Predictive Learning：深度学习与人工智能的未来.pptx

Y LeCunPredictive Learning1st Connectionist Summer School, CMU July 1986Y LeCun1st Connectionist Summer School, CMU July 1986Y LeCunBartlett MelYann LeCunStan DehaeneGerry TesauroMike MozerRichard DurbinBart SelmanAndy BartoJordan PollackDana BallardMike JordanJim HendlerDave TouretzkyDavid Willshaw Jay McClellandTerry SejnowskiGeoff HintonY LeCunSupervised LearningWe can train a machine on lots of examples of tables, chairs, dog, cars, and peopleBut will it recognize table, chairs, dogs, cars, and people it has never seen before?PLANECARCARDeep Learning = The Entire Machine is TrainableY LeCunTraditional Pattern Recognition: Fixed/Handcrafted Feature ExtractorFeature ExtractorTrainable ClassifierMainstream Modern Pattern Recognition: Unsupervised mid-level featuresFeature ExtractorMid-Level FeaturesTrainable ClassifierDeep Learning: Representations are hierarchical and trainedHigh-LevelFeaturesLow-Level FeaturesMid-Level FeaturesTrainable ClassifierY LeCunDeep Convolutional Nets for Object Recognition1 to 10 billion connections, 10 million to 1 billion parameters, 8 to 20 layers.Deep Learning = Learning Hierarchical RepresentationsY LeCunIts deep if it has more than one stage of non-linear feature transformationLow-Level FeatureMid-Level FeatureHigh-Level FeatureTrainable ClassifierFeature visualization of convolutional net trained on ImageNet from [Zeiler Fergus 2013]Very Deep ConvNet ArchitecturesY LeCunSmall kernels, not much subsampling (fractional subsampling).VGGGoogLeNetResNetConvNet for DrivingY LeCun(DARPA LAGR program 2005-2008)[Hadsell et al.,J. of Field Robotics 2009]Input imageStereo LabelsClassifier OutputInput imageStereo LabelsClassifier OutputImage captioning, Semantic Segmentation with ConvNetsY LeCun[Farabet et al. ICML 2011][Farabet et al. PAMI 2013][Lebret, Pinheiro, Collobert 2015][Kulkarni 11][Mitchell 12][Vinyals 14][Mao 14][Karpathy 14][Donahue 14]...Driving Cars with Convolutional NetsY LeCunMobilEyeNVIDIADeepMask: ConvNet Locates and