slide_Large linear classification when data cannot fit in memory

《slide_Large linear classification when data cannot fit in memory》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、LargeLinearClassicationWhenDataCannotFitInMemoryHsiang-FuYuDepartmentofComputerScienceNationalTaiwanUniversityJointworkwithC.-J.Hsieh,K.-W.Chang,andC.-J.LinJuly27,2010Hsiang-FuYu(NationalTaiwanUniv.)June02,20101/35OutlineIntroductionABlockMinimizationFrameworkforLinearSVMsImplementationsforSVMTe

2、chniquestoReducetheTrainingTimeOtherFunctionalitiesExperimentsConclusionsHsiang-FuYu(NationalTaiwanUniv.)June02,20102/35OutlineIntroductionABlockMinimizationFrameworkforLinearSVMsImplementationsforSVMTechniquestoReducetheTrainingTimeOtherFunctionalitiesExperimentsConclusionsHsiang-FuYu(NationalTa

3、iwanUniv.)June02,20103/35LinearClassicationRecentlylinearclassicationisapopularresearchtopicBylinearwemeanthatkernelisnotusedThoughlinearmaynotbeasgoodasnonlinearforsomeproblems:accuracybylinearisasgoodasnonlinear,andtrainingandtestingaremuchfasterThistalkaddressesonlargelinearclassicationHsia

4、ng-FuYu(NationalTaiwanUniv.)June02,20104/35MotivationExistingapproachesforlargelinearclassication:Datasmallerthanmemory:Ecientmethodsarewell-developedDatabeyonddisksize:UsuallyhandledinadistributedwayCanwehavesomethinginthebetween?Asimplesettingmemory

5、od,butonlyfordatawith#features#instancesHsiang-FuYu(NationalTaiwanUniv.)June02,20105/35WhenDataCannotFitInMemoryLIBLINEARonamachinewith1GBmemory:DiskswapcauseslengthytrainingtimeHsiang-FuYu(NationalTaiwanUniv.)June02,20106/35TheGoalGoal:constructlargelinearclassiersforordinaryusersonasinglemach

6、ineAssumptionsmemory

7、eedtopayattentiontothesecondpartLoadingtimemaydominatethetrainingtimeevendatacantinmemory>./liblinear-1.51/trainrcv1_test.binaryrcv1test.binary:>halfmillionsofdocumentsLoadingtime:>1minuteComputingtime:<5secondsHsiang