注冊(cè)會(huì)計(jì)師介紹AI：從算法到深度學(xué)習(xí)你需要了解的內(nèi)容

ACPA’sIntroductiontoAI:FromAlgorithms

toDeepLearning,WhatYouNeedtoKnow

ii

DISCLAIMER

ThispaperwaspreparedbytheCharteredProfessionalAccountantsofCanada(CPACanada)andtheAmericanInstituteofCPAs(AICPA),asnon-authoritativeguidance.

CPACanadaandAICPAdonotacceptanyresponsibilityorliabilitythatmightoccurdirectlyorindirectlyasaconsequenceoftheuse,applicationorrelianceonthismaterial.

?2019CharteredProfessionalAccountantsofCanada

Allrightsreserved.Thispublicationisprotectedbycopyrightandwrittenpermissionisrequiredtoreproduce,storeinaretrievalsystemortransmitinanyformorbyanymeans(electronic,mechanical,photocopying,recording,orotherwise).

Forinformationregardingpermission,pleasecontact

permissions@cpacanada.ca

ACPA’SINTRODUCTIONTOAI:FROMALGORITHMSTODEEPLEARNING,WHATYOUNEEDTOKNOW

PAGE

10

ExecutiveSummary

Althoughnotalwaysnoticeabletothegeneralpublic,ArtificialIntelligence(AI)hasbeenaroundsincethe1950s.However,advancementsinAIhaveacceleratedinthepastdecadeduetotheriseincomputingpower,theavailabilityofdata,andimprovementsinmachinelearningalgorithms.Asaresult,AIcapabilitieshavebeguntoinfiltrateourdailylives—somemoresothanothers.WithAIadoptionincreasingforbothconsumersandbusinesses,itisimportantforCPAstounderstandthefundamentalsofAIandthepotentialimpacttotheirorganizations,clients,andthemselves.

CPAshaveanopportunitytouseAIasanenablerofinnovationandincreasedproductivity.LeveragingAI,CPAscanperformbetterqualityworkwithhigherefficiency.Buttodothat,CPAsneedasolidunderstandingofhowAIworksbeforetheycanidentifytheseopportuni-ties.Forexample,roboticprocessautomationcanincreaseproductivityandappeartoworklikeAI,butisit?WheredomachinelearninganddeeplearningfitintotheworldofAI?

ThispublicationwasdevelopedbyCharteredProfessionalAccountantsofCanada(CPACanada)andtheAmericanInstituteofCPAs(AICPA)tobeusedasafoundationalresource.Itwillexplainthebuzzwordsandtermsyouhavelikelybeenhearing,discusstheevolutionofdata,AIandcomputingpowerandhelpyoutobegintothinkaboutAIandhowitmightimpactyourwork.FuturepublicationsdedicatedtoAIwillrefertothisresourceasitisthefirstofaplannedseriesofpublicationstoexploreAIanditsimpactontheCPAprofession.

CPACanadaandtheAICPAencourageallCPAstocontinuereading,collaboratingwithcolleaguesandlearningaboutAIandothertechnologieswhichmayimpactyouasaCPA.Weareallonthisdigitaljourneytogether!

Introduction

Asyoureadthis,asmartphoneappcalledClimateBasicistellingfarmerswheretoplanttheircrops.Theappdividesthelandintoplots,scansthroughlocaltemperatureanderosionrecords,expectedrainfall,soilqualityandotheragriculturaldatatofigureouthowtomaxi-mizeyields.

Somewhere,someoneissittingontheircouchidlyscrollingthroughNetflix’srecommendations,blissfullyunawaretheyareuniquetothemandweregeneratedbyanalgorithmthatcomparestheirindividualTV-viewingpatternswiththoseofmillionsofusersacross190countries.

Intheaccountingprofessionweseeautomationinaccountspayablewheremanualtaskshavebeenreplacedbysystemsthatdoeverythingfrommatchingpurchaseorderstoflag-ginginvoicesforpayments.Moreover,throughOpticalCharacterRecognitionandmachinelearning,algorithmscanreadthoseinvoicesandclassifyexpensetypesoraccounttypesautomaticallywithahighdegreeofaccuracy.

WelcometoArtificialIntelligence(AI).FromvirtualassistantSiritonavigationappWazetoautonomouscarstohumanoidrobotbankambassadorsandhospitalreceptionists,

toGoogleAssistant’sabilitytocallahairsalonandsetupanappointmentforyou,AIischanginghowweliveandwork.AIiseverywhereandwearestilljustscratchingthesurfaceofhowitwillchangehowweworkandlive.

ButWhatIsAI,Really?

Ashortanswerandthebroadestdefinition:AIisthescienceofteachingprogramsandmachinestocompletetasksthatnormallyrequirehumanintelligence.

Thenewspacerace:GlobalinitiativestowinatAI

Ofthe$15.2billioninvestedgloballyinAIstart-upsin2017,48%wenttoChinaand38%wenttotheU.S.,asperCBInsights.1ThisisindicativeofaglobalracetoleadinAI.TheChinesegovernmentinitsStateCouncilpaperhaslaidoutitsstrategytoemergeasaglobalAIhubby2030.

Ontheotherhand,intheUnitedStates,theDefenseAdvancedResearchProjectsAgency(DARPA),announcedinSeptember2018a$2billioncampaigntodevelopthenextwaveofAItechnologiesfocusedoncon-textualreasoningcapabilities.2

TheEuropeanUnion(EU)inApril2018launcheda$1.8billioninvestmentinAIresearchundertheEU’sHorizon2020fund.3

Francehasannounceda$1.8billioninvestmentpolicythrough2022focusedonAIanddatasharing.4

TheUnitedKingdomgovernmentannounceda￡300millioninvestmentallocationinApril2018and50leadingUKbusinesseshavecomefor-wardtoseta￡1billiondealinAI.5

TheGovernmentofCanadaisallinaswell.In2017,ittaskedtheCanadianInstituteforAdvancedResearchtodevelopa$125millionAIstrategy.6TheCanadiangovernmentisalsoinvesting$950millionintheInnova-tionSuperClusterinitiativewhichincludesSCALE.AI(SupplyChainsandLogisticsExcellence.AI)anindustry-ledconsortiumthatwillshapeanewAI-poweredglobalsupplychainplatform.7

/the-download/610271/chinas-ai-startups-scored-more-funding-than-americas-last-year/

/news-events/2018-09-07

/sites/amyguttman/2018/09/28/three-innovative-european-startups-to-watch/#507861316f37

/article/us-france-tech/france-to-spend-1-8-billion-on-ai-to-compete-with-u-s-china-idUSKBN1H51XP

/business/uk-government-ai-research

www.newswire.ca/news-releases/canada-funds-125-million-pan-canadian-artificial-intelligence-strategy-616876434.html

www.ic.gc.ca/eic/site/093.nsf/eng/home

WhyIsItImportantforCPAstoStayCurrentwithWhatIsHappeninginAI?

Simplyput,AIhasthepotentialtosignificantlyimpactaccountingandassurancejobsandbusinessesinthenearfuture—ifithasnotalready.Transactionaltasksareincreasinglybeingautomated.Algorithmshavetheabilitytospotsubtlecluesindatasetstopinpointfraudoridentifynewopportunitiessuchassavings,improvedefficiencyandprofitability,andnewmarketopportunities.Allthiswhileleveragingandmakingsenseofalargevolumeofinformationcomingfrommultiplesources,commonlyreferredtoasBigData.

Ofcourse,everyonehasanopinionaboutAI.Somefocusonthepositiveandhowitwillimproveeverythingfromtransportationtomedicaldiagnostics.Othersseeproblemsandethicalconcerns:thepotentialeliminationofjobs,andtheviolationofprivacy.HowfarcanwetrusttheautonomyofAIsystems?Forexample,howdoesAIresolveaself-drivingcar’sdilemmawhenitmustdecidewhethertocrashintoanoncomingcarorpedestrianwhenanaccidentisunavoidable?Thereisalsothequestionofdecisiontransparency.WouldeveryAIsystembeabletoexplainintermsweunderstand,thereasonsforitsdecisions?

CPAshaveanopportunitytofocusonAIasanenablerofinnovationandincreasedproduc-tivity.ConsidertheClimateBasicapphelpingfarmers.AccordingtotheU.S.DepartmentofAgriculture,thisuseofAIacrosstheindustryhasproducedthelargestcropsinthecoun-try’shistory.AndNetflixsaysitssophisticatedrecommendationenginesavesthecompany

$1billioninlostsubscribers.8CPAshavetheopportunitytoimproveefficienciesandqual-ityofday-to-dayworkthroughAIthatreadscontracts,performsautomatedtaskssuchasaccuratelyidentifyingandbookingexpenses,oridentifyingriskytransactionsinrealtime.

Fastfacts:ThepotentialimpactofAIisalsofuelinginvestmentsandresearch.TheInternationalDataCorporation(IDC)forecastsAI-relatedinvestmentswillgrowfrom$12billionin2017to$57.6billionin2021.Since2013,machine-learning-basedAIhasbeenthethird-fastest-growingcategoryforpatentfiling.Onthejobfront,employmentsearchengineIndeedreportstheshareofjobsinCanadaandtheU.S.requiringAIskillshassignificantlygrownsince2013.

/netflix-recommendation-engine-worth-1-billion-per-year-2016-6

ThispublicationwasdevelopedbyCPACanadaandtheAICPAtobeusedasafoundationalresource.Itwillexplainthebuzzwordsandtermsyouhavelikelybeenhearing,discusstheevolutionofdata,AIandcomputingpowerandhelpyoutobegintolearnaboutAIandthinkhowitmightimpactyourwork.CPACanadawillalsorefertothispublicationinfuturepubli-cationsdedicatedtoAIasthefirstofaplannedseriesofpublicationstoexploretheimpactofAIintheCPAprofession.Thediagrambelowprovidesanoverviewoftheterminologytobecoveredthroughoutthispublicationandhelpsprovideavisualrepresentationofwherethesetechnologiesfitinrelationtoeachother.

ArtificialIntelligence

MachineLearning

DeepLearning

NoArtificialIntelligence

ArtificialIntelligence

ArtificialIntelligence ArtificialIntelligenceincludingMachine includingDeep

Learning Learning

AlgorithmsusinglargecomplicatedNeuralNetworks

CognitiveTechnologies

ComputerVision

DataAnalytics(Predictive&Prescriptive)

AlgorithmsusingNeuralNetworks

NaturalLanguageProcessing

–CognitiveProcessAutomation(withAI)

BasicDataAnalytics(Descriptive

&Diagnostic)

Automation

RoboticProcessAutomation

BasicAlgorithms

HowWeGotHere

Youcan’thaveMachinelearningwithoutdata.

Itallstartswithdata.Companieshavehistoricallyuseddatatoimprovetheirbusinessperformance(e.g.,customerfeedbacksurveys,planningresourceallocationusingtheCriticalPathMethod,etc.).However,datagatheringwasanexplicit

exerciseandthoseperformingitmightnothavetakenstepstoverifythatthedatacollectedrepresentedwhatwasactuallyhappening.Forexample,itmayhavebeendifficulttodeterminewhethersurveyrespondentstrulyrepresentedthecustomerbaseorwhethertheirbuyinghabitswerefardifferentfromthoseofmoretypicalcustomers.

WiththeInternetandtheevolutiontoadigitalworld,dataacquisitionhasgonefrombeinganexplicitandoftenmanualexercisetoanongoing,automatedpassiveactivity.Websites,appsandsocialmediacantrackeverythingfromyourshoppingbehaviourtoyourheartrate.Manybusinessprocessescanbedigitizedleadingtothecollectionofinformationtohelpreducecosts.Realdatainrealtime.Nosurveysrequiredtobecompletedbyauser,noresponsebias:justfacts.

FromDistributedComputingtoCloudComputing

Withallthispassivedatacollection,datasetshavebecomesomassiveandcomplexthattraditionaldataprocessingsoftwaresimplycannotkeepup.CompaniessuchasGoogle,whosebusinessisbasedondata,setaboutfindingawaytoincreasecomputationalpowerinordertoprocessBigData.Theanswer:DistributedComputing.Ratherthanhavingonesophisticatedcomputerhandlingalldataprocessing,thedataisdividedandsenttohun-dredsorthousandsofcommoditycomputersworkinginconcertbutindependentlythuscuttingdownoverallprocessingtime.

Thissolvedtheproblemforlargetechnologycompanies,butsmallerorganizationsstillhadnowaytoaccessandprocesslargevolumesofdataeffectively.EnterCloudComputing,

arentablenetworkofremoteservershostedontheInternettostore,manageandprocessdata.ThescalabilityofCloudComputinghaseliminatedtheneedtobuyandmaintaincostlycomputersystems;itprovidesaccesstomoreon-demandresourcesandprocessingpowerforeveryone,includingsmallorganizations.Asaresult,allorganizationscannowbenefitfromBigData.

AwordonDataWarehouses,DataLakesandDataSwamps

Howisallthisdatastored?Typically,viaaDataWarehouseoraDataLake.ADataWarehouseisdesignedtostoredataandprovidesbusinessesacentralrepositorytointegrate,manageandanalyzedataatmanylevels.Thedatacancomefrominternaland/orexternalsourcesandislargelystructured.ADataLakeisbothawaytoorganizehugevolumesofwildlydiversedata,bothstruc-turedandmostoftenunstructured,andstoreit.WhenaDataLakebecomesunmanageable,itturnsintoaDataSwamp.

WhenBigDataMeetsAI

ThebasicideabehindAIistoletamachinestatisticallyanalyzeallthedatabeingcollectedtoderiveinsightsmuchfasterandmoreaccuratelythanotherwisepossible.Forexample,ratherthanprogrammingacomputerwithasimplerulesuchas

“50timesthegrowthindataexpectedfrom2010to2020!”9

“morethan12hoursofdaylight=summer;lessthan12hoursofdaylight=winter,”AIanalyzestheparametersofnumeroussummerdays.Usingalgorithmicmodels,itwillthendeterminewhichparametersandatwhatlevelsthoseparametersconsti-tuteasummerday(e.g.,iftemperatureis>18degrees,humidityofx%,etc.).Itwillthenleveragethatunderstandingandbasedonprobabilitiesdeterminewhetheraparticulardayisasummerdayornotbyassessingallthoserelevantparametersagainstwhatithas“l(fā)earned”tobeatypicalsummerday.

ExpertsnowagreethatAIhasevolvedtoapointwhereithasreal,practicalvalue.Thereason?Theconvergenceoffourtrends.Theexplosivegrowthindataismadepossibleby:

digitizationofbusinessprocesses,smartphones,InternetofThings(IoT)andsocialmedia/web2.0

theadvancement(andavailability)inprocessingpowertostoreandcomputedata

thematurityofalgorithmsandAImodels

thehugeupsurgeininvestmentinAI.

/sfamilian/working-with-big-data-jan-2016-part-1/7-CONTEXT_WHATS_BIG_DATABIG_IN

Sowhatdoesallofthismeanforbusinesstoday?

Twoconcepts:improvedproductivityandinnovation.Theseincludesimplificationandautoma-tionoftasksthatfreeupemployeestofocusonmorestimulatingandhighervalue-addingwork.NewinnovativebusinessmodelshavebeenestablishedasaresultofAI,suchasvirtualassistantslikeAlexa,Googleassistantorx.ai.AlthoughnotpopularbeforeAI,theyarenow

amulti-billiondollarindustryatthecentreofthehomeautomationmarket.

Capital-intensivesectorssuchasmanufacturingandtransportationarelikelytoseethelargestproductivitygainsfromAI,giventhatmanyoftheirrepetitiveoperationalprocessesareripeforAI-enabledautomation.10AIrobotsarealreadytakingoverwarehousesacrosstheglobe.AIisalsopoisedtounlockefficienciesinprofessionalservicessuchaslegal,humanresources,accountingandauditwheredataplaysacentralrole.

/Previews/PWC/DocumentAssets/476830.pdf

Withinglobalaccountingandauditingfirms,workisunderwaytogetoutinfrontofandtakeadvantageofthenewparadigmemergingduetoAI.

BDOhassuccessfullypiloteditsBDOLexitranslationapp,usingneuralnetworktechnology,tomanageinformationinmultiplelanguagesduringglobalaudits.

DeloittehaslaunchedOmniaAI,amulti-disciplinarypracticefocusedondeliveryofAIservicesandsolutions.

EYisembeddingAItechnologyacrossitsend-to-endauditprocess,includ-ingtheuseofmachinelearninginEYHelixandEYOptixforpredictiveanalytics,usingdronesforinventoryobservationsandthedevelopmentofbusinessdocument-readingandinterpretationtools.

KPMGhaslauncheditsAItoolkitcalledIgniteandisworkingwithIBM/Watson.

PwCisinvestingindataplatformstosecurelycapture,organizeandfacili-tateanalysisofdataandisworkingwithH2O.ai,aleadingSiliconValleycompany,todevelopAudit.ai,whichwillprovideenhancedcapabilitiestoprovideassurance.

TypesofAI

ThereareessentiallytwokindsofAI:NarrowAIandGeneralAI.Asthenamesuggests,Nar-rowAI,whichcanalsobedescribedasweakAI,ismadeupofnarrowlyintelligentsystemsthatcanexceedhumansinspecifictasks,suchasplayingchessormakingmedicaldiagnoses.Thesenarrowcapabilitiesarenottransferrable(i.e.,anAIchessplayercannotbeusedtoper-formanothertasksuchasamedicaldiagnosis).

GeneralAIorstrongAIreferstohuman-levelintelligencethatisabletotransferknowledgebetweendomains.WhileNarrowAIisallaroundusinlanguageandvisionrecognitionsystemsandrecommendationengines,GeneralAImaybethestuffofsciencefictionandmovies—fornow.

HowAIWorks

InorderforAIprogramstonavigatethroughsituationalcomplexities,differentapproachestocreatingsoftware,withtheabilitytodeterminedifferentoutcomes,arenecessary.

TheLogicandRules-basedapproachusesconditionalinstructionsanddefinedrulestocarryoutataskorsolveaproblemsuchas“ifthis,thenthat.”IthasbeeninpracticeforalongtimeandhasbeentheunderlyingpremiseforAIuntilrecentadvancesinmachinelearninganddeeplearning,whicharetechniqueswithinAI.

Forexample,tocreatesystemstocalculatetaxthroughadigitizeddecision-makingprocess,theLogicandRulesapproachemploysadetailedseriesofcomputerinstructions,includingif/then/elseorprobabilities/weights.ThesedetailedseriesofinstructionsareattheheartofanyAIsystemandareknownasalgorithms.

MachineLearningisthemoreprevalentAIapproachusingalgorithmstoguideitspredic-tions.Itsnameisderivedfromtheabilityofalgorithmsto“l(fā)earn”fromexperience(i.e.,usedatasets)ratherthanrelyingonarules-basedsystem.Thealgorithmscreatecomputationalmodelsthatprocesslargedatasetstopredictoutputsandmakeinferences.Moredataleadstomoreexamples;thishelpsthealgorithmtofinelytuneitsoutputovertime.Inthisway,thealgorithmadjusts—or“l(fā)earns”—bytrialanderror.Eachnovelexampleisanoppor-tunityforthealgorithmtoguesscorrectlyorincorrectlyandlearnfromitsmistakesthususuallyprovidingbetterinsights.Theinsightsarethenfedbacktorefinethealgorithmicmodelsfurtherandmakethemmoreaccurateovertime.Machinelearningisinuseinemailspamfilters,text,imageandvoicerecognitionsystems,searchrankings,spellcheck,andmanymoreapplications.

ACloserLookatMachineLearning

Machinelearningmodelscanbroadlybegroupedintotwotypesofproblems:regressionandclassification.

Regressionproblemsinvolvethepredictionofaquantityorvalue(e.g.,thetaxifaretogofromSantaMonicatodowntownLAduringrushhouronFridaymorning).

Classificationproblemsinvolveclassifyinginput(e.g.,determiningwhetheratransactionistypicaloranomalous).

Threedifferenttechniquesaremostcommonlyusedforamachineto“l(fā)earn”theproblemandbecomesmartatprovidingtheanswer:

SupervisedLearningisamethodtoteachAIsystemsbyexample.Thesystemsareprovidedwithmanypointsofdata.Eachpointistiedtoexpectedoutcomessothatanunderstandingcanstarttodevelopofhowdatarelatestotheexpectedoutcome.Oncetrained,thesystemscantheningestpointsofdataandgiveanoutputthatfitsthelearnedmodel.

Takeaclassificationproblemasanexample.Inthecaseofratingthecreditworthinessofloanapplicants,theAIsystemisprovidedwithdataonthousandsofpastloanapplications,eachonehavingdetailedinformationabouttheprospectiveborrower(i.e.,creditscore,income,maritalstatus,etc.).Becausethedataishistoric,theAIsystemisalsosuppliedwithinformationonwhichloansdefaultedandwhichwererepaidsuc-cessfully.TheAIsystemthenusesstatisticaltechniquestofigureoutasetofrulestodeterminewhenaloanseemstohavegonebadandwhenitdidnot.

Theserulescannowbeappliedtogaugetheprobabilityofanewloanapplicationbeingsuccessfullyrepaidversusnot.Boththeinputandoutputtrainingdataarelabeledsothatoncetrained,theAIsystemcanthenapplywhatithaslearnedtonewdata.Thisprocessshowsthatdecisionsarenotbasedon100%accuracybutonprob-abilitiesinstead.IBMWatson’sdebutontheTVshowJeopardyisanexampleofthis.Whenrespondingtoquestions,Watsoncalculatedprobabilitiesofdifferentanswersandrepliedwiththehighestprobabilityanswer.Thisapproachiscalledsupervisedlearningbecauseanaccuratelyclassifiedtrainingdatasetisusedtosupervisethelearningprocess.

UnsupervisedLearning:Insupervisedlearningthesystemistoldwhichloanswererepaidsuccessfullyandwhichoneswerenotandaskedtoidentifywhattrendsmostaffectedtheoutcome.Ifonlyrawinsightsintothedataarewanted,thecomputerisaskedtofindinsightseffectivelythroughunsupervisedlearning.

Forexample,Netflixmayaskitssystemtofindothercustomerswhoseviewinghabitsaresimilartoyours.Itcanthenusethoseresultstorecommendmoviesthosesimilarcustomersliked.HeretherearenotraineddatasetssuchasaGoodLoanandBadLoanlabel;itisjustaboutlookingatdatatofindclustersofcustomerswhoare“simi-lar,”eventhoughthehumancustomersdidnotspecifywhattheymeantby“similar.”Detectingfraudandanomalyintransactionsoftenfallsinthisrealmbecauseyouareessentiallylookingfortransactionsthatare“nottypical.”

ReinforcementLearningisatechniquebywhichanAIsystemlearnsunderitsownsupervisionwithouttheneedoftraineddatasets.Itmakespredictions,validatesthosepredictionsagainstrealityandcontinuallyadjustsitselftohaveabetteroutputnexttime.Itisinspiredbybehaviouristpsychologyandreferstogoalorreward-orientedalgorithms.

Algorithmsaretrainedtoperformtasksbymaximizingtherewardsorpointsfortheexpectedfavourableoutcomeandcontinuetoexplorenewpossibilities,essentiallythroughtrialanderror,untilthatmaximizedrewardisfound.Forexample,rewardscouldbelinkedtoincreasedvaluesinstockpicking,evaluatingtradingstrategies,pointsinasimulationgameorbeatinganopponentinanadversarialgame.

Moreover,inDeepMind’sAlphaGoZerothesystemlearnedtoplaysimplybyplayinggamesagainstitself,startingfromcompletelyrandomplay.Itquicklysurpassedthehumanlevelofplayanddefeatedthepreviouschampion-defeatingversionofAlphaGoby100gamesto0.11

/blog/alphago-zero-learning-scratch/

Regressionproblemsaresolvedmostlywithsupervisedandreinforcementlearningwhileclassificationproblems,dependingonspecifics,maybeansweredbyusinganyoneofthethreelearningtechniques.

AIAlgorithms

AverageDailyTemp

IceCreamDollarsSpent

Therearedifferentstrategiesusedtoimple-mentthethreedifferenttechniquesdescribedabove.ThesestrategiesareattheheartofwhatallowstheAIsystemto“l(fā)earn”insightsfromthedata.AsaCPAyoumayneverhavetodealwithtechnologyatthislevel,butacursorylookmayprovidesomeappreciationofwhatAIalgorithmsdo.

MostlyBadLoans

MostlyGoodLoans

Applicant’sCreditScore

LoanAmount

Consideraverysimplisticexample.Sayyouhavedataonsalesperdayatanicecreamshopasafunctionofthedailyaverageoftemperature.AnAIsystemcouldidentifyaregressionequationthatprovidesthebestfitfortheprovideddatasetandusethattopre-dictsalesforthenextweekusingprojectedtemperatures.

Mathematically,theproblemisnowaboutfindingacurvethatfitsthegeometryofthecorrelationshownabove.IftheAIsystemfindstheparametersthatdefinedthecurve,itcanpredictanyY($)valuesforagivenX(temperature).

Inclassificationproblems,thecurvewouldtypicallydividetwoclasses(e.g.,goodloansandbadloans).

Source:PwCAnalysis

AIalgorithmsareessentiallyalldesignedtofindtheparametersthatdefinethecurve.

Ofcoursethesesimplecurvesareonlybasedononeortwodimensions.Themorecomplexandmultidimensionalthetask,themorecomplextheAIalgorithm.

WhatistheblackboxofAI?12

Asalgorithmsbecomemorecomplex,itisnotalwaysclearhoworwhyadecisionhasbeenmadebytheAIsystem.

Thelackofinterpretabilityandtransparencyofresultstohumanusers

oftheAIsystemcouldleadtoalackoftrustandlossofconfidenceintheAIsystem.

AIandtheHumanBrain:ATechnicalDiveintoDeepLearning

DeepLearning(alsoknownashierarchicallearning)isasubsetofmachinelearning.ItisanemergingandexcitingformofAIthatcanidentifyrelationshipsandlinkagesinvastvolumesofdatathatwouldbeimpossibleforhumanstoprocessandapplythemtosimilarsitua-tions.Whilestillintheearlystagesofdevelopment,DeepLearninghasthepotentialtotakeautomationtonewlevelsofeffectivenessandimprovedecisionmaking.Buthow?Byusingalgorithmsthatroughlyapproximatethestructuresandfunctionsofthehumanbrain.

Theideaistocreatealgorithmsthatcansimulateanarrayofneuronsinanartificialneuralnetworkthatlearnsfromvastsourcesofdata.Theinterconnectedlayersofnodesorneu-ronsprovidedatatoeachsuccessive

Insimpleterms,DeepLearningisthestackingofalgorithms(bothsupervisedandunsupervisedmodels)ontopofeachotherintomultiplelayers.Thisallowsforcreationofextremelysophisticatedandfinelytunedinput/outputmodels.

layerasanestedhierarchyofcon-ceptsmanylayersdeep.Theconnectionsbetweenthesevirtualneuronsarerepresentedbyanumberorweight.Theseweightsreflecttheimportanceattributedtotheinputdata;thegreatertheweightthemoreimportanttheinputistocarryingoutthedesiredtask.

/emerging-technology/to-open-ai-black-box

Duringthetrainingofneuralnetworks,weightsarevariedandadjusteduntilthedesiredout-comeisarrivedat.Forexample,asshownintheillustrationbelow,whenprocessingthousandsofimagesofhumanfaces,theneuralnetworkrecognizesobjectsbasedonahierarchyofsimplerbuildingblocks:

straightlinesandcurvedlinesatthebasiclevel

theneyes,mouths,andnoses

thenentirefacesand

finally,specificfacialfeatures.

Oncethefirstlevellearnsprimitivefeatures,thisdataisfedtothenextlayer,whichtrainsitselftorecognizemorecomplexfeatures.Theprocessisrepeatedinsuccessivelayersuntilthesystemcanreliablyrecognizeanoseorascar.Trainingthesedeeplearningsystemsrequiresvastamountsofdata;however,thelearningsfromthesesystemscansometimesbereplicatedtoothersituations.Forexample,aneuralnetworktrainedtorecogn