CVPR2012_Deep learning

《CVPR2012_Deep learning》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、CVPR2012TutorialDeepLearningMethodsforVision(draft)HonglakLeeComputerScienceandEngineeringDivisionUniversityofMichigan,AnnArbor1Featurerepresentationspixel1Learningalgorithmpixel2InputMotorbikesInputspace“Non”-Motorbikespixel2pixel12FeaturerepresentationshandleFeatureLearningrepre

2、sentationalgorithmwheelInputMotorbikesInputspace“Non”-MotorbikesFeaturespacepixel2“handle”pixel1“wheel”3Howiscomputerperceptiondone?State-of-the-art:“hand-crafting”FeatureLearningInputdatarepresentationalgorithmObjectdetectionImageLow-levelObjectdetectionvisionfeatures/classificat

3、ion(SIFT,HOG,etc.)AudioclassificationLow-levelAudioSpeakeraudiofeaturesidentification(spectrogram,MFCC,etc.)4ComputervisionfeaturesSIFTSpinimageHoGRIFTTextonsGLOH5ComputervisionfeaturesSIFTSpinimageHand-craftedfeatures:1.NeedsexpertknowledgeHoGRIFT2.Requirestime-consuminghand-tuni

4、ng3.(Arguably)oneofthelimitingfactorsofcomputervisionsystemsTextonsGLOH6LearningFeatureRepresentations•Keyidea:–Learnstatisticalstructureorcorrelationofthedatafromunlabeleddata–Thelearnedrepresentationscanbeusedasfeaturesinsupervisedandsemi-supervisedsettings–Knownas:unsupervisedf

5、eaturelearning,featurelearning,deeplearning,representationlearning,etc.•Topicscoveredinthistalk:–RestrictedBoltzmannMachines–DeepBeliefNetworks–DenoisingAutoencoders–Applications:Vision,Audio,andMultimodallearning7LearningFeatureHierarchy•DeepLearning3rdlayer–Deeparchitecturescanb

6、e“Objects”representationallyefficient.–Naturalprogressionfrom2ndlayerlowleveltohighlevel“Objectparts”structures.1stlayer–Cansharethelower-level“edges”representationsformultipletasks.Input8Outline•RestrictedBoltzmannmachines•DeepBeliefNetworks•DenoisingAutoencoders•ApplicationstoVi

7、sion•ApplicationstoAudioandMultimodalData9LearningFeatureHierarchy[Relatedwork:Hinton,Bengio,LeCun,Ng,andothers.]Higherlayer:DBNs(Combinationsofedges)Firstlayer:RBMs(edges)Inputimagepatch(pixels)11Note:Noexplicit“pooling.”Describemoreconcretely..RestrictedBoltzmannMachineswithbina

8、ry-valuedinputdata•Representation