资源描述：
《Accelerating Very Deep Convolutional》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、1AcceleratingVeryDeepConvolutionalNetworksforClassificationandDetectionXiangyuZhang,JianhuaZou,KaimingHey,andJianSunAbstract—Thispaperaimstoacceleratethetest-timecomputationofconvolutionalneuralnetworks(CNNs),especiallyverydeepCNNs[1]thathavesubstantiallyimpactedthecomputervisioncomm
2、unity.Unlikepreviousmethodsthataredesignedforapproximatinglinearfiltersorlinearresponses,ourmethodtakesthenonlinearunitsintoaccount.Wedevelopaneffectivesolutiontotheresultingnonlinearoptimizationproblemwithouttheneedofstochasticgradientdescent(SGD).Moreimportantly,whilepreviousmethod
3、smainlyfocusonoptimizingoneortwolayers,ournonlinearmethodenablesanasymmetricreconstructionthatreducestherapidlyaccumulatederrorwhenmultiple(e.g.,10)layersareapproximated.ForthewidelyusedverydeepVGG-16model[1],ourmethodachievesawhole-modelspeedupof4withmerelya0.3%increaseoftop-5err
4、orinImageNetclassification.Our4acceleratedVGG-16modelalsoshowsagracefulaccuracydegradationforobjectdetectionwhenpluggedintotheFastR-CNNdetector[2].IndexTerms—ConvolutionalNeuralNetworks,Acceleration,ImageClassification,ObjectDetectionF1INTRODUCTIONtheacceleratednetworksasgenericfeatu
5、reextractorsforotherrecognitiontasks[2],[12]remainunclear.TheaccuracyofconvolutionalneuralnetworksItisnontrivialtospeedupwhole,verydeepmodels(CNNs)[3],[4]hasbeencontinuouslyimproving[5],forcomplextaskslikeImageNetclassification.Acceler-[6],[7],[1],[8],butthecomputationalcostoftheseat
6、ionalgorithmsinvolvenotonlythedecompositionnetworksalsoincreasessignificantly.Forexample,theoflayers,butalsotheoptimizationsolutionstotheverydeepVGGmodels[1],whichhavewitnesseddecomposition.Data(response)reconstructionsolversgreatsuccessinawiderangeofrecognitiontasks[9],[17]basedonst
7、ochasticgradientdescent(SGD)and[2],[10],[11],[12],[13],[14],aresubstantiallyslowerbackpropagationworkwellforsimplertaskssuchthanearliermodels[4],[5].Real-worldsystemsmayascharacterclassification[17],butarelesseffectivesufferfromthelowspeedofthesenetworks.ForforcomplexImageNetmodels(a
8、swewilldiscussedexa
