Presentation 2011-03-08
Comparison between the Parameter and the Hidden Variable Space for Calculation of the Marginal Likelihood
Takushi MIKI, Keisuke YAMAZAKI, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) The marginal likelihood has important information for model selection and optimization of a prior distribution. In practical situations, a numerical calculation based on the Markov Chain Monte Carlo(MCMC) method is necessary because the analytical calculation is not straightforward in hierarchical learning modes. There are two MCMC methods; one is on the parameter space and the other is on the hidden variable space. However, it has not been studied yet how the space difference affects the numerical accuracy. In this paper, we calculate the marginal likelihood of a mixture of normal distributions using the methods, and compare the two spaces.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Marginal likelihood / Markov Chain Monte Carlo method / mixture of normal distributions / hidden variable
Paper # NC2010-176
Date of Issue

Conference Information
Committee NC
Conference Date 2011/2/28(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) Comparison between the Parameter and the Hidden Variable Space for Calculation of the Marginal Likelihood
Sub Title (in English)
Keyword(1) Marginal likelihood
Keyword(2) Markov Chain Monte Carlo method
Keyword(3) mixture of normal distributions
Keyword(4) hidden variable
1st Author's Name Takushi MIKI
1st Author's Affiliation Department of Computer Science, Tokyo Tech.()
2nd Author's Name Keisuke YAMAZAKI
2nd Author's Affiliation PI Lab, Tokyo tech.
3rd Author's Name Sumio WATANABE
3rd Author's Affiliation PI Lab, Tokyo tech.
Date 2011-03-08
Paper # NC2010-176
Volume (vol) vol.110
Number (no) 461
Page pp.pp.-
#Pages 6
Date of Issue