CMP242 Course Description
Spring 1998

This course offers an introduction to the design and analysis of machine learning algorithms. It covers learning models from the fields of statistical decision theory and pattern recognition, artificial intelligence, and theoretical computer science. Topics include graphical models (which will get special emphasis this quarter), maximum likelihood and Bayes estimation, Bayes decision theory, classification learning, nonlinear regression and neural networks, density estimation and other kinds of unsupervised learning, as well as formal models of learning.

Prerequisites for the course are knowledge of multivariate calculus, probability, and analysis of algorithms. Knowledge of analysis of algorithms should be at the level of CMPS201. Knowledge of probability should be at least at the senior undergraduate math course level. You should know about discrete probability spaces and counting, conditional probabilities, Bayes' theorem, independence, random variables and distributions (binomial, geometric, negative binomial and Poisson), expectation, variance, Chebychev's inequality, the weak law of large numbers and the central limit theorem. A possible book to review these areas would be S. Ross, "A First Course in Probability," McMillan, 1984, or his other book "An Introduction to Probability Models" (on reserve in the Science Library), but there are many others.

The requirements for the course are: project or paper (60%), homework and exams (40%). This is not a seminar course. Students will not be required to give talks. The project will be due the last day of classes. A one page midterm report on the project will be due at the end of the fifth week. Most projects for CIS242 come in one of three forms:

  1. Code up and test a learning algorithm, or test many learning algorithms that other people have coded up.
  2. Establish a new analytical performance bound for a learning algorithm
  3. Write a critical review of current research in one area of Machine Learning
Options 2 and 3 involve writing a paper that is at least of the quality of a short (7-15 pages) technical report or journal paper (see board office for examples). Papers for option 2 will be shorter than those for option 3. Option 1 involves mostly coding and testing, although programs must be accompanied by a brief (2-3 page) user guide describing what the software is and how it is used, and a brief (5-6 page) summary of the experimental results, presented in the form of a very short scientific paper. Most students choose 1. I have tried to compile a list of ideas for class projects as well.

There are two texts for this class. Both are on reserve in the Science library (or will be soon). The first is "An Introduction to Bayesian Networks" by Finn V. Jensen, published by Springer. We will use this for about one third the course. There are copies at the Baytree bookstore. The second is "Pattern recognition and Neural Networks" by B. D. Ripley. This should be at the bookstore by April 14. The remainder of the course material will be handouts and readings put on reserve.

Questions regarding about page content should be directed to haussler@cse.ucsc.edu.
Last modified April 2, 1998.

Back to the CMP 242 Class Page.