Presentation 2004/6/18
Nonlinear noisy independent component analysis
Shin-ichi MAEDA, Shin ISHII,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In this study, we discuss a noisy nonlinear extension of independent component analysis (ICA). There have been proposed several extensions of the original noise-free linear ICA, e. g., noisy ICA or nonlinear ICA. There are few studies dealing with both noisy and nonlinear situations, however, because of the difficulty in integral calculation of the likelihood. In this study, we approximate the integral by a Taylor expansion and a Laplace approximation. The derived algorithm formulated as an expectation-maximization (EM) algorithm generalizes several of existing ICA algorithms. We also derive an optimal step size for our EM algorithm and discuss the reason why various noisy linear ICA algorithms based on maximum likelihood estimation are unsuccessful in being the noise-free linear ICA in the noiseless limit.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) independent component analysis / nonlinear mixture / Laplace approximation / EM algorithm
Paper # NC2004-35
Date of Issue

Conference Information
Committee NC
Conference Date 2004/6/18(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) Nonlinear noisy independent component analysis
Sub Title (in English)
Keyword(1) independent component analysis
Keyword(2) nonlinear mixture
Keyword(3) Laplace approximation
Keyword(4) EM algorithm
1st Author's Name Shin-ichi MAEDA
1st Author's Affiliation Nara Institute of Science and Technology()
2nd Author's Name Shin ISHII
2nd Author's Affiliation Nara Institute of Science and Technology
Date 2004/6/18
Paper # NC2004-35
Volume (vol) vol.104
Number (no) 140
Page pp.pp.-
#Pages 6
Date of Issue