Statistical Learning Theory and ApplicationsLearning: Brains and MachinesLearning: math, engineering, neuroscienceClass9.520 Statistical Learning Theory and ApplicationsSlide 6Slide 7Overview of overviewSlide 9Learning from examples: goal is not to memorize but to generalize, eg predict.Binary classification caseReason to learn some learning theoryNotesSlide 14Slide 15Slide 16Thus….the key requirement (main focus of learning theory) to solve the problem of learning from examples: generalization (and possibly even consistency).Slide 18Thus….two key requirements to solve the problem of learning from examples: well-posedness and generalizationSlide 20Slide 21We have used a simple algorithm -- that ensures generalization -- in most of our applications…Slide 24Theory summarySlide 26Slide 27Learning from Examples: engineering applicationsSlide 29Slide 30Slide 31Another application: using learning algorithms to decrypt the brain codeGoal (analysis): Can we “read-out” the subject’s object percept?Slide 34Slide 35Recording at each recording site during passive viewingSlide 37Training a classifier on neuronal activity.Decoding the neural code … population response (using a classifier)Slide 40Slide 41Slide 42Slide 47Learning Object Detection: Finding Frontal Faces ...Learning Face DetectionFace detection:…Face identification: exampleSlide 52People classification/detection: training the systemSlide 54The system was tested in a test car (Mercedes)Slide 56A pedestrian detection system in a test car (Mercedes, ~10 years ago) is now becoming a product (MobilEye)Slide 58Slide 59Face classification/detection: training the systemFace identification: training the systemComputer vision: new StreetScenes Project Learning Algorithms for Scene UnderstandingSlide 63Slide 64Image AnalysisImage SynthesisReconstructed 3D Face Models from 1 imageSlide 68Slide 69Extending the same basic learning techniques (in 2D): Trainable Videorealistic Face Animation (voice is real, video is synthetic)Trainable Videorealistic Face AnimationSlide 73A Turing test: what is real and what is synthetic?Slide 80Slide 81Slide 82Some numbersSlide 85A theory of the ventral stream of visual cortexSlide 87The Ventral Visual Stream: From V1 to ITSlide 90The modelModel accounts for rapid categorization by humansSlide 94Slide 95Slide 96Slide 97Slide 98Slide 99Slide 100Slide 101Model’s early predictions: neurons become view-tuned during recognitionRecording Sites in Anterior ITThe Cortex: Neurons Tuned to Object Views as predicted by modelView-tuned cells: scale invariance (one training view only)!Predictions of the model for view tuned IT cells vs Logothetis’ data (using same stimuli)Slide 108Slide 109…when we compared its performance with machine vision…Sample Results on the CalTech 101-object datasetThe model performs at the level of the best computer vision systemsSlide 113Experiment: rapid (to avoid backprojections) animal detection in natural imagesRapid categorization (not real experiment)Targets and distractorsHumans achieve model-level performanceSlide 121Slide 1283. Limitations: recognition in “clutter”: recognition in multi-object (770 images becomes increasingly difficult for feedforward arhitecturesSome remarksThus…main msg of the talkSlide 133Slide 134Slide 135Slide 136Slide 137Slide 138Slide 139Beyond 50 ms: Feedback loops9.520, spring 2007Statistical Learning Theory and Statistical Learning Theory and Applications Applications Jake Bouvrie and Lorenzo Rosasco and Ryan Rifkin + tomaso poggio9.5209.520, spring 2007Learning: Brains and Learning: Brains and MachinesMachinesLearning is the gateway to understanding the brain and to making intelligent machines. Problem of learning: a focus for o modern math o computer algorithms o neuroscienceLearning theory+ algorithmsComputational Neuroscience: models+experimentsENGINEERING APPLICATIONS• Bioinformatics • Computer vision • Computer graphics, speech synthesis, creating a virtual actorHow visual cortex works – and how it may suggest better computer vision systemsLearning:math, engineering, neuroscience211min ( , ( ))i iKf HiV y f x fm�=� �+� �� ��llTheorems on foundations of learning:Predictive algorithms9.520, spring 2007ClassClassRules of the game: problem sets (2) final project (min = review; max = j. paper) grading participation! mathcamps? Monday late afternoon?Web site: http://www.mit.edu/~9.520/Slides on the Web siteStaff mailing list is [email protected] Student list is [email protected] Please fill form!9.520, spring 20039.520 Statistical Learning Theory and 9.520 Statistical Learning Theory and Applications Applications Class 25: Project presentations2:30—2:45 "Adaboosting SVMs to recover motor behavior from motor data", Neville Sanjana2:45-3:00 "Review of Hierarchical Learning", Yann LeTallec:3:00—3:15 "An analytic comparison between SVMs and Bayes Point Machines", Ashis Kapoor3:15-3:30 "Semi-supervised learning for tree-structured data", Charles Kemp:3:30—3:45 “Unsupervised Clustering with Regularized Least Square classifiers" - Ben Recht:3:40—3:50 "Multi-modal Human Identification." Brian Kim 3:50—4:00 "Regret Bounds, Sequential Decision-Making and Online Learning", Sanmay Das9.520, spring 20039.520 Statistical Learning Theory and 9.520 Statistical Learning Theory and Applications Applications Class 26: Project presentations2:35-2:50 "Learning card playing strategies with SVMs", David Craft and Timothy Chan2:50-3:00 "Artificial Markets: Learning to trade using Support Vector Machines“, Adlar Kim3:00-3:10 "Feature selection: literature review and new development'‘, Wei Wu3:10—3:25 "Man vs machines: A computational study on face detection" Thomas Serre9.520, spring 20079.520, spring 2007Overview of overviewOverview of overviewo The problem of so The problem of supervised learning: “real” math behind ito Examples of engineering applications (from our group)o Learning and the brain (example of object Learning and the brain (example of object recognition)recognition)Learning: much more than memory Role of learning (theory and applications in many different domains) has grown substantially in CS Plasticity and learning have a central stage in the neurosciences Until now math and engineering of learning has developed independently of neuroscience…but it may begin to change: we will see the example of learning+computer vision…Learning from examples: goal is not to