Statistical Learning Theory and ApplicationsLearning: Brains and MachinesLearning: much more than memoryLearning: math, engineering, neuroscienceClass9.520 Statistical Learning Theory and Applications (2007) 9.520 Statistical Learning Theory and Applications (2003) Overview of overview Learning from examples: goal is not to memorize but to generalize, eg predict.Binary classification caseReason to learn some learning theorySlide Number 13Slide Number 14Slide Number 15 Thus….the key requirement (main focus of classical learning theory) to solve the problem of learning from examples: generalization Slide Number 17 Thus….two key requirements to solve the problem of learning from examples: well-posedness and generalization. How are they related?Slide Number 20We have used a simple algorithm -- that ensures generalization -- in most of our applications…Slide Number 23Winning against the curse of dimensionality: new research directions in learningSlide Number 25Slide Number 26What are the principles of learning from few data in high dimensional spaces?Slide Number 28Slide Number 29Overview Overview of overviewLearning from Examples: engineering applicationsSlide Number 34Slide Number 35Learning from Examples: engineering applicationsObject recognition for computer vision: (personal) historical perspectiveExamples: Learning Object Detection: Finding Frontal FacesLearning Object Detection: Finding Frontal Faces ...Learning Face DetectionFace detection:…Slide Number 43The system was tested in a test car (Mercedes)Slide Number 45~10 year old CBCL computer vision work: SVM-based pedestrian detection system in Mercedes test car… now becoming a product (MobilEye)Slide Number 47People classification/detection: training the systemFace classification/detection: training the systemFace identification: training the systemWhat about the model and computer vision? The street scene projectSlide Number 52Learning from Examples: engineering applicationsAnother application: using learning algorithms to decrypt the brain codeGoal (analysis): Can we “read-out” the subject’s object percept?Slide Number 56Slide Number 57Recording at each recording site during passive viewingSlide Number 59 Training a classifier on neuronal activity.Decoding the neural code … population response (using a classifier)Slide Number 62Slide Number 63Slide Number 64Learning from Examples: engineering applicationsImage AnalysisImage SynthesisReconstructed 3D Face Models from 1 imageReconstructed 3D Face Models from 1 imageSlide Number 70Extending the same basic learning techniques (in 2D): Trainable Videorealistic Face Animation (voice is real, video is synthetic)Trainable Videorealistic Face AnimationSlide Number 74A Turing test: what is real and what is synthetic?Overview of overviewSlide Number 77Slide Number 78Some numbersThe problem: recognition in natural images (e.g., “is there an animal in the image?”)How does visual cortex solve this problem? How can computers solve this problem?Slide Number 82A “feedforward” version of the problem: rapid categorization A model of the ventral stream, which is also an algorithm… …”solves” the problem (if the mask forces feedforward processing)… Extensive comparison w| neural dataSlide Number 87Slide Number 88Formalizing the hierarchy: towards a theoryIt is just possible that the brain ….Statistical Learning Theory and ApplicationsLorenzo Rosasco + Jake Bouvrie +Ryan Rifkin + Tomaso PoggioMcGovern Institute for Brain ResearchCenter for Biological and Computational LearningDepartment of Brain & Cognitive SciencesMassachusetts Institute of TechnologyCambridge, MA 02139 USA9.250Learning: Brains and MachinesLearning is the gateway to understanding the brain and to making intelligent machines. Problem of learning: a focus for o modern matho computer algorithmso neuroscienceLearning: 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 in the class the situation in vision…Learning theory+ algorithmsComputationalNeuroscience: 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 ( , ( ))iiKfHiVy fx fμ∈=⎡⎤+⎢⎥⎣⎦∑llTheorems on foundations of learning:Predictive algorithmsClassRules of the game: problem sets (2)final project (min = review; max = j. paper)gradingparticipation!Web site: http://www.mit.edu/~9.520/Slides on the Web siteStaff mailing list is [email protected] Student list will be [email protected] fill form!9.520 Statistical Learning Theory and Applications (2007) Class 26: Project presentations (past examples)10:30 - Simon Laflamme “Online Learning Algorithm for Structural Control using Magnetorheological Actuators”- Emily Shen “Time series prediction”- Zak Stone “Facebook project”- Jeff Miller “Clustering features in the standard model of cortex”- Manuel Rivas "Learning Age from Gene Expression Data“- Demba Ba “Sparse Approximation of the Spectrogram via Matching Pursuits: Applications to Speech Analysis”- Nikon Rasumov "Data mining in controlled environment and real data"9.520 Statistical Learning Theory and Applications (2003) Class 26: Project presentations (past examples)2:35-2:50 "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 SerreOverview of overviewo o The problem of sThe problem of supervised learning: “real” math behind ito Examples of engineering applications (from our group)o Learning and the brain Learning and the brainLearning from examples: goal is not to memorize but to generalize, eg predict.INPUTOUTPUTfGiven Given a set of a set of llexamples (data)examples (data)QuestionQuestion: find function : find function ffsuch that such that is a is a good predictorgood predictorof of yyfor a for a