Lectures

Lecture Notes for CS429/529, Spring 2007

This page will house lecture notes, slides, code, URLs, etc., as they become available.

Lecture 02: Jan 18: Introduction: the basic (supervised) learning problem; taxonomy of ML; definitions.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 03: Jan 23: HW1 assigned; hypothesis spaces; loss functions; the supervised learning problem; decision trees; the DT learning algorithm part I.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 04: Jan 25: DTs continued: splitting functions; learning bias; entropy; information gain; practicalities: pruning and multi-way splits.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 05: Jan 30: Mathematical aside: optimization of multivariate functions; measuring performance: empirical vs. generalization; the first amendment of machine learning.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 06: Feb 01: Measuring performance: holdout data, stratification, randomization, cross-validation, learning curves, variance; vector spaces, norms, and inner products.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 07: Feb 06: Homework 2 assigned; intro to linear regression; homogeneous coordinates; the squared-error loss function; vector/matrix derivative identities; minimization of the loss function; relation to Hilbert spaces.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 08: Feb 08: Reading 1 assigned; a LSE example; linear discriminants via LSE; multi-class classification from linear discriminants; intro to support vector machines.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 09: Feb 13: HW1 returned; nonlinearly separable data; projection functions; nonlinear surfaces; geometry of hyperplanes; the maximum-margin objective function; quadratic programming formulation; nonseparable data and slack variables; the dual form; kernel functions.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 10: Feb 15: Support vector machines finalized: kernel functions; example kernels; interpretation of kernels; law of cosines; angles in projected space; support vector classifier form; putting it together: all the steps in one place; final notes: SVMs and dimensionality.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 11: Feb 20: HW3 assigned; ML Trivia; meditations on assumptions; gender classification; statistical decision theory (intro); the Bayes optimal decision rule; example.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 12: Feb 22: Review: Bayes-optimal decision boundary; model assumptions and parameter priors; the 1-d Gaussian distribution; the D-d Gaussian distribution; exercise: Gaussian decision surfaces.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 13: Feb 27: Final project proposal assigned; midterm announced; statistical model learning; the parameter learning problem; principle of maximum likelihood; the likelihood function; joint vs. marginal PDFs; statistical independence; IID samples; the joint likelihood function; exercise: maximum likelihood estimation.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 14: Mar 06: MLE exercise completed; the likelihood function in action; the full statistical modeling pipeline; interpretations of Gaussian distributions; conditional distributions and Bayes' rule; intro to Bayesian posterior analysis.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 15: Mar 08: The Bayesian modeling worldview; philosophy; example: Bayesian coin flipping; Bayes' rule in action; generative, prior, data, and posterior distributions; the Beta distribution; the binomial distribution; conjugate priors and posteriors; exercise: airspeed of an unladen (African) swallow.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 16: Mar 20: Bayesian swallows, finalized; examination of the Gaussian posterior; metric functions; example metrics; the nearest-neighbor classifier; properties; k-NN; nearest-neighbors as density estimators; curse of dimensionality; metric spaces, Banach spaces, and Hilbert spaces.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 17: Mar 27: Exam back; Readin 2 assigned; dissertation announcements; Intro to reinforcement learning: actions, outcomes, rewards, states; reward aggregation; SARS tuples; histories; policies; value functions: finite and infinite horizon value.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 18: Mar 29: RL continued: stochastic outcomes; transition functions; probability of histories; history dependence of transition function and parameter size; Markov chain refresher; definition: Markov decision process (MDP); exercise: MDP formulation.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 19: Apr 03: Value in stochastic worlds: expected value; principle of maximum expected utility; the RL problem concretely; taxonomy: planning vs. learning; planning sub-problems: policy evaluation and policy maximization; policy evaluation: the Bellman equation; solution of the Bellman equation.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 20: Apr 05: R3 assigned; solutions of the Bellman equation; example: navigation in a gridworld; formalization; cute pictures: optimal policy, value function; maze variants; the policy iteration algorithm; visualization of PI in action; properties and variant policy solution algorithms.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 21: Apr 10: Q functions; policy iteration in terms of Q; Q learning; Q learning in action; measuring the performance of RL algorithms; drawbacks of Q learning.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 22: Apr 12: Notes on technical writing; Q learning revisited; what Q learning is doing; action selection and exploration; off-policy vs. on-policy learning; drawbacks to basic Q learning; eligibility traces; SARSA; failure modes of off-policy learning.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 23: Apr 17: Assignment of presentation dates; presentation tips; eligibility: "radioactive breadcrumbs"; the forward view of eligibility; replacing traces; discussion of R3.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 24: Apr 19: ICES surveys; tips on final project report; model free vs. model based RL; the E^3 algorithm; optimisim under uncertainty; exploration vs. exploitation; Chernoff bounds and convergence of T estimates; poly-time RL.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 25: Apr 24: Unsupervised learning: intro to topics; clustering; generative models and hidden/latent variables; coin flipping and dice rolling; the expectation-maximization meta-algorithm; Gaussian mixture models; (intuitive) derivation of the M and E steps; examples.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|
Lecture 26: Apr 26: (Brief) intro to dimensionality reduction; objective functions; feature subset selection and the wrapper method; linear projections; PCA; nonlinear dimensionality reduction (NLDR); LLE.
Lecture Notes
|Keynote slides| |PDFified slides| |PowerPoint slides|

Last updated: Thu 26 Apr 2007 03:35:56 PM MDT