Presentation 2014-11-21
Hyper-parameter estimation for compressive sensing with a Bernoulli-Gauss prior distribution
Toshiyuki WATANABE, Jun-ichi INOUE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Compressive sensing is a theory that estimates sparse information signals which has few non-zero elements from less observations. In terms of Bayesian estimation, Laplasian distribution (L_1-regularization) is normally chosen for the prior distribution. However, if a true distribution is a Bernoulli normal distribution whose zero or non-zero element is generated by a Bernoulli distribution of a non-zero rate (sparse rate as a 'hyper-parameter') and the non-zero elements is normally distributed, the Bernoulli normal distribution should be chosen for a candidate of the prior distribution in the sense of Bayesian optimality. In this paper, we evaluate a dependence of the hyper-parameter on the mean-square error by replica method and discuss the dynamics of hyper-parameter estimation by means of EM algorithm to maximize the marginal likelihood indirectly.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Compressive sensing / Statistical mechanics / Replica method / EM algorithm / Markov chain Monte Carlo method
Paper # NC2014-28
Date of Issue

Conference Information
Committee NC
Conference Date 2014/11/14(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) Hyper-parameter estimation for compressive sensing with a Bernoulli-Gauss prior distribution
Sub Title (in English)
Keyword(1) Compressive sensing
Keyword(2) Statistical mechanics
Keyword(3) Replica method
Keyword(4) EM algorithm
Keyword(5) Markov chain Monte Carlo method
1st Author's Name Toshiyuki WATANABE
1st Author's Affiliation Graduate School of Information Science and Technology, Hokkaido University /()
2nd Author's Name Jun-ichi INOUE
2nd Author's Affiliation
Date 2014-11-21
Paper # NC2014-28
Volume (vol) vol.114
Number (no) 326
Page pp.pp.-
#Pages 6
Date of Issue