AI執(zhí)行者指南報告

Anexecutive’sguidetoAI

PAGE

1

Anexecutive’sguidetoAI

Stayingaheadintheacceleratingartificial-intelligenceracerequiresexecutivestomakenimble,informeddecisionsaboutwhereandhowtoemployAIintheirbusiness.Onewaytopreparetoactquickly:knowtheAIessentialspresentedinthisguide.

Artificialintelligence:Adefinition

AIistypicallydefinedastheabilityofamachinetoperformcognitivefunctionsweassociatewithhumanminds,suchasperceiving,reasoning,learning,andproblemsolving.ExamplesoftechnologiesthatenableAItosolvebusinessproblemsareroboticsandautonomousvehicles,computervision,language,virtualagents,andmachinelearning.

Machinelearning:Adefinition

MostrecentadvancesinAIhavebeenachievedbyapplyingmachinelearningtoverylargedatasets.Machine-learningalgorithmsdetectpatternsandlearnhowtomakepredictionsandrecommendationsbyprocessingdataandexperiences,ratherthanbyreceivingexplicitprogramminginstruction.Thealgorithmsalsoadaptinresponsetonewdataandexperiencestoimproveefficacyovertime.

Understandingthemajortypesofmachinelearning

Supervisedlearning

Whatitis

Analgorithmusestrainingdataandfeedbackfromhumanstolearntherelationshipofgiveninputstoagivenoutput(eg,howtheinputs“timeofyear”and“interestrates”predicthousingprices)

Whentouseit

Youknowhowtoclassifytheinputdataandthetypeofbehavioryouwanttopredict,butyouneedthealgorithmtocalculateitforyouonnewdata

Howitworks

Ahumanlabelseveryelementoftheinputdata(eg,inthecaseof

predictinghousingprices,labelstheinputdataas“timeofyear,”“interestrates,”etc)anddefinestheoutputvariable(eg,housingprices)

Thealgorithmistrainedonthedatatofindtheconnectionbetweentheinputvariablesandtheoutput

Oncetrainingiscomplete—typicallywhenthealgorithmissufficientlyaccurate—thealgorithmisappliedtonewdata

Unsupervisedlearning

Analgorithmexploresinputdatawithoutbeinggiven

anexplicitoutputvariable(eg,explorescustomerdemographicdatatoidentifypatterns)

Youdonotknowhowtoclassifythedata,andyouwantthealgorithmtofindpatternsandclassifythedataforyou

Thealgorithmreceivesunlabeleddata(eg,asetofdatadescribingcustomerjourneysonawebsite)

Itinfersastructurefromthedata

Thealgorithmidentifiesgroupsofdatathatexhibitsimilarbehavior(eg,formsclusters

ofcustomersthatexhibitsimilarbuyingbehaviors)

Reinforcementlearning

Analgorithmlearnstoperformatasksimplybytryingto

maximizerewardsitreceivesforitsactions(eg,maximizespointsitreceivesforincreasingreturnsofaninvestmentportfolio)

Youdon’thavealotoftrainingdata;youcannotclearlydefinetheidealendstate;ortheonlywaytolearnabouttheenvironmentistointeractwithit

Thealgorithmtakesanactionontheenvironment(eg,makesatradeinafinancialportfolio)

Itreceivesarewardiftheactionbringsthemachineastepclosertomaximizingthetotalrewardsavailable(eg,thehighesttotalreturnontheportfolio)

Thealgorithmoptimizesforthebestseriesofactionsbycorrectingitselfovertime

Supervisedlearning:Algorithmsandsamplebusinessusecases1

Algorithms Samplebusinessusecases

Linearregression

Highlyinterpretable,standardmethodformodel-ingthepastrelationshipbetweenindependentinputvariablesanddependentoutputvariables(whichcanhaveaninfinitenumberofvalues)tohelppredictfuturevaluesoftheoutputvariables

Logisticregression

Extensionoflinearregressionthat’susedforclassifationtasks,meaningtheoutputvariableisbinary(eg,onlyblackorwhite)ratherthan

continuous(eg,aninfinitelistofpotentialcolors)

Linear/quadraticdiscriminantanalysisUpgradesalogisticregressiontodealwithnonlinearproblems—thoseinwhichchangestothevalueofinputvariablesdonotresultinproportionalchangestotheoutputvariables.

Understandproduct-salesdriverssuchascompetitionprices,distribution,advertisement,etc

Optimizepricepointsandestimateproduct-priceelasticities

Classifycustomersbasedonhowlikelytheyaretorepayaloan

Predictifaskinlesionisbenignormalignantbasedonitscharacteristics(size,shape,color,etc)

Predictclientchurn

Predictasaleslead’slikelihoodofclosing

Decisiontree

Highlyinterpretableclassificationorregressionmodelthatsplitsdata-featurevaluesintobranchesatdecisionnodes(eg,ifafeatureisacolor,eachpossiblecolorbecomesanewbranch)untilafinaldecisionoutputismade

Provideadecisionframeworkforhiringnewemployees

Understandproductattributesthatmakeaproductmostlikelytobepurchased

NaiveBayes

ClassificationtechniquethatappliesBayestheorem,whichallowstheprobabilityofaneventtobecalculatedbasedonknowledgeoffactorsthatmightaffectthatevent(eg,ifanemailcontainstheword“money,”thentheprobabilityofitbeingspamishigh)

Analyzesentimenttoassessproductperceptioninthemarket

Createclassifierstofilterspamemails

1We’velistedsomeofthemostcommonlyusedalgorithmstoday—thislistisnotintendedtobeexhaustive.Additionally,anumberofdifferentmodelscanoftensolvethesamebusinessproblem.Conversely,thenatureofanavailabledatasetoftenprecludesusingamodeltypicallyemployedtosolveaparticularproblem.Forthesereasons,thesamplebusinessusecasesaremeantonlytobeillustrativeofthetypesofproblemsthesemodelscansolve.

Algorithms Samplebusinessusecases

Supportvectormachine

Atechniquethat’stypicallyusedforclassificationbutcanbetransformedtoperformregression.Itdrawsanoptimaldivisionbetweenclasses(aswideaspossible).Italsocanbequicklygeneralizedtosolvenonlinearproblems

Randomforest

Classificationorregressionmodelthatimprovestheaccuracyofasimpledecisiontreebygeneratingmultipledecisiontreesandtakingamajorityvoteofthemtopredicttheoutput,whichisacontinuousvariable(eg,age)foraregressionproblemandadiscretevariable(eg,eitherblack,white,orred)forclassification

Predicthowmanypatientsahospitalwillneedtoserveinatimeperiod

Predicthowlikelysomeoneistoclickonanonlinead

Predictcallvolumeincallcentersforstaffingdecisions

Predictpowerusageinanelectrical-distributiongrid

AdaBoost

Classificationorregressiontechniquethatusesamultitudeofmodelstocomeupwithadecisionbutweighsthembasedontheiraccuracyinpredictingtheoutcome

Detectfraudulentactivityincredit-cardtransactions.Achievesloweraccuracythandeeplearning

Simple,low-costwaytoclassifyimages(eg,recognizelandusagefromsatelliteimagesforclimate-changemodels).Achievesloweraccuracythandeeplearning

Gradient-boostingtrees

Classificationorregressiontechniquethatgeneratesdecisiontreessequentially,whereeachtreefocusesoncorrectingtheerrorscomingfromtheprevioustreemodel.Thefinaloutputisacombinationoftheresultsfromalltrees

Forecastproductdemandandinventorylevels

Predictthepriceofcarsbasedontheircharacteristics(eg,ageandmileage)

Simpleneuralnetwork

Modelinwhichartificialneurons(software-basedcalculators)makeupthreelayers(aninputlayer,ahiddenlayerwherecalculationstakeplace,andanoutputlayer)thatcanbeusedtoclassifydataorfindtherelationshipbetweenvariablesinregressionproblems

Predicttheprobabilitythatapatientjoinsahealthcareprogram

Predictwhetherregistereduserswillbe

willingornottopayaparticularpriceforaproduct

Unsupervisedlearning:Algorithmsandsamplebusinessusecases2

Algorithms Samplebusinessusecases

K-meansclustering

Putsdataintoanumberofgroups(k)thateachcontaindatawithsimilarcharacteristics(asdeterminedbythemodel,notinadvancebyhumans)

Gaussianmixturemodel

Ageneralizationofk-meansclusteringthatprovidesmoreflexibilityinthesizeandshapeofgroups(clusters)

Hierarchicalclustering

Splitsoraggregatesclustersalongahierarchicaltreetoformaclassificationsystem

Segmentcustomersintogroupsbydistinctcharateristics(eg,agegroup)—forinstance,tobetterassignmarketingcampaignsorpreventchurn

Segmentcustomerstobetterassignmarketingcampaignsusingless-distinctcustomercharacteristics(eg,productpreferences)

Segmentemployeesbasedonlikelihoodofattrition

Clusterloyalty-cardcustomersintoprogressivelymoremicrosegmentedgroups

Informproductusage/developmentbygroupingcustomersmentioningkeywordsinsocial-mediadata

Recommendersystem

Oftenusesclusterbehaviorpredictiontoidentifytheimportantdatanecessaryformakingarecommendation

Recommendwhatmoviesconsumersshouldviewbasedonpreferencesofothercustomerswithsimilarattributes

Recommendnewsarticlesareadermightwanttoreadbasedonthearticlesheorheisreading

2We’velistedsomeofthemostcommonlyusedalgorithmstoday—thislistisnotintendedtobeexhaustive.Additionally,anumberofdifferentmodelscanoftensolvethesamebusinessproblem.Conversely,thenatureofanavailabledatasetoftenprecludesusingamodeltypicallyemployedtosolveaparticularproblem.Forthesereasons,thesamplebusinessusecasesaremeantonlytobeillustrativeofthetypesofproblemsthesemodelscansolve.

Anexecutive’sguidetoAI

PAGE

6

Reinforcementlearning:Samplebusinessusecases3

Optimizethetradingstrategyforanoptions-tradingportfolio

Balancetheloadofelectricitygridsinvaryingdemandcycles

Stockandpickinventoryusingrobots

Optimizethedrivingbehaviorofself-drivingcars

Optimizepricinginrealtimeforanonlineauctionofaproductwithlimitedsupply

Deeplearning:Adefinition

Deeplearningisatypeofmachinelearningthatcanprocessawiderrangeofdataresources,requireslessdatapreprocessingbyhumans,andcanoftenproducemoreaccurateresultsthantraditionalmachine-learningapproaches.Indeeplearning,interconnectedlayersofsoftware-basedcalculatorsknownas“neurons”formaneuralnetwork.Thenetworkcaningestvastamountsofinputdataandprocessthemthroughmultiplelayersthatlearnincreasinglycomplexfeaturesofthedataateachlayer.Thenetworkcanthenmakeadeterminationaboutthedata,learnifitsdeterminationiscorrect,andusewhatithaslearnedtomakedeterminationsaboutnewdata.Forexample,onceitlearnswhatanobjectlookslike,itcanrecognizetheobjectinanewimage.

3Thesamplebusinessusecasesaremeantonlytobeillustrativeofthetypesofproblemsthesemodelscansolve.

Understandingthemajordeeplearningmodelsandtheirbusinessusecases4

Convolutionalneuralnetwork Recurrentneuralnetwork

Whatitis

Amultilayeredneuralnetworkwithaspecialarchitecturedesignedtoextractincreasinglycomplexfeaturesofthedataateachlayertodeterminetheoutput

Whentouseit

Whenyouhaveanunstructureddataset(eg,images)andyouneedtoinferinformationfromit

Amultilayeredneuralnetworkthatcanstoreinformationincontextnodes,allowingittolearndatasequencesandoutputanumberoranothersequence

Whenyouareworkingwithtime-seriesdataorsequences(eg,audiorecordingsortext)

4Thesamplebusinessusecasesaremeantonlytobeillustrativeofthetypesofproblemsthesemodelscansolve.

Convolutionalneuralnetwork Recurrentneuralnetwork

Howitworks

Processinganimage

Theconvolutionalneuralnetwork(CNN)receivesanimage—forexample,oftheletter“A”—thatitprocessesasacollectionofpixels

Inthehidden,innerlayersofthemodel,itidentifiesuniquefeatures,forexample,theindividuallinesthatmakeup“A”

TheCNNcannowclassifyadifferentimageastheletter“A”ifitfindsinittheuniquefeaturespreviouslyidentifiedasmakinguptheletter

Predictingthenextwordinthesentence“Areyoufree ?”

Arecurrentneuralnetwork(RNN)neuronreceivesacommandthatindicatesthestartofasentence

Theneuronreceivestheword“Are”andthenoutputsavectorofnumbersthatfeedsbackintotheneurontohelpit“remember”thatitreceived“Are”(andthatitreceiveditfirst).Thesameprocessoccurswhenitreceives“you”and“free,”

withthestateoftheneuronupdatinguponreceivingeachword

Afterreceiving“free,”theneuronassignsaprobabilitytoeverywordintheEnglishvocabularythatcouldcompletethesentence.Iftrainedwell,theRNNwillassigntheword“tomorrow”oneofthehighestprobabilitiesandwillchooseittocompletethesentence

Businessusecases

Diagnosehealthdiseasesfrommedicalscans

Detectacompanylogoinsocialmediatobetterunderstandjointmarketing

opportunities(eg,pairingofbrandsinoneproduct)

Understandcustomerbrandperceptionandusagethroughimages

Detectdefectiveproductsonaproductionlinethroughimages

Generateanalystreportsforsecuritiestraders

Providelanguagetranslation

Trackvisualchangestoanareaafteradisastertoassesspotentialdamageclaims(inconjunctionwithCNNs)

Assessthelikelihoodthatacredit-cardtransactionisfraudulent

Generatecaptionsforimages

Powerchatbotsthatcanaddressmorenuancedcustomerneedsandinquiries

Timeline:WhyAInow?

Aconvergenceofalgorithmicadvances,dataproliferation,andtremendousincreasesincomputingpowerandstoragehaspropelledAIfromhypetoreality.

Algorithmicadvancements

+

1805–Legendrelaysthegroundworkformachinelearning

FrenchmathematicianAdrien-MarieLegendrepublishestheleastsquaremethodforregression,whichheusedtodetermine,fromastronomicalobservations,theorbitsofbodiesaroundthesun.Althoughthismethodwasdevelopedasastatisticalframework,itwouldprovidethebasisformanyoftoday’smachine-learningmodels.

+

1958–Rosenblattdevelopsthefirstself-learningalgorithm

AmericanpsychologistandcomputerscientistFrankRosenblattcreatestheperceptronalgorithm,anearlytypeofartificialneuralnetwork(ANN),whichstandsasthefirstalgorithmicmodelthatcouldlearnonitsown.AmericancomputerscientistArthurSamuelwouldcointheterm“machinelearning”thefollowingyearforthesetypesofself-learningmodels(aswellasdevelopagroundbreakingcheckersprogramseenasanearlysuccessinAI).

+

1965–Birthofdeeplearning

UkrainianmathematicianAlexeyGrigorevichIvakhnenkodevelopsthefirstgeneralworkinglearningalgorithmsforsupervisedmultilayerartificialneuralnetworks(ANNs),inwhichseveralANNsarestackedontopofoneanotherandtheoutputofoneANNlayerfeedsintothenext.Thearchitectureisverysimilartotoday’sdeep-learningarchitectures.

+

1986–Backpropagationtakeshold

AmericanpsychologistDavidRumelhart,

BritishcognitivepsychologistandcomputerscientistGeoffreyHinton,andAmericancomputerscientistRonaldWilliamspublishonbackpropagation,popularizingthis

keytechniquefortrainingartificialneuralnetworks(ANNs)thatwasoriginallyproposedbyAmericanscientistPaulWerbosin1982.

BackpropagationallowstheANNtooptimizeitselfwithouthumanintervention(inthiscase,itfoundfeaturesinfamily-treedatathatweren’tobviousorprovidedtothealgorithminadvance).Still,lackofcomputationalpower

andthemassiveamountsofdataneededtotrainthesemultilayerednetworkspreventANNsleveragingbackpropagationfrombeingusedwidely

Explosionofdata

+ +

1991–OpeningoftheWorldWideWeb

TheEuropeanOrganizationforNuclearResearch(CERN)beginsopeninguptheWorldWideWebtothepublic.

Early2000s–BroadbandadoptionbeginsamonghomeInternetusers

BroadbandallowsusersaccesstoincreasinglyspeedyInternetconnections,upfromthepaltry

56kbpsavailablefordownloadingthroughdial-upinthelate1990s.Today,availablebroadbandspeedscansurpass100mbps

(1mbps=1,000kbps).Bandwidth-hungryapplicationslike

YouTubecouldnothavebecomecommerciallyviablewithouttheadventofbroadband.

Exponentialincreasesincomputingpowerandstorage

+

1965–Moorerecognizesexponentialgrowthinchippower

IntelcofounderGordonMoorenoticesthatthenumberoftransistorspersquareinchonintegratedcircuitshasdoubledeveryyearsincetheirinvention.HisobservationbecomesMoore’slaw,whichpredictsthetrendwillcontinueintotheforeseeablefuture(althoughitlaterprovestodosoroughlyevery18months).Atthetime,state-of-the-artcomputationalspeedisintheorderofthreemillionfloating-pointoperationspersecond(FLOPS).

+

1997–IncreaseincomputingpowerdrivesIBM’sDeepBlue

victoryoverGarryKasparovDeepBlue’ssuccessagainsttheworldchesschampionlargelystemsfrommasterfulengineeringandthetremendouspowercomputerspossessatthattime.DeepBlue’scomputerachievesaround11gigaFLOPS(11billionFLOPS).

+

1999–MorecomputingpowerforAIalgorithmsarrives…butnoonerealizesityet

NvidiareleasestheGeForce256graphicscard,marketedastheworld’sfirsttruegraphicsprocessingunit(GPU).Thetechnologywilllaterprove

fundamentaltodeeplearningbyperformingcomputationsmuchfasterthancomputerprocessingunits(CPUs).

9 Anexecutive’sguidetoAI

+

1989–BirthofCNNsforimagerecognitionFrenchcomputerscientistYannLeCun,nowdirectorofAIresearchforFacebook,andotherspublishapaperdescribinghowatypeofartificialneuralnetworkcalledaconvolutionalneuralnetwork(CNN)iswellsuitedfor

shape-recognitiontasks.LeCunandteamapplyCNNstothetaskofrecognizinghandwrittencharacters,withtheinitialgoalofbuildingautomaticmail-sortingmachines.Today,CNNsarethestate-of-the-artmodelforimagerecognitionandclassification.

+

+

1992–UpgradedSVMsprovideearlynatural-language-processingsolutionComputerengineersBernhardE.Boser(Swiss),IsabelleM.Guyon(French),andRussianmathematicianVladimirN.Vapnikdiscoverthatalgorithmicmodelscalledsupportvectormachines(SVMs)canbeeasilyupgradedtodealwithnonlinearproblemsbyusingatechniquecalledkerneltrick,leadingtowidespreadusageofSVMsinmanynatural-language-processingproblems,suchasclassifyingsentimentandunderstandinghumanspeech.

1997–RNNsgeta“memory,”positioningthemtoadvancespeechtotext

+

+

In1991,GermancomputerscientistSeppHochreitershowedthataspecialtypeofartificialneuralnetwork(ANN)calledarecurrentneuralnetwork(RNN)canbeusefulinsequencingtasks(speechtotext,forexample)ifitcouldrememberthebehaviorofpartsequencesbetter.In1997,HochreiterandfellowcomputerscientistJürgenSchmidhubersolvetheproblembydevelopinglongshort-termmemory(LSTM).Today,RNNswithLSTMareusedinmanymajorspeech-recognitionapplications.

+

1998–BrinandPagepublishPageRankalgorithm

+

+

Thealgorithm,whichrankswebpageshigherthemoreotherwebpageslinktothem,formstheinitialprototypeofGoogle’ssearchengine.ThisbrainchildofGooglefoundersSergeyBrinandLarryPagerevolutionizesInternetsearches,openingthedoortothecreationandconsumptionofmorecontentanddataontheWorldWideWeb.Thealgorithmwouldalso

goontobecomeoneofthemostimportantforbusinessesastheyvieforattentiononanincreasinglysprawlingInternet.

2004–Facebookdebuts

HarvardstudentMarkZuckerbergandteamlaunch“Thefacebook,”asitwasoriginallydubbed.Bytheendof2005,thenumberofdata-generating

Facebookusersapproachessixmillion.

2004–Web2.0hitsitsstride,launchingtheeraofuser-generateddata

Web2.0referstotheshiftingoftheInternetparadigmfrompassivecontentviewingtointeractiveandcollaborativecontentcreation,socialmedia,blogs,video,andotherchannels.PublishersTimO’ReillyandDaleDoughertypopularize

theterm,thoughitwascoinedbydesignerDarcyDiNucciin1999.

2005–YouTubedebuts

Withinabout18months,thesitewouldserveupalmost100millionviewsperday.

2005–NumberofInternetusersworldwidepassesone-billionmark

2002—Amazonbringscloudstorageandcomputingtothemasses

+

AmazonlaunchesitsAmazonWebServices,offeringcloud-basedstorageandcomputingpowertousers.Cloudcomputingwouldcometorevolutionizeanddemocratizedatastorageandcomputation,givingmillions

ofusersaccesstopowerfulITsystems—previouslyonlyavailabletobigtechcompanies—atarelativelylowcost.

+

2004–DeanandGhemawatintroducetheMapReducealgorithmtocopewithdataexplosion

WiththeWorldWideWebtakingoff,Googleseeksoutnovel

ideastodealwiththeresultingproliferationofdata.ComputerscientistJeffDean(currentheadofGoogleBrain)andGooglesoftwareengineerSanjayGhemawatdevelopMapReducetodealwithimmenseamountsofdatabyparallelizingprocessesacross

largedatasetsusingasubstantialnumberofcomputers.

+

2005–Costofonegigabyteofdiskstoragedropsto$0.79,from$277tenyearsearlierAndthepriceofDRAM,atypeofrandom-accessmemory(RAM)commonlyusedinPCs,dropsto

$158pergigabyte,from$31,633in1995.

10 Anexecutive’sguidetoAI

+

2006–Hintonreenergizestheuseofdeep-learningmodelsTospeedthetrainingofdeep-

learningmodels,GeoffreyHintondevelopsawaytopretrainthemwithadeep-beliefnetwork(aclassofneuralnetwork)beforeemployingbackpropagation.Whilehismethodwouldbecomeobsoletewhencomputationalpowerincreased

toalevelthatallowedforefficientdeep-learning-modeltraining,Hinton’sworkpopularizedtheuseofdeeplearningworldwide—andmanycredithimwithcoiningthephrase“deeplearning.”

2007–IntroductionoftheiPhonepropelssmartphonerevolution—andampsupdatageneration

+

ApplecofounderandCEOSteveJobsintroducestheiPhoneinJanuary2007.Thetotalnumberofsmartphonessoldin2007reachesabout122million.Theeraofaround-the-clockconsumptionandcreationofdataandcontentbysmartphoneusersbegins.

2006–CuttingandCafarellaintroduceHadooptostoreandprocessmassiveamountsofdata

+

InspiredbyGoogle’sMapReduce,computerscientistsDougCuttingandMikeCafarelladeveloptheHadoopsoftwaretostoreandprocessenormousdatasets.

Yahoousesitfirst,todealwiththeexplosionofdatacomingfromindexingwebpagesandonlinedata.

2009–UCBerkeleyintroducesSparktohandlebigdatamodelsmoreefficientlyDevelopedbyRomanian-CanadiancomputerscientistMateiZahariaatUCBerkeley’sAMPLab,SparkstreamshugeamountsofdataleveragingRAM,makingitmuchfasteratprocessingdatathansoftwarethatmustread/writeonharddrives.Itrevolutionizestheabilitytoupdatebigdataandperformanalyticsinrealtime.

++

2009–NgusesGPUstotraindeep-learningmodelsmoreeffectively

AmericancomputerscientistAndrewNgandhisteamatStanfordUniversityshowthattrainingdeep-beliefnetworkswith100millionparametersonGPUsismorethan70timesfasterthandoingsoonCPUs,afindingthatwouldreducetrainingthatoncetookweekstoonlyoneday.

2010–Numberofsmartphonessoldintheyearnears300million

+

Thisrepresentsanearly2.5timesincreaseoverthenumbersoldin2007.

+

2010–MicrosoftandGoogleintroducetheirclouds

CloudcomputingandstoragetakeanothersteptowardubiquitywhenMicrosoftmakesAzureavailableandGooglelaunchesitsGoogleCloudStorage(theGoogleCloudPlatformwouldcomeonlineaboutayearlater).

+

2010–WorldwideIPtrafficexceeds20exabytes(20billiongigabytes)permonthInternetprotocol(IP)trafficisaidedbygrowingadoptionofbroadband,particularlyintheUnitedStates,whereadoptionreaches65percent,accordingtoCisco,whichreportsthismonthlyfigureandtheannualfigureof242exabytes.

11 Anexecutive’sguidetoAI

2012–NumberofFacebookusershitsonebillion

Theamountofdataprocessedbythecompany’ssystemssoarspast500terabytes.

2012–Deep-learningsystemwinsrenownedimage-classificationcontestforthefirsttime

GeoffreyHinton’steamwinsImageNet’simage-classificationcompetitionbyalarge

margin,withanerrorrateof15.3percentversusthesecond-besterrorrateof26.2percent,usingaconvolutionalneuralnetwork(CNN).Hinton’steamtraineditsCNNon1.2millionimagesusing

2011–IBMWatsonbeatsJeopardy!

+

+

IBM’squestionansweringsystem,Watson,defeatsthetwogreatestJeopardy!champions,BradRutterandKenJennings,byasignificantmargin.IBMWatsonusestenracksofIBMPower750serverscapableof80teraFLOPS(that’s