Data Workflows for Machine Learning Presentation.pdf

《Data Workflows for Machine Learning Presentation.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、DataWorkflowsforMachineLearning!PacoNathan@pacoidhttp://liber118.com/pxn/Whyisthistalkhere?MachineLearninginproductionappsislessandlessaboutalgorithms(eventhoughthatworkisquitefunandvital)!Performingrealworkismoreabout:!socializingaproblemwithinanorganization!featureengineering(“BeyondProductM

2、anagers”)!tournamentsinCI/CDenvironments!operationalizinghigh-ROIappsatscale!etc.!SoI’lljustcrawloutonalimbandstatethatleveraginggreatframeworkstobuilddataworkflowsismoreimportantthanchasingafterdiminishingreturnsonhighlynuancedalgorithms.!BecauseInterwebs!DataWorkflowsforMachineLearningMiddlewa

3、rehasbeenevolvingforBigData,andtherearesomegreatexamples—we’llreviewseveral.Processhasbeenevolvingtoo,rightalongwiththeusecases.!PopularframeworkstypicallyprovidesomeMachineLearningcapabilitieswithintheircorecomponents,oratleastamongtheirmajorusecases.!Let’sconsiderfeaturesfromEnterpriseDataWo

4、rkflowsasabasisforwhat’sneededinDataWorkflowsforMachineLearning.!Theirrequirementsforscale,robustness,costtrade-offs,interdisciplinaryteams,etc.,serveasguidesingeneral.CaveatAuditorIwon’tclaimtobeexpertwitheachoftheframeworksandenvironmentsdescribedinthistalk.Expertwithafewofthemperhaps,butmoret

5、othepoint:embroiledinmanyusecases.!Thistalkattemptstodefinea“scorecard”forevaluatingimportantMLdataworkflowfeatures:what’sneededforusecases,compareandcontrastofwhat’savailable,plussomeindicationofwhichframeworksarelikelytobebestforagivenscenario.!Seriously,thisisaworkinprogress.Outline•Definition

6、:MachineLearning•Definition:DataWorkflows•Awholebuncho’examplesacrossseveralplatforms•Ninepointstodiscuss,leadinguptoascorecard•BecauseNotebooks•Questions,comments,flyingtomatoes…DataWorkflowsforMachineLearning:Framethequestion…!“ABasisforWhat’sNeeded”Definition:MachineLearning“Machinelearningalgo

7、rithmscanfigureouthowtoperformimportanttasksbygeneralizingfromexamples.Thisisoftenfeasibleandcost-effectivewheremanualprogrammingisnot.Asmoredatabecomesavailable,moreambitiousproblemscanbetackled.Asaresult,machinelearningiswidelyusedinco