Are There Local Maxima in the Infinite-Sample Likelihood of Gaussian Mixture Estimation?
Abstract
Consider the problem of estimating the centers of a uniform mixture of unit-variance spherical Gaussians in , 1 from i.i.d. samples x _1,..., x _ m drawn from this distribution. This can be done by maximizing the (average log) likelihood . Maximizing the likelihood is guaranteed to recover the correct centers, in the large-sample limit, for any mixture model of the form (1). Unfortunately, maximizing the likelihood is hard in the worst case, and we usually revert to local search heuristics such as Expectation Maximization (EM) which can get trapped in the many local minima the likelihood function might have.
Cite
Text
Srebro. "Are There Local Maxima in the Infinite-Sample Likelihood of Gaussian Mixture Estimation?." Annual Conference on Computational Learning Theory, 2007. doi:10.1007/978-3-540-72927-3_47Markdown
[Srebro. "Are There Local Maxima in the Infinite-Sample Likelihood of Gaussian Mixture Estimation?." Annual Conference on Computational Learning Theory, 2007.](https://mlanthology.org/colt/2007/srebro2007colt-there/) doi:10.1007/978-3-540-72927-3_47BibTeX
@inproceedings{srebro2007colt-there,
title = {{Are There Local Maxima in the Infinite-Sample Likelihood of Gaussian Mixture Estimation?}},
author = {Srebro, Nathan},
booktitle = {Annual Conference on Computational Learning Theory},
year = {2007},
pages = {628-629},
doi = {10.1007/978-3-540-72927-3_47},
url = {https://mlanthology.org/colt/2007/srebro2007colt-there/}
}