THINKPIECE

Opportunitiesfor

ArtificialIntelligence

InMediaMeasurement

November2024

CONTENTS

Introduction3

AboutCIMM

4

AbouttheAuthor5

TheBuildingBlocksofAI

6

BestFitApplicationsforAIandML

10

UsingAIforMediaMeasurement1

7

PuttingAItoWork2

4

ChallengestoAdoption2

7

Conclusions3

0

Introduction

TherehasbeenmuchexcitementabouttheapplicationofAItomanyfields

overthelast18months,andmedia

measurementisnoexception.Inthispaper,Ihaveconsultedacrosstheindustryand

beyondtounderstandthecurrentuseofAItechnologies,bothplannedandinproductiontoday.

Ihavefoundmanyexcitingandinnovativeapplicationsthatarelikelytocontinuetodrivetheevolutionofmediameasurement.Thispaperwillprovideapractical,

in-depthintroductiontothearrayofAI

technologiesformediaresearchersandanappraisalofthecurrentstateofthe

art,likelyfutureinnovations,andriskstoconsiderwhendeployingAIinyourmediameasurementsolutions.

–TomWeiss,author

●

3

AboutCIMM

TheCoalitionforInnovativeMediaMeasurementisanon-partisan,

pan-industrysubsidiaryoftheAdvertisingResearchFoundation,focusedon

cultivatingandsupportingimprovements,bestpracticesandinnovationsin

measurementandcurrency,data

collaborationandenablement,andthe

useofnewmetricsandapproachesto

understandingthevalueofmedia.CIMMembracestheentiremediaandadvertisingecosystemandprioritizeseffective

collaborationtodelivermeaningfulchange.Tofindoutmore,visit

●

4

orcontact

infro@

.

AbouttheAuthor

TomWeissisatechnologist,datascientist,andserial

entrepreneurwithover30yearsofexperienceacross

mediaandtechnology.CurrentlyservingasChief

TechnologyOfficeratRun3TVandBoardMemberatMX8Inc.,TomhasbeenattheforefrontofAI-drivenconsumerresearchandsmartbroadcastinginnovations.Hehas

foundedandledmultipleglobaltechnologyventures,

developedindustry-leadingdatascienceteams,and

holdsseveralpatentsinmediaandadvertising.WithanMAinPhysicsfromOxfordUniversity,Tomblendsdeeptechnicalexpertisewithapassionforhelpingbrands

betterunderstandandengagewiththeiraudiences.Youcanfollowhimonsubstackat

Tomwouldliketothankthefollowingwhohaveall

providedinputandinspirationforthispaper:

?AndyBrownfromtheAttentionCouncil

?AndyPrincepfromMarketCast

?AshwinNavinfromSamba

?BrianWestfromNBCU

?ChrisWilsonofHyphametrics

?EmilyMcReynoldsofAdobe

?JimBisbeefromVanderbiltUniversity

?JohnBrauerfromComcast

?JonWattsofCIMM

?JustinEvansfromSamsungAds

?KeithSmithfromMSI

?KenWilburfromUCSanDiegoRadySchoolofManagement

?KevinKohnfromKinetiq

?MatthewMcGranaghanfromtheUniversityofDelaware

?MichaelVinsonofComscore

?NickNorthfromtheBBC

?PaulDonatofromtheARF

?YanLiufromTVision

Theviewsandopinionsexpressedinthisreportare

solelythoseoftheauthor,TomWeiss,anddonot

necessarilyreflecttheviewsorpositionsofanyother

individualorentity.Thisreportisintendedsolely

forinformationalpurposes.NeitherCIMMnorthe

authorsmakeanyrepresentationsastotheaccuracy

orcompletenessofanyinformationcontainedinthis

reportorinanyreportorwebsitelinkedtointhisreport.NeitherCIMMnortheauthorswillbeliableforany

errorsoromissionsinthisinformationorforanylosses,injuries,ordamagesincurredfromthedisplayoruseofthisinformation.

5

6

TheBuildingBlocksofAI

ThehistoryofAIdatestothemid-20th

century.Theconceptof“artificialintelligence”wasfirstcoinedin1956byJohnMcCarthy

attheDartmouthConference,markingthebirthofAIasanacademicdiscipline.

Keymilestonesincludethedevelopment

ofearlyprogramminglanguages,creating

thefirstneuralnetworkinthe1940s,and

machinelearninginthe1980s.The1990s

and2000ssawsignificantadvancements,

withthedevelopmentofmoresophisticatedalgorithmsandincreasedcomputational

power.NoteworthynewstoriesthatcapturedthepopularimaginationincludeIBM’s

DeepBluedefeatingchesschampionGarryKasparovin1997,theintroductionofdeeplearninginthe2000s,andthevictoryofAIsystemsincomplexgameslikeJeopardy!

andGo.Morerecently,AIfromGoogle’s

DeepMindhasmadebreakthroughsinseveralcomputationallyexpensiveareasofscience

earningitscreatoraNobelPrizeinChemistry,andOpenAI’sChatGPThaswowedmillionswithitshuman-likeresponses.

Asoftheendof2024,wenowhaveAIswithnaturallanguageprocessingabilitiesthat

canpassthebarexamintheUSAandsolveMathsandPhysicsproblemsatacomparableleveltoundergraduatestudents.Wecan

createrealisticvideosbasedonasingle

imageoratextdescriptionofthevideoandeditphotosandvideossimplybyaskingtheAIwhattodo.

Noneoftheseseemedfeasibleatthe

beginningof2021.Theirdevelopmenthas

driventhecreationof“Transformer”-basedmodels,whichwerefirstpostulatedin2017.Theyhavemassivelyincreasedthevolumeofdatawecanfeedintoaneuralnetwork-stylemodel.Witharoundaquarterofamillion

academicpapersbeingpublishedeachyear,novelapproacheswilllikelybedevelopedinthecomingyearstoincreasethesemodels’efficacyfurther.

BeforeweconsiderhowtoapplythesenewtechnologiestoAI,let’sreview

theirfundamentals.

TheBuildingBlocksofAI

WhatdowemeanbyAI?

ThereareseveralschoolsofthoughtwhenitcomestoAI.Initially,AIreliedonlogicandrulestomakedecisions.Inrecentyears,approachessuchasneuralnetworksand

geneticalgorithmsthatsimulatehowthehumanbrain

workshaveemerged.Machinelearning,asubsetofAI,

focusesondevelopingalgorithmstolearnfromandmakepredictionsondata.Deeplearning,asubsetofmachine

learning,useslayeredneuralnetworkstoanalyze

variousfactorsinlargeamountsofdata.ReinforcementlearningisanotherapproachwhereAIsystemslearn

tomakedecisionsbytrialanderror.Together,these

approachesformthefoundationofmodernAIresearchandapplication.

Forthispaper,wearetakingthebroadestpossibleviewofartificialintelligenceandconsideringittoincludeallareasofmachinelearningandstatisticalinference:

Figure1:ThedifferenttechnologiesinvolvedinAI

AIisfundamentallyaboutteachingcomputersto

learnfromdata,allowingthemtomakepredictionsor

decisionsautonomouslywithoutbeingprogrammedforeveryspecifictask.Thesemodelsimproveandadaptastheyarefedmoreinformation.Theabilitytoevolveandenhanceitsperformancewitheachnewpieceofdata

makesmachinelearningsoeffective.

Thefieldisbroadlycategorizedintoseveralareas,eachwithitsstrengthsandweaknesses.

SupervisedLearning

Supervisedlearningislikeateacherguidingastudentthroughalesson,wheretheanswersarealreadyknown,

andthetaskistolearnhowtoanswerthequestions

correctly.Insupervisedlearning,algorithmsaretrainedondatasetsthatincludeboththequestions(inputs)andthecorrectanswers(outputs).Theaimistoteachthe

algorithmtomakeaccuratepredictionsonnew,unseendata.It’slikelearningtorecognizewhetheramanorawomanissittinginfrontoftheTVbystudyinglabeledphotosandthentryingtoidentifythepersoninanewimage.Thisapproachpowerstechnologieslikevoice

recognitionandemailspamfilters.

●

7

TheBuildingBlocksofAI

UnsupervisedLearning

Unsupervisedlearningisclosertostudying

acomplextopicindependently,withouta

teacherorguidebook.Unsupervisedlearningdealswithunlabeleddata,strivingtouncoverhiddenpatternsorstructures.It’sakinto

observingagardenfullofplantsandgroupingthemintocategoriesbasedontheirfeatures,eventhoughyou’veneverbeentoldwhich

featuresaremostdistinctive.Thismethodisusedtodiscovercustomersegments

inmarketingordetectunusualpatternsindicatingfraud.

ReinforcementLearning

Reinforcementlearningislearningbydoing,wheretheconsequencesofyouractions

teachyouhowtobehaveinthefuture.Themodelknowstomakedecisionsbytryingdifferentactionsandseeingwhathappens,aimingtomaximizecumulativereward.It’sliketeachingachildtoplaychess;thechildlearnsthemovesthatbringthemclosertocheckmateovertime.Applications

includeautonomousdrivingandstrategicgame-playing.

Semi-Supervisedand

Self-SupervisedLearning

Theseapproachesblendsupervisedand

unsupervisedlearningelements.Theyare

helpfulwhenthere’sonlyasmallamountoflabeleddataandalargequantityofunlabeleddata.It’slikelearningtopaintwithsome

instructionsforafewtechniquesbutthen

exploringandexperimentingonyourown

tocreateamasterpiece.Thesemethods

helpimprovelearningefficiencyandare

particularlybeneficialinfieldswherelabeleddataisrareorexpensivetogather.

semi-supervisedlearningpredictscustomerbehaviorbyleveragingasmalldatasetof

labeleduserinteractions(e.g.,clicks,views,orpurchases)alongsidealargerpoolof

unlabeledwebtrafficdata.Similarly,

self-supervisedlearningtechniquescanbeemployedtoanalyzevideocontentoraudiostreams,identifyingpatternslikerecurring

themesorcommonviewerreactionswithoutneedingextensivemanualtagging.

Forinstance,indigitaladvertising,

8

NeuralNetworks

Neuralnetworksareinspiredbythestructureand

functionofthehumanbrain.Theyconsistoflayers

ofinterconnectednodesor‘neurons,’eachlayer

transformingitsinputdataintoamoreabstractand

compositerepresentation.The“deep”indeeplearningreferstothenumberoflayersthroughwhichthedataistransformed.Morelayersallowformorecomplex

learningandunderstanding,enablingthesenetworkstoidentifyandlearnpatternsindataatmultiplelevelsofabstraction.

Asimpleneuralnetworkislikeamachinewithaninput(likeaphoto),severalprocessesitgoesthrough(the

hiddenlayers),andanoutput(identifyingiftheimageisofacarorabicycle).Thesenetworksadjusttheir

innerworkingsbasedonthedatathey’refed,learningtoimprovetheiraccuracyovertime.It’slikelearningtorecognizedifferenttypesofcarsbylookingatmanypicturesandgettingbetterwitheachone.

Whenweaddmorelayers,we’vemadedeepneural

networks.Thesecanpickapartandunderstanddatainincrediblydetailedways.Forexample,inrecognizingaface,onelayermightfocusondetectingedges,anotheronshapes,andyetanotheronspecificfeatureslikeeyesorasmile.

TypesofneuralnetworksincludeConvolutionalNeural

Networks(CNNs)andRecurrentNeuralNetworks(RNNs).CNNsareexpertsinhandlingimages,breakingdown

visualdataintopiecesitcananalyzeandunderstand,layerbylayer.RNNsareallaboutsequences,like

sentencesortime-seriesdata,wheretheorderof

thingsmatters.Theyhaveamemory,allowingthemtouseinformationfromearlierinthesequencetomakedecisionslater.

Largelanguagemodels(LLMs)likeChatGPT,together

withmostofthelatestinnovationsinAI,havebeendrivenbytransformermodels,introducedina2017papercalledAttentionisallyouneed1.Theyarenotablebecause,

althoughtheyaresequentialmodels,theycanbetrainedinparallel,allowingmuchmoreextensivetrainingsets

thanwerepossiblewithpreviousneuralnetworksthatrequiredsequentialtraining.

Thestrengthofdeeplearninganditsneuralnetworks

aretheincrediblycomplexfiltersthroughwhichwepushdatatoextractinsightsandanswers.Thedownsideis

thattheinnerworkingsaresoopaquethat,muchlikethebrain,wecannotalwaysexplainwhytheygivetheresultstheyprovide.Inanindustrylikeaudiencemeasurement,wheretransparencyhasbeenkeyforsomanyyears,thisisasignificantbarriertotheuseofAI.

Inlargelanguagemodels,thisalsoleadsisthe

phenomenonof“hallucinations.”AhallucinationreferstoAIgeneratingincorrectorfabricatedinformationbyamodel,evenwhensuchresponsesseemplausibleorconfident.Thisissueisparticularlyprevalentinnaturallanguageprocessing(NLP)modelslikeChatGPT.Forexample,anLLMmightgenerateareferencetoastudythatdoesn’texistorfabricatedetailsaboutahistoricalevent.Hallucinationsoccurbecausethemodeldoesn’ttruly“understand”thedata;itpredictsthemostlikelynextwordoroutputbasedonitstraining.

Thesemisstepscanpresentsignificantchallengesin

applicationslikemediameasurement,whereaccuracyandreliabilityareparamount.Addressingthisrequires

refiningtrainingprocesses,incorporatingbetter

constraints,andusingtechniqueslikehuman-in-the-

loopvalidationtoensurethemodelremainsgroundedinfactualdata.

Overfitting

ThemainpitfallintraininganyAImodelsisoverfitting.It’shardbecauseanoverfittedmodelhaspositivemetrics

supportingitsaccuracybutdoesn’tperforminthewild,andallbutthemostexperienceddatascientistswill

struggletoidentifywhenit’soccurring.

Thefundamentalproblemisthatthemodelistrained

sotightlyonitstrainingdatathatitcannotmakeany

predictionsoutsidethetrainingset.Ifyou’retryingto

teachamodeltoidentifydogphotos,you’dshowit

hundredsofimages,somewithdogsandsomewithout,andtellitwhichoneshavedogs.Ifthemodellearnstoidentifydogsbasedonstandardfeaturesofdogs(likehavingfur,fourlegs,andatail),itwilldowellwhenit

seesnewphotos.

Overfittingiswhenthemodelpaystoomuchattention

tounimportantdetails.Forexample,ifmanyofthedog

photosyoushowedweretakeninyourbackyard,itmightstartthinkingthat“beinginyourbackyard”isintegraltoidentifyingadog.Thismeansitwouldperformwellwithphotostakeninyourbackyard(eventhenon-dogones)becauseitmemorizedthisdetail.Itwould,however,

performpoorlywithphotostakenelsewherebecauseitfocusedtoomuchonthespecificsofthetrainingexamplesratherthanlearningthegeneralfeaturesofwhatmakesadogadog.

1

Vaswanietal.AttentionIsAllYouNeed,

/abs/1706.03762

●

9

BestFitApplicationsforAIandML

EachoftheAItechnologiescanbeappliedtodifferentapplications.Wecancategorizetheapplicationsinaverysimilarwaytohowwethinkaboutthetechnologies.Thissectionprovidesanoverviewoftheseapplications,

illustratingthediversewaysAIcanbeappliedtosolvereal-worldchallenges.

Figure2:ThedifferentAIusecases

Clustering

Clusteringisthetaskofgroupingasetofobjectssothatobjectsinthesamegrouparemorelikeeachotherthanthoseinothergroups.It’sdifferentfromclassificationinthatwedon’tknowwhatlabelswewantforeachgroup.

IfIwalkedintoagatheringataUnitedNations

delegation,Imightstartnoticinggroupsbasedonwherethey’refromwithoutaskingthemdirectly.Istartseeingpatterns-thelanguagespeoplespeak,thestylesof

clothing,perhapseventhefoodstheyprefer.Thistaskofsorting,basedonsimilaritiesyouobserve,isakintowhatwecallclusteringindatascience.It’saboutgrouping

objectssothatthosewithineachgrouparemorelikeeachotherthantheyaretoobjectsindifferentgroups.

Onewaytotackleourhypotheticalproblemwouldbe

theK-meansmethod.Inthiscase,wedecideupfront

thatwewanttoformacertainnumberofgroups(let’s

sayfive)andthenmovepeoplearounduntilwehavefivedistinctgroupswhereeveryoneisascloseaspossibleto

othersfromtheirregion.K-meansdoessomethingsimilarwithdatapoints,partitioningthemintokclustersbasedontheirfeatures,aimingforhighsimilaritywithinclustersanddifferencesbetweenthem.

Ifwedon’tknowupfronthowmanygroupswewant,

wecouldusehierarchicalclustering,whichisabit

likedrawingafamilytreeoftherelationshipsbetween

differentgroups.Insteadoflumpingpeopleintoseparategroups,youstartbyfindingthetwopeoplemostlike

eachotherandgroupthem.Then,youseethenext

closestpersonorgroup,andsoon,untileveryoneisincludedinabig,branchingfamilytreethatshowsnotjustthegroupsbuthowthey’rerelatedtoeachother.

Thismethodunderstandsnotjustthegroupingsbutthehierarchyorlevelsofsimilaritybetweenthem.

Density-basedclustering,suchasDBSCAN,istoday’smosteffectiveapproach.It’slikenoticingwherepeoplenaturallygatherinclustersatourparty.Somegroups

mightbetightlypacked,chattingindensecircles,whileothersmightbemorelooselygathered.

10

BestFitApplicationsforAIandML

Thismethodidentifiesclustersbasedondenselypackeddatapoints,allowingforgroupsofvariousshapesand

sizesandidentifyingdatapointsthatdon’tbelongtoanygroup(likesomeonesittingalone,perhaps).

Lastly,imagineifyoucouldguesswherepeoplewere

frombasedonspecificcharacteristics,likeassuming

someonemightbefromacoldercountryifthey’re

wearingheavywinterclothing.Model-basedclusteringoperatesunderasimilarprincipleifdatacomesfrom

differentsources(likecountries)andtriestofigureoutthesesourcesandtheircharacteristics.It’slikehavingasetofassumptions(models)aboutwherepeople

mightbefromandseeinghowwelleachpersonfitsthoseassumptions.

Inessence,clusteringhelpsusmakesenseofcomplexdatabyfindingnaturalgroupsorpatterns,muchlike

organizingadiversegatheringintogroupsbytheir

similarities.Whetherthroughsimplepartitioning,

drawingfamilytrees,spottinggenuinegatherings,ormakingeducatedguessesaboutorigins,eachmethodoffersawaytounderstandandsimplifytheworldof

dataaroundus.

FeatureExtraction

Featureextractiontransformsrawdataintoamoreusefulformatbyemphasizingthemostrelevantfeaturesor

patterns.Ratherthanfocusingonrawdetails,feature

extractionhelpsusdistillcomplexinformationintoits

essentialelements,makingiteasiertoanalyzeand

interpret.Thinkofitascreatingasummaryorfindingthekeypointsinalongbookratherthantryingtoremembereveryword.

Forinstance,dimensionalityreductionislikefocusingonlyonthemostimportantpagesofthatbook.When

workingwithlargedatasetswithmanyvariables,not

everypieceofdataisalwaysusefulforanalysis.Imagineyou’reobservingseveralpeopleandtrackingevery

minutedetail-fromtheirshoesizetotheirfavoritecolor

tothenumberoffrecklestheyhave.However,perhaps

onlyafewcharacteristics,likeheightandlanguage

spoken,arerelevantforgroupingthemorunderstandingtheirbehavior.Dimensionalityreductiontechniques,likePrincipalComponentAnalysis(PCA),helpuszeroinonthesefeatures,discardinglessvaluabledetailswithout

losingmuchinformation.It’saboutsimplifyingwithoutsacrificingtheessenceofthedata.

Associationrulesuncoverhiddenpatternsindatathatpointtorelationshipsbetweendifferentitems.Thinkofitasfiguringoutthatwhensomeonebuysbread,theyoftenalsobuybutter.Theserules,frequentlyusedinmarket

basketanalysis,helpusdiscoverconnectionsthatmightnotbeimmediatelyobviousbuthaveahighprobabilityofoccurringtogether.It’slikenoticingthatwhentwopeopleatourgatheringareconversinginthesamelanguage,

they’realsolikelydiscussingsimilartopics.ThefamousApriorialgorithmisapopularmethodtodiscovertheseassociations,helpinguspredictthelikelihoodofeventsbasedonpastdata,asiscommonlyseenwhenaretailerrecommendsproductsyoumightliketopurchase.

AnomalydetectionislikespottingtheonepersonatourUnitedNationsgatheringwho’snotquitefitting

in-perhapsthey’redressedunexpectedlyorbehavingdifferentlyfromtherest.Indatascience,anomalies

11

representrareorunusualpatternsthatstandoutfromthenorm,whichcouldindicateanythingfromfraud

BestFitApplicationsforAIandML

detectioninbankingtounusualmachine

behaviorinindustrialmonitoring.Techniquesforanomalydetection,suchasautoencodersorclustering-basedmethods,focuson

identifyingtheseoutliers,flaggingthemas

differentfromthetypicaldatapoints,much

likepickingouttheonepersonatapartywhoisn’tminglingintheusualway.

Inessence,featureextractionallows

ustosiftthroughcomplexdatatofind

what’sessential-reducingunnecessary

details,findingrelationships,oridentifyingoutliers.Thesetechniqueshelpussimplify,analyze,andultimatelyunderstanddata

moremeaningfully.Likeunderstanding

thefundamentaldynamicsatagathering,

featureextractionhighlightsthemostcriticalpatterns,connections,andodditiesinthe

datawe’reworkingwith.

Predictivemodels

Predictivemodelspredictcontinuous

outcomesbasedononeormorevariables.

It’stheAIequivalentofbeingaweather

forecaster,tryingtopredicttomorrow’s

temperatureorafinancialanalystforecastingstockprices.Aparticulartoolsetforthese

kindsoftasksisregressionmodels.Based

onpastdata,thesemodelsareallabout

predictingnumbers-liketemperatures,prices,oranycontinuousoutcome.

Weareallfamiliarwithlinearregression,themoststraightforwardapproach.Wedidthis

inhighschoolwhenwedrewastraightline

throughascatterofdotsonagraphtofitthedatabest.Thislinehelpsusunderstandandpredicttherelationshipbetweenthings-for

example,howsalesmightincreasewithmoreadvertisingspend.It’sthego-tomethodwhentherelationshipbetweenourdatapointsis

directanduncomplicated.

Buttypically,thingsaren’tsostraightand

simple.Real-worldrelationshipsoftencurveandtwistunexpectedly,andweneedto

applynon-linearregressionmodels.Thesecanfitcurvesandcomplexshapestofitthepeculiarpatternsofourdata.Whetherit’sthepredictivecustomerlifetimevalueorcreatingacreditrating,non-linearmodelshelpus

graspandpredicttheseintricatepatterns.

12

Timeseriesforecastingistheapplicationofnon-linearanalysistopredictthefuture.Byexaminingpatterns

overtime-suchastheebbandflowofoceantidesortheseasonalupsanddownsinhotelbookings-thisapproachhelpspredictwhatwillhappennext.

Lastly,wehaveaseriesofmodelsthatareoptimized

nottopredicttheoutcomebuttopredicttheprobabilityoftheoutcomeoccurring.Thisisthefieldofrisk

assessmentandprediction.AIcanhelpinsurance

companies,healthcareproviders,andsafetyengineersmakeinformedpredictionsbyanalyzingpastaccidentsorhealthoutcomes.

Inessence,regressionmodelsinAIareaboutmakinginformedguessesaboutthefuturebasedonpast

patterns.Whetherdrawingstraightlines,sketchingcurves,peeringintothefuture,orcalculatingrisks,thesetoolshelpusconfidentlynavigateaworld

ofuncertainties.

ClassificationProblems

Classificationisafundamentaltaskinmachinelearning,separatingdataintopredefinedclasses.Theclosest

analogyissortingahugepileofphotographsintoalbumsbasedonwho’sinthem.Inmachinelearning,sorting-orclassifying-dataintopredefinedcategoriescanbedoneinseveralways.

First,there’sbinaryclassification.Thisisthesimple

problemofdecidingwhethersomethingisonethingoranother-likedeterminingwhetheraphotoisofacatoradog.Inthedigitalworld,thiscouldidentifywhether

anemailisspamorifamedicaltestindicatesadiseaseispresentorabsent.It’sastraightforward,either/

orsituation.

Thisbecomesharderinmulti-classclassification.

Insteadoftwocategories,wehaveawholespectrum-likesortingphotosintoalbumsforcats,dogs,birds,andfish.Eachimagegoesintojustonealbumbasedonwhat’sinit.Computersusethisapproachfortaskslikerecognizinghandwrittendigits(isthatsquigglea1,a2,ora9?)or

figuringoutthelanguageofatext.

Whenaphotocontainsacatandadog,multi-labelclassificationcomesintoplay,allowingeachpieceofdatatobelongtomultiplecategoriessimultaneously.

Forinstance,anewsarticlemightbetaggedas

simultaneouslyrelatingtopolitics,economics,andinternationalaffairs.

Finally,wehaveimbalancedclassification,wheresomecategoriesaremuchlesscommonthanothers.It’sa

significantchallengeinareaslikefrauddetection,wheretherareevent(afraudulenttransaction)isprecisely

whatneedstobeidentifiedamongaseaoflegitimate

transactions.Thechallengehereistoprovideenough

trainingdatafortherareeventstothemodelwithoutthemodellearningthattheseeventsaremorecommonthantheyare.

Inessence,classificationisaboutteachingamachinetorecognizepatternsandmakedecisions,whetherasimpleyes/no,pickingoneofmanyoptions,jugglingmultiple

labelsatonce,orspottingtherareneedleinahaystack.

DeepLearningApplications

Withitsabilitytoprocessandlearnfromvastamountsofdata,deeplearninghasmanyapplications.

Firstup,wehaveimageandvideorecognition.We’ve

alreadylookedatteachingacomputertorecognizeyourfaceorthedifferencebetweenacatandadog.Deep

learningmodels,particularlyConvolutionalNeural

Networks,havebecomeincrediblygoodatmaking

senseofpixelsandpatternstoidentifyobjectsand

actionsinimagesandvideos.Thistechnologypowers

everythingfromfacialrecognitiononyoursmartphonetoautomatedtaggingoffriendsinsocialmediaphotos.

NaturalLanguageProcessinginvolvesteaching

machinestounderstand,interpret,andrespondto

humanlanguagemeaningfullyandhelpfully.We’ve

seentremendousadvancesthankstodeeplearning,

particularlymodelslikeTransformers.Machinescannoweasilytranslatelanguages,gaugethesentimentbehindyoursocialmediaposts,orevensummarizelongarticles,makinginformationmoreaccessible.

Combiningthesemodels,deeplearninghas

revolutionizedspeechrecognitionandsynthesis.

Rememberwhenyouha