Seminar on Computational Learning and Adaptation


  Lloyd Clustering of Gauss Mixtures

Robert M. Gray
Information Systems Lab
Department of Electrical Engineering
Stanford University

An early application of statistical clustering was Stuart Lloyd's 1959 algorithm for designing optimal quantizers, an algorithm commonly known in communications and signal processing applications as the "Lloyd-Max" algorithm. Quantization, or source coding with a fidelity criterion as it is known in Shannon information theory, strongly resembles a variety of problems that have arisen through the years in communications, signal processing, statistics, and mathematics. Included are several statistical clustering approaches such as k-means, the problem of sums of moments, and the problem of approximation of continuous probability distributions by discrete ones. The goal of this talk will be to describe the general quantization problem as it is typically formulated in information theory and to survey the state of the theory and design algorithms. Brief mention of similar problems in other fields will be made, but the specific examples used to illustrate the ideas will be the "worst case" role played by Gauss and Gauss mixture models and a resulting approach to designing Gauss mixture models from learning data via Lloyd clustering with a relative entropy distortion measure.



Date: Thursday, October 17

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule