Summary: A Two­Round Variant of EM for Gaussian Mixtures
Sanjoy Dasgupta 
AT&T Labs -- Research
Leonard J. Schulman
Georgia Institute of Technology
Abstract
We show that, given data from a mixture of k
well­separated spherical Gaussians in R n , a sim­
ple two­round variant of EM will, with high
probability, learn the centers of the Gaussians to
near­optimal precision, if the dimension is high
(n  log k). We relate this to previous theoreti­
cal and empirical work on the EM algorithm.
1 Introduction
At present EM is the method of choice for learning mix­
tures of Gaussians. A series of theoretical and experimen­
tal studies over the past three decades have contributed to
the collective intuition about this algorithm. We will rein­
terpret a few of these results in the context of a new perfor­
mance guarantee.

Source: Abu-Mostafa, Yaser S. - Department of Mechanical Engineering & Computer Science Department, California Institute of Technology

Collections: Computer Technologies and Information Sciences