580.691/491 Learning Theory
Course Instructor: Reza Shadmehr


Overview: This course introduces the probabilistic foundations of learning theory. We will discuss topics in regression, estimation, optimal control, system identification, Bayesian learning, and classification. Our aim is to first derive some of the important mathematical results in these topics, and then apply the framework to problems in biology, particularly animal learning and control of action.

 

Lecture times: Spring semester, 2008. Mondays and Wednesdays, 4:30 – 5:45 PM, Ames 219.

Teaching Assistant: Vincent Ethier

 

Exams: Midterm on March 12, Final on April 30.

 

Useful mathematical identities

 

Course Outline:

            Introduction

          Lecture 1: (intro.ppt) Introduction: adaptation vs. learning; linear classifiers; types of adaptation: supervised, unsupervised, reinforcement.
Homework: digit classification and cross validation

          Lecture 2: Review of probability theory.  Bayes rule, expected value and variance of random variables and sum of random variables, expected value of random variables raised to a power, Binomial distribution, Poisson distribution, Normal distribution.
Homework: probability theory

            Regression, generalization, and maximum likelihood

          Lecture 3: (LMS_1.ppt) Loss function as mean squared error; batch learning and the normal equation; Cross validation, batch vs. online learning, steepest descent algorithm, LMS, convergence of LMS.
Homework: (simulation) classify using regression.  Data set.

          Lecture 4: (LMS_2.ppt) Newton-Raphson, LMS and steepest descent with Newton-Raphson, weighted least-squares, regression with basis functions.
Homework: moving centers of Gaussian bases.

          Lecture 5: (generalization.ppt) generalization function, examples from psychophysics, estimation of generalization function from sequence of errors (linear technique)
Paper to discuss: Poggio T, Fahle M, Edelman S. (1992) Fast perceptual learning in visual hyperacuity.  Science 1992 May 15, 256:1018-21.
Homework: (simulation) estimate generalization function from record of errors.  Data set.

          Lecture 6: (ML_1.ppt) Maximum likelihood estimation; likelihood of data given a distribution; ML estimate of model weights and model noise.
Homework: derive online estimates of model weights and model noise.

          Lecture 7: (ML_2.ppt) Distribution of the ML estimate of model weights and model noise; multi-variate normal distribution; variance of model weights as a measure of model uncertainty.
Homework:  ML estimate of coin-toss probability, bias of the ML estimate of an exponential distribution.

 

            State estimation of linear stochastic systems

          Lecture 8:  (Kalmanfilter.ppt) Optimal parameter estimation, parameter uncertainty, state noise and measurement noise, adjusting learning rates to minimize model uncertainty.  Derivation of the Kalman filter algorithm.
Homework: Convergence of the Kalman gain and uncertainty.

          Lectures 9 and 10:  (same file as for lecture 9) Application of the optimal estimation algorithm to biological data: classical conditioning in animals, data fusion and combining data from multiple sensors, fast and slow memory systems, massed vs. spaced learning.  Forward models and integration of predicted sensory outcomes with measured sensory outcomes.
Homework: Simulation of a learner with fast, medium, and slow learning systems.
Homework: Simulation of a forward model with sensory feedback.

          Optimal feedback control of linear stochastic system

          Lecture 11:  (OptimalControl.ppt) Introduction to optimal control; method of Lagrange multipliers, open-loop optimal control, relating continuous and discrete linear systems.
Homework: non-uniform weighting of cost of motor commands and simulation of a saccade.

          Lecture 12:  (same file as for lecture 11) Optimal feedback control, optimal stochastic feedback control with Gaussian noise.
Homework: reaching to targets that might move.

          Lecture 13: (same file as for lecture 12) duality of optimal feedback control and the Kalman filter, signal dependent noise, optimal feedback control with signal dependent noise.

 

          Identification of linear stochastic systems

          Lecture 14:  (SubSpace.ppt) Introduction to subspace analysis; projection of row vectors of matrices, singular value decomposition, system identification of deterministic systems using subspace methods.
Homework: system identification of a deterministic system
Data set
Overschee and De Moor (1996) Subspace identification for linear systems: theory, implementation, applications.  Kluwer Academic, The Netherlands

 

          Bayesian integration

          Lecture 15:  (Bayes_1.ppt) “Single stage” maximum a posteriori (MAP) estimators with examples from selected distributions: coin toss with Beta distributed prior, classification with Gaussian distributed prior.
Homework: naïve Bayes classifier.

          Lecture 16:  (Bayes_2.ppt) Gaussian distribution and linear regression. Matrix inversion lemma; Bayesian integration of jointly distributed random variables; linear regression with a prior.
Homework: Posterior distribution with general variance covariance matrices; posterior distribution with two observed data points; maximizing the posterior directly.

          Lecture 17:  (Bayes_3.ppt) Bayesian learning in the central nervous system.  Derivation of LMS in the Bayesian case. 
Papers to discuss: KP Koerding, DM Wolpert (2004) Bayesian integration in sensorimotor learning.  Nature 427:244-247.  JM Hillis, MO Ernst, MS Banks, MS Landy (2002) Combining sensory information: mandatory fusion within, but not between, senses.  Science 298:1627-1630.
Homework: simulation of bimodal priors; optimal learning rates.

          Classification via Bayesian estimation

          Lecture 18:  Introduction to classification; Fisher linear discriminant, classification using posterior probabilities with explicit models of densities, confidence and error bounds of the Bayes classifier, Chernoff error bounds. 
Homework: Bayesian classification of a binary decision

          Lecture 19:  Linear and quadratic decision boundaries.  Equal-variance Gaussian densities (linear discriminant analysis), unequal-variance Gaussian densities (quadratic discriminant analysis), Kernel estimates of density.
Homework: Classification using assumptions of equal and unequal Gaussian distributions; classification using kernel density estimates.

          Lecture 20:  Logistic regression as a method to model posterior probability of class membership as a function of state variables; batch algorithm: Iterative Re-weighted Least Squares; on-line algorithm.
Homework: logistic regression with multiple classes of unequal variance.

          Lecture 21:  Neural mechanisms of classification learning; generalization in classification learning.  Basal ganglia damage disrupts classification learning but not cerebellar or medial-temporal lobe damage.  Cerebellar damage disrupts a form of regression learning but not basal ganglia damage.
Papers to discuss: Knowlton et al. (1996) A neo-striatal habit learning system in humans.  Science 273:1399.  Poldrack et al. (2001) Interactive memory systems in the human brain.  Nature 414:546. 

 

          Expectation Maximization

          Lecture 22:  Unsupervised classification.  Mixture models, K-means algorithm, and Expectation-Maximization (EM). 
Homework: image segmentation.  Imagedata 

          Lecture 23:  EM and conditional mixtures.  EM as maximizing the expected complete log-likelihood; method of Lagrange multipliers; selecting number of mixture components; mixture of experts.

 

          Reinforcement learning

          Lecture 24:  Introduction to reinforcement learning; value functions and Bellman equations; generalized policy iteration
Homework: rat maze problem. Mazedata .

          Lecture 25:  Temporal difference learning; policy improvement theorem; addiction and reinforcement learning.
Homework.  Randomwalkdata  Schultzpaper

          Lecture 26:  TD-lambda and eligibility trace.