Deep Residual Learning for Image Recognition.pdf

《Deep Residual Learning for Image Recognition.pdf》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、DeepResidualLearningforImageRecognitionKaimingHeXiangyuZhangShaoqingRenJianSunMicrosoftResearchfkahe,v-xiangz,v-shren,jiansung@microsoft.comAbstract202056-layerDeeperneuralnetworksaremoredifficulttotrain.We20-layerpresentaresiduallearningframeworktoeasethetraining10105

2、6-layerofnetworksthataresubstantiallydeeperthanthoseusedtesterror(%)trainingerror(%)20-layerpreviously.Weexplicitlyreformulatethelayersaslearn-0001234560123456ingresidualfunctionswithreferencetothelayerinputs,in-iter.(1e4)iter.(1e4)steadoflearningunreferencedfunctions

3、.Weprovidecom-Figure1.Trainingerror(left)andtesterror(right)onCIFAR-10prehensiveempiricalevidenceshowingthattheseresidualwith20-layerand56-layer“plain”networks.Thedeepernetworkhashighertrainingerror,andthustesterror.Similarphenomenanetworksareeasiertooptimize,andcanga

4、inaccuracyfromonImageNetispresentedinFig.4.considerablyincreaseddepth.OntheImageNetdatasetweevaluateresidualnetswithadepthofupto152layers—8deeperthanVGGnets[41]butstillhavinglowercomplex-greatlybenefitedfromverydeepmodels.ity.Anensembleoftheseresidualnetsachieves3.57%

5、errorDrivenbythesignificanceofdepth,aquestionarises:IsontheImageNettestset.Thisresultwonthe1stplaceonthelearningbetternetworksaseasyasstackingmorelayers?ILSVRC2015classificationtask.WealsopresentanalysisAnobstacletoansweringthisquestionwasthenotoriousonCIFAR-10with100an

6、d1000layers.problemofvanishing/explodinggradients[1,9],whichThedepthofrepresentationsisofcentralimportancehamperconvergencefromthebeginning.Thisproblem,formanyvisualrecognitiontasks.Solelyduetoourex-however,hasbeenlargelyaddressedbynormalizedinitial-tremelydeepreprese

7、ntations,weobtaina28%relativeim-ization[23,9,37,13]andintermediatenormalizationlayersprovementontheCOCOobjectdetectiondataset.Deep[16],whichenablenetworkswithtensoflayerstostartcon-residualnetsarefoundationsofoursubmissionstoILSVRCvergingforstochasticgradientdescent(S

8、GD)withback-&COCO2015competitions1,wherewealsowonthe1stpropagation[22].placesonthetasksofImageNetdetection,ImageNetlocal-Whe