Presentation | 2004/10/12 Expectation maximum algorithm and statistical physics Jun-ichi INOUE, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | When we attempt to make statistical inference by using Markov random fields or Bayesian networks, theprobabilistic model includes hyper-parameters and some missing data which are impossible to be observed. In suchcases, we should choose the strategy that maximizes so-called marginal likelihood function and regard the solution ofthis optimization problem as the best suitable values of the hyper-parameters. To maximize the marginal likelihoodfunction systematically, EM algorithm (Expectation Maximum algorithm) is widely used. In this talk, I present severalaspects of this algorithm from a view point of statistical mechanics. I also show the relation between the EM andstatistical physics with the following key concepts in mind : marginal likelihood function and free energy, MCMCmethod and meanfield approximation, replica method etc. for several problems of information processing. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | EM algorithm / Maximum likelihood estimate / Statistical mechanics / Gaussian mixture / MRFs |
Paper # | NC2004-76 |
Date of Issue |
Conference Information | |
Committee | NC |
---|---|
Conference Date | 2004/10/12(1days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | |
Vice Chair | |
Secretary | |
Assistant |
Paper Information | |
Registration To | Neurocomputing (NC) |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Expectation maximum algorithm and statistical physics |
Sub Title (in English) | |
Keyword(1) | EM algorithm |
Keyword(2) | Maximum likelihood estimate |
Keyword(3) | Statistical mechanics |
Keyword(4) | Gaussian mixture |
Keyword(5) | MRFs |
1st Author's Name | Jun-ichi INOUE |
1st Author's Affiliation | Graduate School of Information Science and Technology, Hokkaido University() |
Date | 2004/10/12 |
Paper # | NC2004-76 |
Volume (vol) | vol.104 |
Number (no) | 349 |
Page | pp.pp.- |
#Pages | 6 |
Date of Issue |