four moves to machine learning at scale
TRANSCRIPT
FourMovestoScalableMachineLearningRobertWelborn,USAA
Movefrommonolithstodistributedcompute• Buildsthebasisforaccesstotoolsacrosstheenterpriseinsteadofonappliances
• Pointyouranalyticinfrastructureatoperationaldatainsteadofanalyticdata
• Transactionaldatadrivesyourmarketingmodels
MovefromBatchtoRealTime• Thelatencyisn’ttheproblem,thebatchphilosophyistheproblem• Processwhentheinformationisavailable,notonaschedule,thespeedwillgetbetterasyourprocessesimprove• Modelsthathaveworkedforyearswillneedtoberetooled(scoringcan’ttakeallnight)
MovefromStaticCoefficientstoResponsive
• Createtheabilityforfastfeedbackacrosstheenterprise
• The“learningloop”themodellearnsfromtheresultsofthelastexperiment
• Ifyouareinaregulatedindustry,youhavetobringyourcompliancefolksalongforthisjourney
• Humansmakedecisionsaboutwhattotestnotwhattoshowwhom
MovefromonthePremisestotheCloud• Don’treinventscaling,scalingwasdesignedforthecloud,atbest,youaretryingtodothingsbetterthanpeoplewhoalreadysolvedtheproblem,atworst,you’re“scaling”unsustainably• Allocatetherighttasksinthecloudandonpremises