资源描述：
《GrCAN Gradient Boost ConvolutionalGrCAN：梯度提升卷积 具有神经决策森林的自动编码器》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、JOURNALOFLATEXCLASSFILES,VOL.14,NO.8,AUGUST20151GrCAN:GradientBoostConvolutionalAutoencoderwithNeuralDecisionForestManqingDong,LinaYao,XianzhiWang,BoualemBenatallah,andShuaiZhangAbstract—Randomforestanddeepneuralnetworkaretwoschoolsofeffectiveclassificationmethodsinmachinelearning.Whiletheran
2、domforestisrobustirrespectiveofthedatadomain,deepneuralnetworkhasadvantagesinhandlinghighdimensionaldata.Inviewthatadifferentiableneuraldecisionforestcanbeaddedtotheneuralnetworktofullyexploitthebenefitsofbothmodels,inourwork,wefurthercombineconvolutionalautoencoderwithneuraldecisionforest,wh
3、ereautoencoderhasitsadvantagesinfindingthehiddenrepresentationsoftheinputdata.Wedevelopagradientboostmoduleandembeditintotheproposedconvolutionalautoencoderwithneuraldecisionforesttoimprovetheperformance.Theideaofgradientboostistolearnandusetheresidualintheprediction.Inaddition,wedesignastruc
4、turetolearntheparametersoftheneuraldecisionforestandgradientboostmoduleatcontiguoussteps.Theextensiveexperimentsonseveralpublicdatasetsdemonstratethatourproposedmodelachievesgoodefficiencyandpredictionperformancecomparedwithaseriesofbaselinemethods.IndexTerms—Gradientboost,Convolutionalautoen
5、coder,Neuraldecisionforest.F1INTRODUCTIONACHINElearningtechniqueshaveshowngreatpowerisnotdifferentiable,researcherstrytotransformittoaMandefficacyindealingwithvarioustasksinthepastdifferentiableoneandadditintoneuralnetworks.Atypicalfewdecades[20].Amongexistingmachinelearningmodels,workisporpo
6、sedbyJohanneset.al[16],whopointoutRandomforest[41]anddeepneuralnetwork[50]aretwothatanydecisiontreescanberepresentedasatwo-layerpromisingclassesofmethodsthathavebeenprovensuc-ConvolutionalNeuralNetwork(CNN)[40],wherethefirstcessfulinmanyapplications.Randomforestistheensemblelayerincludesthede
7、cisionnodesandthesecondlayerleafofdecisiontrees[41].Itcanalleviatethepossibleoverfittingnodes.Theythenimplementacastrandomforest,whichofdecisiontreestotheirtrainingset[27]andthereforeisperformsbetterthantraditionalrandomforestandneuralrobustirrespec
