深度学习及其应用_ 课件 0405自监督-预任务.pptxVIP

Self-supervisedLearning:pretasklearning

SupervisedandUnsupervisedLearningSJTUDeepLearningLecture.2SupervisedLearningUnsupervisedLearning

Self-supervisedLearningPretexttasksContrastivebasedLearningGenerativeReconstructionbasedLearning2015-2022-Populartime

UnsupervisedVisualRepresentationLearningbyContextPrediction(ICCV2015Oral)CarlDoersch,AbhinavGupta,AlexeiA.EfrosExperimentshowsthelearnedsemanticinformationcanusedforclassificationinalinearevaluation.Twoconvolutionalnetworksaresharedweights.Self-supervisedLearningwithPretextTask

UnsupervisedLearningofVisualRepresentationsbySolvingJigsawPuzzles(ECCV2016)MehdiNoroozi,PaoloFavaro1.randomlycropa225×225pixelwindowfromanimage(reddashedbox)2.divideitintoa3×3grid,andrandomlypicka64×64pixeltilesfromeach75×75pixelcell3.9tilesarereorderedviaarandomlychosenpermutationfromapredefinedpermutationsetandarethenfedtotheCFN4.predicttheindexofthechosenpermutationSelf-supervisedLearningwithPretextTaskJigsawPuzzles

UnsupervisedRepresentationLearningbyPredictingImageRotations(ICLR2018)SpyrosGidaris,PraveerSingh,NikosKomodakisSelf-supervisedLearningwithPretextTask

BoostingSelf-SupervisedLearningviaKnowledgeTransfer(CVPR2018)MehdiNoroozi,AnanthVinjimoor,PaoloFavaro,HamedPirsiavashSelf-supervisedLearningwithPretextTask

S4L:Self-SupervisedSemi-SupervisedLearning(ICCV2019)XiaohuaZhai,AvitalOliver,AlexanderKolesnikov,LucasBeyerCombinationofsemi-supervisedlearningandself-supervisedlearningObjectivefunctionIncludetwotermsRotationLossEntropyLossTomakethemodelmorerobusttoadversarialperturbations.ThefirsttermontherightistoreducetheKLdivergence,whichcanbewrittenas:Thesecondtermisastandardcross-entropyloss(byclustering-basedpseudolabels):Self-supervisedLearningwithPretextTask

Summary:DesignthePretextt