UnderstandingAlTechnology

Aconcise,practical,andreadableoverviewofArtificialIntelligence

andMachineLearningtechnologydesignedfornon-technical

managers,officers,andexecutives

April2020

By:GregAllen,ChiefofStrategyandCommunications

JointArtificialIntelligenceCenter(JAIC)

DepartmentofDefense

ForewordbyJAICDirectorLtGenJackShanahan

Acknowledgments

Theauthorwouldliketothankthefollowingindividualsfortheirassistancereviewingearlierdrafts

ofthisdocument:

?Dr.JeffAlstott(IARPA)

?Dr.NateBastian(Major,U.S.Army,DoDJointAlCenter)

?Dr.StevenL.Brunton(UniversityofWashington)

?Dr.MatthewDaniels(GeorgetownUniversity)

?Dr.EdFelten(PrincetcnUniversity)

?Mr.RobJasper(PacificNorthwestNationalLaboratory)

?Dr.JohnLaunchbury(Galois,andformerlyDARPA)

Disclaimer

Theviewsexpressedinthisdocumentarethoseoftheauthoraloneanddonotnecessarilyreflect

thepositionoftheDepartmentofDefenseortheUnitedStatesGovernment.

J)

Website:/

Twitter:@DoDJAIC

Linkedln:/company/dod-ioint-artificial-intelliqence-center/

□A1CUnderstandingArtificialIntelligenceTechnology2

FOREWORDBYLIEUTENANTGENERALJACKSHANAHAN

ItishardformetodescribethesteepslopeofthelearningcurveIfacedwhenI

startedtheProjectMavenjourneyoverthreeyearsago.WhileinmanywaysIstill

considermyselfanArtificialIntelligenceneophytetoday,whatIknewaboutthe

subjectbackthencouldbarelyfillthefirstfewlinesofasinglepageinmytrusty

notebook.Myjourneyofdiscoverysincethenhasbeenchallenging,tosaythe

least.IonlywishGregAllen'sguideto"UnderstandingAlTechnology11hadbeen

availabletomeinlate2016asweembarkedonourfirstAI/MLpilotprojectforISR

full-motionvideoanalysis.

Greghasperformedaninestimableservicebywritingthisguide.Alischanging

nationalsecurity,andit'sessentialthatDoDleadershaveafirmgraspofthe

technology'sbuildingblocks.AsIlearnedbackin2017andamremindeddailyin

myroleastheDirectoroftheJointAlCenter(JAIC),Alisnotanelixir.Itisan

enabler-onethatiscriticaltoourfuturenationalsecurity.Itisimportantforallof

ustosharethesamefundamentalunderstandingofAltechnology.Greg'sguide

balancesbreadthanddepthinjusttherightway.Itisclear,concise,andcogent.

IamconfidentitwillbeavaluableresourceforeveryoneinDoDandbeyond.

LieutenantGeneralJohnN.T.“Jack”Shanahan

Director,JointArtificialIntelligenceCenter

DepartmentofDefense

April2020

□A1CGregoryC.Allen|DoDJointAlCenter2

□A1CUnderstandingArtificialIntelligenceTechnology3

EXECUTIVESUMMARY

ManyofficialsthroughouttheDepartmentofDefenseareaskedtomake

decisionsaboutAlbeforetheyhaveanappropriateunderstandingofthe

technology'sbasics.Thisguidewillhelp.

TheDoDAlStrategydefinesAlas“theabilityofmachinestoperformtasksthat

normallyrequirehumanintelligence."Thisdefinitionincludesdecades-oldDoD

Al,suchasaircraftautopilots,missileguidance,andsignalprocessingsystems.

ThoughmanyAltechnologiesareold,therehavebeenlegitimatetechnological

breakthroughsoverthepasttenyearsthathavegreatlyincreasedthediversityof

applicationswhereAlispractical,powerful,anduseful.Mostofthebreakthroughs

andexcitementaboutAlinthepastdecadehavefocusedonMachineLearning

(ML),whichisasubfieldofALMachineLearningiscloselyrelatedtostatisticsand

allowsmachinestolearnfromdata.

ThebestwaytounderstandMachineLearningAlistocontrastitwithanolder

approachtoAl,HandcraftedKnowledgeSystems.HandcraftedKnowledge

SystemsareAlthatusetraditional,rules-basedsoftwaretocodifysubjectmatter

knowledgeofhumanexpertsintoalongseriesofprogrammed“ifgivenxinput,

thenprovideyoutput"rules.Forexample,theAlchesssystemDeepBlue,which

defeatedtheworldchesschampionin1997,wasdevelopedincollaboration

betweencomputerprogrammersandhumanchessgrandmasters.The

programmerswrote(literallytypedbyhand)acomputercodealgorithmthat

consideredmanypotentialmovesandcountermovesandreflectedrulesfor

strongchessplaygivenbyhumanexperts.

MachineLearningsystemsaredifferentinthattheir“knowledge“isnot

programmedbyhumans.Rather,theirknowledgeislearnedfromdata:a

MachineLearningalgorithmrunsonatrainingdatasetandproducesanAl

model.Toalargeextent,MachineLearningsystemsprogramthemselves.Even

so,humansarestillcriticalinguidingthislearningprocess.Humanschoose

algorithms,formatdata,setlearningparameters,andtroubleshootproblems.

MachineLearninghasbeenaroundalongtime,butitpreviouslywasalmost

alwaysexpensiveandcomplicatedwithlowperformance,sotherewere

comparativelyfewapplicationsandorganizationsforwhichitwasagoodfit.

Thankstotheever-increasingavailabilityofmassivedatasets,massivecomputing

power(bothfromusingGPUchipsasacceleratorsandfromthecloud),open

sourcecodelibraries,andsoftwaredevelopmentframeworks,theperformance

andpracticalityofusingMachineLearningAlsystemshasincreaseddramatically.

TherearefourdifferentfamiliesofMachineLearningalgorithms,whichdiffer

basedonaspectsofthedatatheytrainon.Itisimportanttounderstandthe

differentfamiliesbecauseknowingwhichfamilyanAlsystemwillusehas

implicationsforeffectivelyenablingandmanagingthesystem'sdevelopment.

□A1CGregoryC.Allen|DoDJointAlCenter3

□A1CUnderstandingArtificialIntelligenceTechnology5

Purpose:Bynow,nearlyallDoDofficialsunderstandthattheriseofAlisan

importanttechnologytrendwithsignificantimplicationsfornationalsecurity,but

manystruggletogivesimpleandaccurateanswerstobasicquestionslike:

?WhatisAl?

?HowdoesAlwork?

?WhyisnowanimportanttimeforAl?

?WhatarethedifferenttypesofMachineLearning?Howdotheydiffer?

?WhatareNeuralNetworksandDeepLearning?

?WhatarethestepsofbuildingandoperatingAlsystems?

?WhatarethelimitationsandrisksofusingofAlsystems?

Contrarytopopularbelief,youdonotneedtounderstandadvanced

mathematicsorknowcomputerprogramminglanguagestobeabletoanswer

theabovequestionsaccuratelyandtodevelopapracticalunderstandingofAl

relevanttoyourorganization'sneeds.Thisguidewillcovereverythingthatthevast

majorityofDoDleadersneedtoknow.

WHATISAl?

TheDoDAlStrategystatesthat"Alreferstotheabilityofmachinestoperform

tasksthatnormallyrequirehumanintelligence."Thisdefinitionissoobviousthat

manyareconfusedbyitssimplicity.Infact,however,thisdefinitionisverysimilar

totheonesusedbymanyleadingAltextbooksandleadingresearchers.Thefirst

thingtonoteaboutthisdefinitionisthatAlisanextremelybroadfield,onethat

coversnotonlythebreakthroughsofthepastfewyears,butalsothe

achievementsofthefirstelectroniccomputersdatingbacktothe1940s.

Thedefinitionof"ArtificialIntelligence,11isabitofamovingtarget.When

somethingisnewandexciting,peoplehavenoqualmsaboutlabelingit**Artificial

Intelligence.nOncethecapabilitiesofaparticularAlapproacharefamiliar,

though,theyareoftencalledmerely“software.”Thispaperwillreturnlatertothe

subjectofhowmodernAlapproachesaredifferentandwhynowisacritical

momentforAltechnology.Fornow,justunderstandthatevenoldtechnology

canstillbeAl.

HOWDOESAlWORK?

Alisdefinedbyasetofcapabilities,ratherthanaspecifictechnicalapproachto

achievingthosecapabilities.Therearemanydifferentapproachestodeveloping

Alsystems,andvariousapproachesworkdifferentlywithdifferentstrengthsand

weaknesses.DARPA,alongtimepioneerinAlresearch,hashelpfullygrouped

manyoftheseapproachesintotwobroadcategories:(1)Handcrafted

Knowledgeand(2)MachineLearning.MachineLearningsystemsarethenewer

ofthetwoapproaches(thoughstilldecadesold)andareresponsibleforthe

dramaticimprovementsinAlcapabilitiesoverthepasttenyears.Ifyou'veheard

□A1CGregoryC.Allen|DoDJointAlCenter5

□A1CUnderstandingArtificialIntelligenceTechnology6

somepersonorcompanyclaimthattheirsystemMusesAI,Mmostlikelytheymean

thattheirsystemisusingMachineLearning,whichisafarcryfromtheirsystem

beinganautonomousintelligenceequaltoorgreaterthanhumanintellectinall

categories.Still,recentprogressinMachineLearningisabigdeal,with

implicationsfornearlyeveryindustry,includingdefenseandintelligence.The

easiestwaytounderstandMachineLearningsystems,however,isbycontrasting

themwithHandcraftedKnowledgesystems,sothispaperwillbeginthere.

HandcraftedKnowledgeAl

HandcraftedKnowledgeSystemsaretheolderofthetwoAlapproaches,nearly

asoldaselectroniccomputers.Attheircore,theyaremerelysoftwaredeveloped

incooperationbetweencomputerprogrammersandhumandomainsubject

matterexperts.HandcraftedKnowledgeSystemsattempttorepresenthuman

knowledgeintoprogrammedsetsofrulesthatcomputerscanusetoprocess

information.Inotherwords,the"intelligence"oftheHandcraftedKnowledge

Systemismerelyaverylonglistofrulesintheformof“ifgivenxinput,thenprovide

youtput/1Whenhundredsorthousandsormillionsofthesedomain-specificrules

arecombinedsuccessfully-into“theprogram”-theresultisamachinethatcan

seemquitesmartandcanalsobeveryuseful.

Awell-knownexampleofaHandcraftedKnowledgeAlSysteminwidespreaduse

istaxpreparationsoftware.Byrequiringuserstoinputtheirtaxinformation

accordingtopre-specifieddataformatsandthenprocessingthatdata

accordingtotheformallyprogrammedrulesofthetaxcode(developedin

cooperationbetweenhumansoftwareengineersandaccountants),theoutput

canbegoodenoughtopassanIRSaudit.Whenfirstintroducedinthe1980s,tax

preparationsoftwarewasverysuccessfullymarketedasArtificialIntelligence.

Nowthatithasbeeninwidespreadusefordecades,however,callingit“Art而cial

Intelligence“hasfallenoutoffashion.Nevertheless,itstillfallswithinboththeDoD

definitionofAlandtheformaldefinitionusedbymostresearchersinthefield.

AnotherfamousexampleofaHandcraftedKnowledgeSystemis“DeepBlue/*

theIBM-developed,chess-playingAlthatdefeatedthehumanworldchess

championin1997.DeepBluewasdevelopedincooperationbetweenIBM's

softwareengineersandseveralchessgrandmasters,whohelpedtranslatetheir

humanchessexpertiseintotensofthousandsofcomputercoderulesforplaying

grandmaster-levelchess.

BothtaxpreparationAlsystemsandDeepBlueareaspecifictypeofHandcrafted

KnowledgeAlknownasanExpertSystem.AnothertypeofHandcrafted

KnowledgeAlisaFeedbackControlSystem,whichuseshuman-authoredrulesto

computesystemoutputbasedonsensormeasurementinputs.FeedbackControl

SystemshavebeeninwidespreadusebytheDepartmentofDefensefor

decades.Aircraftautopilots,missileguidancesystems,andelectromagnetic

□A1CGregoryC.Allen|DoDJointAlCenter6

□A1CUnderstandingArtificialIntelligenceTechnology7

signalprocessingsystemsarejustafewofthethousandsofhigh-performing,

extremelyreliableFeedbackControlAlSystemsthattheDepartmentofDefense

anditspartnershavedevelopedandoperatedoverthepasteightdecades.In

thissense,HandcraftedKnowledgeSystemsarevictimsoftheirownsuccess.They

aresocommonthattheyaregenerallynolongerreferredtoas“Al”incommon

discourse.Nevertheless,HandcraftedKnowledgeSystemsremainimportantand

useful.Insomeareas,suchastaxpreparation,theystillhavefarhigher

performancethanMachineLearningsystems.Inotherareas,suchaschess

playing,languagetranslation,andimageclassification,HandcraftedKnowledge

AlsystemshavebeengreatlysurpassedinperformancebyMachineLearningAl

systems.Regardless,HandcraftedKnowledgesystemswillcontinuetoimprove

andseewideusefordecadestocome.

MachineLearningAl

ThekeydifferencebetweenaHandcraftedKnowledgeSystemandaMachine

Learningsystemisinwhereitreceivesitsknowledge.Ratherthanhavingtheir

knowledgebeprovidedbyhumansintheformofhand-programmedrules,

MachineLearningsystemsgeneratetheirownrules.ForMachineLearning

systems,humansprovidethesystemtrainingdata.Byrunningahuman-

generatedalgorithmonthetrainingdataset,theMachineLearningsystem

generatestherulessuchthatitcanreceiveinputxandprovidecorrectoutputy.

Inotherwords,thesystemlearnsfromexamples(trainingdata),ratherthanbeing

explicitlyprogrammed.ThisiswhydataissovitalinthecontextofALDataisthe

mainrawmaterialoutofwhichhigh-performingMachineLearningAlsystemsare

built.Forthisreason,thequality,quantity,representativenessanddiversityofdata

willdirectlyimpacttheoperationalperformanceoftheMLsystem.Algorithmsand

computinghardwarearealsoimportant,butnearlyallMLsystemsrunon

commoditycomputinghardware,andnearlyallofthebestalgorithmsarefreely

availableworldwide.Hence,havingenoughoftherightdatatendstobethekey.

Whileitistruethat-toalargeextent-MachineLearningsystemsprogram

themselves,humansarestillcriticalinguidingthislearningprocess:humans

choosealgorithmsanddatasets,formatdata,setlearningparameters,and

troubleshootproblems.

Atthisstage,manyreadersmayaskthemselves,Msowhat?WhyisMachine

Learningimportant?"

Thereasonisthattherearemanyapplicationsv/heretaskautomationwouldbe

useful,butwherehumanprogrammingofallofthesoftwarerulestoimplement

automationiseitherimpracticalorgenuinelyimpossible.Sometimeshuman

expertsareunabletofullytranslatetheirintuitiondecision-makingintofixedrules.

Further,forasurprisinglylargesubsetoftheseapplications,theperformanceof

□A1CGregoryC.Allen|DoDJointAlCenter7

□A1CUnderstandingArtificialIntelligenceTechnology8

MachineLearningsystemsisveryhigh,muchhigherthanwaseverachievedwith

HandcraftedKnowledgeSystemsorindeedbyhumanexperts.Thisdoesnot

meanthatHandcraftedKnowledgesystemsareobsolete.Formanyapplications

theyremainthecheapestand/orhighestperformingapproach.

Figure1:SimplifiedDiagramofAlApproaches

HandcraftedKnowtedgeAl

Operaftonai

InputData

MachineLearningAl

OperoAonai

InputOak)

MOCfWW

Loaming

Patternrecognition,imageanalysis,languagetranslation,contentgeneration,

andspeechtranscriptionarejustafewnoteworthyexampleswherethepast

performanceofHandcraftedKnowledgeAlwasverylow,buttheperformance

ofMachineLearningAlisextremelyhigh,oftenbetterthanhumanperformance.

BecauseoftheincreasedperformanceandenhancedproductivityMachine

Learningenables,therearemanypracticalapplicationsthroughoutthe

economyandindustry.WearenotafterusingAlforitsownsake.Weareafter

increasedperformanceandenhancedproductivity.Ifsthatsimple.

WHYISNOWANIMPORTANTTIMEFORAl?

Alhasbeenaroundfordecades.So,whyhaseveryonebeentalkingaboutit

constantlyinrecentyears?Itboilsdowntothis-forMachineLearningAlsystems

-therehasbeenamassiveincreaseinthenumberofreal-worldapplications

whereAlisnowpracticalandpowerful.Therearefourmainreasonswhythisis

truenowbutwasnottruetenyearsago:

□A1CGregoryC.Allen|DoDJointAlCenter8

□A1CUnderstandingArtificialIntelligenceTechnology9

1)MoreMassiveDatasets:MachineLearningalgorithmstendtorequirelarge

quantitiesoftrainingdatainordertoproducehighperformanceAlmodels.

Forexample,somefacialrecognitionAlsystemscannowroutinelyoutperform

humans,buttodosorequirestensofthousandsormillionsoflabeledimages

offacesfortrainingdata.WhenMachineLearningwasfirstdeveloped

decadesago,therewereveryfewapplicationswheresufficientlylarge

trainingdatawasavailabletobuildhighperformancesystems.Today,an

enormousnumberofcomputersanddigitaldevicesandsensorsare

connectedtotheinternet,wheretheyareconstantlyproducingandstoring

largevolumesofdata,whetherintheformoftext,numbers,images,audio,or

othersensordatafiles.

Ofcourse,moredataonlyhelpsifthedataisrelevanttoyourdesired

application.Ifyou'retryingtodevelopabetteraircraftautopilot,thena

bunchofconsumerloanapplicationdataisn'tgoingtohelp,nomatterhow

muchyouhave.Ingeneral,trainingdataneedstomatchthereal-world

operationaldatavery,verycloselytotrainahigh-performingAlmodel.

2)IncreasedComputingPower:ToafargreaterextentthanHandcrafted

KnowledgeSystems,MachineLearningAlsystemsrequirealotofcomputing

powertoprocessandstorealltheabove-mentioneddata.Aroundtenyears

ago,computinghardwarestartedgettingpowerfulenoughandcheap

enoughthatitwaspossibletorunMachineLearningalgorithmsonmassive

datasetsusingcommodityhardware.Oneespeciallyimportantturningpoint

around2010wasdevelopingeffectivemethodsforrunningMachineLearning

algorithmsonGraphicsProcessingUnits(GPUs)ratherthanontheCentral

ProcessingUnits(CPUs)thathandlemostcomputingworkloads.Originally

designedforvideogamesandcomputergraphics,GPUsarehighly

parallelized,whichmeanstheycanperformlargenumbersofsimilar

calculationsatthesametime.Itturnsoutthatmassiveparallelismisextremely

usefulinspeedingupthetrainingofMachineLearningAlmodelsandin

runningthosemodelsoperationally.FormanytypesofMachineLearning,

usingGPUscanspeedupthetrainingprocessby10-20xwhilereducing

computerhardwarecosts.Accesstothecloudisalsoveryhelpful,since

organizationscanrapidlyaccessmassivecomputingresourcesondemand

(fortherelativelyshortamountsoftimeneededfortraining)andlimit

purchasesofcomputingpowertoonlywhattheyneed,whentheyneedit.

3)ImprovedMachineLearningAlgorithms:ThefirstMachineLearningalgorithms

aredecadesold,andsomedecades-oldalgorithmsremainincrediblyuseful.

Inrecentyears,however,researchershavediscoveredmanynewalgorithms

thathavegreatlysharpenedthefield'scutting-edge.Thesenewalgorithms

havemadeMachineLearningmodelsmoreflexible,morerobust,andmore

capableofsolvingdifferenttypesofproblems.

□A1CGregoryC.Allen|DoDJointAlCenter9

□A1CUnderstandingArtificialIntelligenceTechnology10

4)OpenSourceCodeLibrariesandFrameworks:Thecutting-edgeofMachine

Learningisnotonlybetterthanever,butalsomoreeasilyavailable.Foralong

time,MachineLearningwasaspecializednichewithincomputerscience.

DevelopingMachineLearningsystemsrequiredalotofspecificexpertiseand

customsoftwaredevelopmentthatmadeitoutofreachformost

organizations.Now,however,therearemanyopensourcecodelibrariesand

developertoolsthatalloworganizationstouseandbuildupontheworkof

externalcommunities.Asaresult,noteamororganizationhastostartfrom

scratch,andmanypartsthatusedtorequirehighlyspecializedexpertisehave

beenlargelyautomated.ThedifficultyofdevelopinganAlmodelhasfallen

tothepointwhere-forsomeapplications-evennon-expertsandbeginners

cancreateusefulAltools.Insomecases,opensourceMLmodelscanbe

entirelyreused.

Figure2.KeyFactorsDrivingRecentImprovementsinMLPerformance

Inshort,usingMachineLearninggenerallyusedtobeexpensiveand

complicated,sotherewerecomparativelyfewapplicationsandorganizations

forwhichitwasagoodfit.Now,however,usingMachineLearningispractical

andpowerfulforafarmorediversesetofapplications.Thankstotheever-

increasingavailabilityofmoremassivedatasets,increasedcomputingpower,

improvedMachineLearningalgorithms,andimprovedopensourcecode

librariesandsoftwaredevelopmentframeworks,thingsthatusedtobenearly

impossible,suchasautomatedfacialrecognition,arenowpossible.Programs

thatusedtohaveterribleperformance,suchasautomatictranslation,nowhave

significantlybetterperformance.Finally,Alsystemsthatusedtobeextremely

□A1CGregoryC.Allen|DoDJointAlCenter10

□A1CUnderstandingArtificialIntelligenceTechnology11

expensivetodevelop,suchasimageryclassification,areoftennowaffordable

andsometimesevencheap.

Despitetheirhugepotential,Alsolutionsarenotagreatfitforalltypesof

problems.IfyouhaveanapplicationwhereyouthinkusingAlcouldbebeneficial,

knowingwhetherornotanyparticularsystemthatisclaimingtouse“Al”isusing

MachineLearningisimportantforseveralreasons.Foronething,Machine

Learningworksdifferentlyfromtraditionalsoftware,andithasdifferentstrengths

andweaknessestoo.Moreover,MachineLearningtendstobreakandfailin

differentways.Abasicunderstandingofthesestrengths,weaknesses,andfailure

modescanhelpyouunderstandwhetherornotyourparticularproblemsarea

goodfitforaMachineLearningAlsolution.

WHATARETHEDIFFERENTTYPESOFMACHINELEARNING?HOWDOTHEYDIFFER?

LikeArtificialIntelligence,MachineLearningisalsoanumbrellaterm,andthere

arefourdifferentbroadfamiliesofMachineLearningalgorithms.Therearealso

manydifferentsubcategoriesandcombinationsunderthesefourmajorfamilies,

butagoodunderstandingofthesefourbroadfamilieswillbesufficientforthe

vastmajorityofDoDemployees,includingseniorleadersinnon-technicalroles.

Thefourcategories-Figure3.LabeledandUnlabeledTrainingData

explainedmoreon

thefollowingLabeledDataUnlabeledData

differbasedonwhat

typesofdatatheir

algorithmscanwork

with.However,the

importantdistinction

isnotwhetherthe

dataisaudio,images,

text,ornumbers.

Rather,theimportant

distinction

iswhetherornotthetrainingdataislabeledorunlabeledandhowthesystem

receivesitsdatainputs.Figure3providesasimpleillustrationoflabeledand

unlabeledtrainingdataforaclassifierofimagesofcatsanddogs.

Dependinguponwhetherornotdataislabeled,adifferentfamilyofalgorithms

applies.ThefourmajorfamiliesofalgorithmsareSupervisedLearning,

UnsupervisedLearning,Semi-SupervisedLearning,andReinforcementLearning.

SupervisedLearning:"Supervised”meansthat；beforethealgorithmprocesses

thetrainingdata,some,supervisor),(whichmaybeahuman,groupofhumans,

oradifferentsoftwaresystem)hasaccuratelylabeledeachofthedatainputs

□A1CGregoryC.Allen|DoDJointAlCenter11

□A1CUnderstandingArtificialIntelligenceTechnology12

withitscorrectassociatedoutput.Forexample,ifthegoaloftheAlsystemisto

correctlyclassifytheobjectsindifferentimagesaseither“cat”or“dog,"the

labeledtrainingdatawouldhaveimageexamplespairedwiththecorrect

classificationlabel.SupervisedLearningsystemscanalsobeusedforidentifying

thecorrectlabelsofcontinuousnumericaloutputs.Forexample,"giventhiswing

shapeinput,predicttheoutputairdragcoefficient.”

ManySupervisedLearningsystemscanachieveextremelyhighperformance,but

theyrequireverylargelabeleddatasetstodoso.Usingimageclassificationasan

example,acommonruleofthumbisthatthealgorithmneedsatleast5,000

labeledexamplesofeachcategoryinordertoproduceanAlmodelwithdecent

performance.Acquiringallofthislabeleddatacanbeeasyorverydifficult,

dependingupontheapplication.Inthecaseoffacialrecognitionalgorithms,

mostcompaniesusepaidhumanstomanuallylabelimages.Inthecaseofonline

shoppingrecommendationengines,thecustomersareactuallyprovidingthe

datalabelsthroughthenormalcourseoftheirshopping.Thedatainputsarethe

recommendeditemsdisplayedtothecustomersandthecustomer'sprofile

information,whiletheoutputsaretheactualpurchasesmadeornotmade.This

isoneofthemajorreasonswhyinternetcompanieswereattheforefrontofthe

adoptionofMachineLearningAl:theiruserswereconstantlyproducingvaluable

datasets-bothlabeledandunlabeled-andtheonlineenvironmentallowedfor

rapidexperimentationwithMachineLearning-enabledanalysisandautomation.

Notethatpre-labeleddataisonlyrequiredforthetrainingdatathatthealgorithm

usestotraintheAlmodel.TheAlmodelinoperationalusewithnewdatawillbe

generatingitsownlabels,theaccuracyofwhichwilldependontheAl'straining.

Ifthetrainingdatasetwassufficientlylarge,highquality,andrepresentativeof

thediversitypresentintheoperationalenvironment,thentheperformanceofthe

Almodelingeneratingtheselabelscanbeatorabovehumanperformance.

UnsupervisedLearning:Unsupervisedalgorithmsarethosethatcanextract

featuresfromthedatawithouttheneedforaground-truthlabelfortheresults.

Usingtheaforementionedexampleofanimageclassifier,theAlmodelproduced

byanunsupervisedalgorithmwouldnotreturnthataspecificinputimagewasof

a“cat”ora"dog."Rather,themodelwouldsortthetrainingdatasetintovarious

groupsbasedontheirsimilarity.Onesortedgroupmightbethedesiredgroupsof

catsanddogs,butimagesmightinsteadbesortedbasedonundesired

categoriessuchaswhetherornottheyhaveablueskyinthebackgroun