Split and merge EM algorithm for improving Gaussian ixture density estimates

The EM algorithm for Gaussian mixture models often gets caught in local max ima of the likelihood which involve having too many Gaussians in one part o f the space and too few in another, widely separated part of the space. We present a new EM algorithm which performs split and merge operations on the Gaussians to escape from these configurations. This algorithm uses two nov el criteria for efficiently selecting the split and merge candidates. Exper imental results on synthetic and real data show the effectiveness of using the split and merge operations to improve the likelihood of both the traini ng data and of held-out test data.