资源描述：
《Towards Convolutional Neural Networks Compression via Global Error Reconstruction》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、ProceedingsoftheTwenty-FifthInternationalJointConferenceonArtificialIntelligence(IJCAI-16)TowardsConvolutionalNeuralNetworksCompressionviaGlobalErrorReconstructionShaohuiLin1,2,RongrongJi1,2⇤,XiaoweiGuo3,XuelongLi41FujianKeyLaboratoryofSensingandComputingforSmartCity,XiamenUniversity,361005
2、,China2SchoolofInformationScienceandEngineering,XiamenUniversity,361005,China3BestImage,TencentTechnology(Shanghai)Co.,Ltd,China4Xi’anInstituteofOpticsandPrecisionMechanics,ChineseAcademyofSciences,Xi’an,Chinalinshaohui007@126.com,rrji@xmu.edu.cn,scorpioguo@tencent.com,xuelongli@ieee.orgAbs
3、tract1998;SimonyanandZisserman,2014;C.SzegedyandRa-binovich,2015;ZeilerandFergus,2014;Y.JiaandDar-Inrecentyears,convolutionalneuralnetworksrell,2014;K.HeandSun,2015],objectdetection[R.Gir-(CNNs)haveachievedremarkablesuccessinvar-shickandMalik,2014;K.HeandSun,2014]andimagere-iousapplications
4、suchasimageclassification,ob-trieval[Y.GongandLazebnik,2014].Despitethelonghis-jectdetection,objectparsingandfacealignment.toryofneuralnetworkresearchintheliterature[Fukushima,SuchCNNmodelsareextremelypowerfultodeal1980],thesignificantsuccessofCNNsismainlydrivenbywithmassiveamountsoftrainingd
5、atabyusingmil-theadvancedcomputingresourcesavailablenowadays.Forlionsandbillionsofparameters.However,theseinstance,totrainadiscriminativeCNNmodellikeAlexNetmodelsaretypicallydeficientduetotheheavycost[A.KrizhevskyandHinton,2012]orVGGNet[Simonyaninmodelstorage,whichprohibitstheirusageonandZis
6、serman,2014],itistypicallyrequiredtosethun-resource-limitedapplicationslikemobileorem-dredsofmillionsofparameters,whicharetunedusingmas-beddeddevices.Inthispaper,wetargetatcom-sivelabelledorunlabelleddataviaapproximatedoptimiza-pressingCNNmodelstoanextremewithoutsig-tion(e.g.stochasticgradi
7、entdescent)throughGPUordis-nificantlylosingtheirdiscriminability.Ourmaintributedsettings[J.DengandLi,2009].Tothateffect,vari-ideaistoexplicitlymodeltheoutputreconstructionousimplementationsofCNNsareintroducedintheliterature,errorbetweentheoriginalandcompr
