Deep Learning & Feature Learning Methods for Vision

Slide 1

Deep Learning & Feature Learning Methods for VisionRob Fergus (NYU)Kai Yu (Baidu)MarcAurelio Ranzato (Google)Honglak Lee (Michigan)Ruslan Salakhutdinov (U. Toronto)Graham Taylor (University of Guelph)CVPR 2012 Tutorial: 9am-5:30pmTutorial Overview9.00am:IntroductionRob Fergus (NYU)10.00am:Coffee Break10.30am:Sparse CodingKai Yu (Baidu)11.30am:Neural NetworksMarcAurelio Ranzato (Google)12.30pm:Lunch1.30pm:Restricted Boltzmann Honglak Lee (Michigan)Machines 2.30pm:Deep BoltzmannRuslan Salakhutdinov (Toronto) Machines 3.00pm:Coffee Break3.30pm:Transfer Learning Ruslan Salakhutdinov (Toronto)4.00pm:Motion & Video Graham Taylor (Guelph)5.00pm:Summary / Q & AAll5.30pm:EndOverviewLearning Feature Hierarchies for VisionMainly for recognition

Many possible titles:Deep LearningFeature LearningUnsupervised Feature Learning

This talk: Basic concepts Links to existing vision approaches

Existing Recognition ApproachHand-designedFeature ExtractionTrainableClassifierImage/VideoPixelsFeatures are not learned

Trainable classifier is often generic (e.g. SVM)ObjectClassSlide: Y.LeCunMotivationFeatures are key to recent progress in recognition

Multitude of hand-designed features currently in useSIFT, HOG, LBP, MSER, Color-SIFT.

Where next? Better classifiers? Or keep building more features?

Felzenszwalb, Girshick, McAllester and Ramanan, PAMI 2007

Yan & Huang (Winner of PASCAL 2010 classification competition)5What Limits Current Performance?Ablation studies on Deformable Parts Model Felzenszwalb, Girshick, McAllester, Ramanan, PAMI10

Replace each part with humans (Amazon Turk):

Also removal of part deformations has small (