深度迁移度量学习.pptx

Deep Transfer Metric LearningJiang Taoran Li MingjiaAlgorithmsIntroductionExperiments1、Three learning methods2、Three shortcomings1、DTML2、DSTML1、Face Verification2、Person Re-IdentificationIntroduction此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑迁移学习度量学习此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑此页所有文字及形状均可修改与编辑深度学习关键词关键词关键词关键词Problemsseek a singlelinear distance132nonlinear relationshipexplicit nonlinear mapping functionsDeep Metric learningtraining andtest samplessameLinear feature spaceMetric learningkernel trickdeep modelsimilarity measureassumedistributions learn feature representationsTransfer learningDeep learningDSTMLDTMLh, W, bDTMLTwo constraints on the outputs of all training samples at the top of the network: 1) the inter-class variations are maximized and the intra-class variations are minimized, and 2) the distribution divergence between the source domain and the target domain at the top layer of the network is minimized.DMLConstructing a deep neural network to compute the representations of each sample x by passing it to multiple layers of nonlinear transformations.DMLDMLintra-class variations are minimizedinter-class variations are maximizedDMTLstochastic sub-gradient descent methodDMTLDMTLDMTLDMTLintra-class variations are minimizedinter-class variations are maximizedDeep learningTransfer learningDSTMLdiscriminativeinformation from the output of all layers.Face VerificationFace VerificationoutputWDRefLBPLBPLFWDTMLFace VerificationPerson Re-IdentificationPerson Re-Identificationcolor and texture histogramscolor and texture histogramsoutputSource domainTarget domainDTMLPerson Re-IdentificationPerson Re-Identification The nonlinear mapping function can be explicitly obtained： is a nonlinear activation function which operates component-wisely. For the first layer, we assume. For each pair of samplesand, their distan