Gaussian Mixture Models (GMMs) are among the most statistically mature methods for clustering (though they are also used intensively for density estimation). In this tutorial, we introduce the concept of clustering, and see how one form of clustering…in which we assume that individual datapoints are generated by first choosing one of a set of multivariate Gaussians and then sampling from them…can be a well-defined computational operation. We then see how to learn such a thing from data, and we discover that an optimization approach not used in any of the previous Andrew Tutorials can help considerably here. This optimization method is called Expectation Maximization (EM). We'll spend some time giving a few high level explanations and demonstrations of EM, which turns out to be valuable for many other algorithms beyond Gaussian Mixture Models (we'll meet EM again in the later Andrew Tutorial on Hidden Markov Models). The wild'n'crazy algebra mentioned in the text can be found (hand-written) Download Tutorial Slides (PDF format).
Powerpoint Format: The Powerpoint originals of these slides are freely available to anyone who wishes to use them for their own work, or who wishes to teach using them in an academic institution. Please email Andrew Moore at firstname.lastname@example.org if you would like him to send them to you. The only restriction is that they are not freely available for use as teaching materials in classes or tutorials outside degree-granting academic institutions.
Advertisment: I have recently joined Google, and am starting up the new Google Pittsburgh office on CMU's campus. We are hiring creative computer scientists who love programming, and Machine Learning is one the focus areas of the office. If you might be interested, feel welcome to send me email: email@example.com .