Speeding up Convolutional Neural Networks

《Speeding up Convolutional Neural Networks》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、JADERBERG,VEDALDI,ZISSERMAN:SPEEDINGUPCONVOLUTIONALNEURAL...1SpeedingupConvolutionalNeuralNetworkswithLowRankExpansionsMaxJaderbergVisualGeometryGroupmax@robots.ox.ac.ukDepartmentofEngineeringScienceAndreaVedaldiUniversityofOxfordvedaldi@robots.ox.ac.ukOxfo

2、rd,UKAndrewZissermanaz@robots.ox.ac.ukAbstractThefocusofthispaperisspeedinguptheapplicationofconvolutionalneuralnet-works.Whiledeliveringimpressiveresultsacrossarangeofcomputervisionandma-chinelearningtasks,thesenetworksarecomputationallydemanding,limitingt

3、heirde-ployability.Convolutionallayersgenerallyconsumethebulkoftheprocessingtime,andsointhisworkwepresenttwosimpleschemesfordrasticallyspeedinguptheselayers.Thisisachievedbyexploitingcross-channelorfilterredundancytoconstructalowrankbasisoffiltersthatarerank-

4、1inthespatialdomain.Ourmethodsarearchitectureag-nostic,andcanbeeasilyappliedtoexistingCPUandGPUconvolutionalframeworksfortuneablespeedupperformance.Wedemonstratethiswitharealworldnetworkde-signedforscenetextcharacterrecognition[15],showingapossible2.5speed

5、upwithnolossinaccuracy,and4.5speedupwithlessthan1%dropinaccuracy,stillachievingstate-of-the-artonstandardbenchmarks.1IntroductionManyapplicationsofmachinelearning,andmostrecentlycomputervision,havebeendis-ruptedbytheuseofconvolutionalneuralnetworks(CNNs).T

6、hecombinationofanend-to-endlearningsystemwithminimalneedforhumandesigndecisions,andtheabilitytoefficientlytrainlargeandcomplexmodels,haveallowedthemtoachievestate-of-the-artperformanceinanumberofbenchmarks[10,14,19,33,37,38].However,thesehighper-formingCNNsc

7、omewithalargecomputationalcostduetotheuseofchainsofseveralconvolutionallayers,oftenrequiringimplementationsonGPUs[16,19]orhighlyoptimizeddistributedCPUarchitectures[40]toprocesslargedatasets.Theincreasinguseofthesenet-worksfordetectioninslidingwindowapproac

8、hes[9,28,33]andthedesiretoapplyCNNsinreal-worldsystemsmeansthespeedofinferencebecomesanimportantfactorforappli-cations.Inthispaperweintroduceaneasy-to-implementmethodforsignificantlyspeedinguppre-traine