlywhenitisappliedtotimeseriesdata.Oneexampleofcontinuousestimationisforecastingthesalesdemandforaproduct,basedonasetofinputdatasuchasprevioussalesfigures,consumersentiment,andweather.Anotherexampleispredictingthepriceofrealestate,suchasabuilding,usingdatadescribingthepropertycombinedwithphotosofit.

Clustering.Theseproblemsrequireasystemtocreateasetofcategories,forwhichindividualdatainstanceshaveasetofcommonorsimilarcharacteristics.Anexampleofclusteringiscreatingasetofconsumersegmentsbasedondataaboutindividualconsumers,includingdemographics,preferences,andbuyerbehavior.

Allotheroptimization.Theseproblemsrequireasystemtogenerateasetofoutputsthatoptimizeoutcomesforaspecificobjectivefunction(someoftheotherproblemtypescanbeconsideredtypesofoptimization,sowedescribetheseas“allother”optimization).Generatingarouteforavehiclethatcreatestheoptimumcombinationoftimeandfueluseisanexampleofoptimization.

Anomalydetection.Givenatrainingsetofdata,determinewhetherspecificinputsareoutoftheordinary.Forinstance,asystemcouldbetrainedonasetofhistoricalvibrationdataassociatedwiththeperformanceofanoperatingpieceofmachinery,andthendeterminewhetheranewvibrationreadingsuggeststhatthemachineisnotoperatingnormally.Notethatanomalydetectioncanbeconsideredasubcategoryofclassification.

Ranking.Rankingalgorithmsareusedmostoftenininformationretrievalproblemsinwhichtheresultsofaqueryorrequestneedstobeorderedbysomecriterion.

Recommendationsystemssuggestingnextproducttobuyusethesetypesofalgorithmsasafinalstep,sortingsuggestionsbyrelevance,beforepresentingtheresultstotheuser.

Recommendations.Thesesystemsproviderecommendations,basedonasetoftrainingdata.Acommonexampleofrecommendationsaresystemsthatsuggestthe“nextproducttobuy”foracustomer,basedonthebuyingpatternsofsimilarindividuals,andtheobservedbehaviorofthespecificperson.

Datageneration.Theseproblemsrequireasystemtogenerateappropriatelynoveldatabasedontrainingdata.Forinstance,amusiccompositionsystemmightbeusedtogeneratenewpiecesofmusicinaparticularstyle,afterhavingbeentrainedonpiecesofmusicinthatstyle.

Exhibit2illustratestherelativetotalvalueoftheseproblemtypesacrossourdatabaseofusecases,alongwithsomeofthesampleanalyticstechniquesthatcanbeusedtosolveeachproblemtype.Themostprevalentproblemtypesareclassification,continuousestimation,andclustering,sugge