Deconvolutional Networks(解卷积网络).pptx

Deconvolutional NetworksMatthew D. ZeilerDilip Krishnan, Graham W. TaylorRob FergusDept. of Computer Science, Courant Institute, New York UniversityMatt ZeilerOverviewUnsupervised learning ofmid and high-level image representationsFeature hierarchy built from alternating layers of:Convolutional sparse coding (Deconvolution) Max poolingApplication to object recognitionMotivationGood representations are key to many tasks in visionEdge-based representations are basis of many modelsSIFT [Lowe’04], HOG [Dalal Triggs ’05] othersYan Huang (Winner of PASCAL 2010 classification competition)Felzenszwalb, Girshick, McAllester and Ramanan, PAMI 2007Beyond Edges? Mid-level cuesContinuationParallelismJunctionsCorners“Tokens” from Vision by D.MarrHigh-level object parts:Two ChallengesGrouping mechanismWant edge structures to group into more complex formsBut hard to define explicit rulesInvariance to local distortionsCorners, T-junctions, parallel lines etc. can look quite differentTalk OverviewSingle layerConvolutional Sparse CodingMax PoolingMultiple layersMulti-layer inferenceFilter learningComparison to related methodsExperimentsTalk OverviewSingle layerConvolutional Sparse CodingMax PoolingMultiple layersMulti-layer inferenceFilter learningComparison to related methodsExperimentsRecap: Sparse Coding (Patch-based)Over-complete linear decompositionof input using dictionaryInputDictionary regularization yields solutionswith few non-zero elementsOutput is sparse vector:Single Deconvolutional LayerConvolutional form of sparse codingSingle Deconvolutional LayerSingle Deconvolutional LayerSingle Deconvolutional Layer1Single Deconvolutional Layer1Single Deconvolutional Layer1Single Deconvolutional Layer1Top-down DecompositionSingle Deconvolutional Layer1Single Deconvolutional Layer1Toy ExampleFeaturemapsFiltersObjective for Single Layermin z:= Input, = Feature maps, = FiltersSimplify Notation::Filters are parameters of model (shared across all images)Fe