A robust objective function of joint approximate diagonalization

Authors:
Yoshitatsu Matsuda;Kazunori Yamaguchi
Affiliations:
Department of Integrated Information Technology, Aoyama Gakuin University, Sagamihara-shi, Kanagawa, Japan;Department of General Systems Studies, Graduate School of Arts and Sciences, The University of Tokyo, Meguro-ku, Tokyo, Japan
Venue:
ICANN'12 Proceedings of the 22nd international conference on Artificial Neural Networks and Machine Learning - Volume Part II
Year:
2012

Citing 5
Cited 0

High-order contrasts for independent component analysis

Neural Computation
Independent component analysis using an extended infomax algorithm for mixed subgaussian and supergaussian sources

Neural Computation
Adaptive Blind Signal and Image Processing: Learning Algorithms and Applications

Adaptive Blind Signal and Image Processing: Learning Algorithms and Applications
An adaptive threshold in joint approximate diagonalization by assuming exponentially distributed errors

Neurocomputing
An information theoretic approach to joint approximate diagonalization

ICONIP'11 Proceedings of the 18th international conference on Neural Information Processing - Volume Part I

Quantified Score

Hi-index	0.00

Visualization

Abstract

Joint approximate diagonalization (JAD) is a method solving blind source separation, which can extract non-Gaussian sources without any other prior knowledge. However, it is not robust when the sample size is small because JAD is based on an algebraic objective function. In this paper, a new robust objective function of JAD is derived by an information theoretic approach. It has been shown in previous works that the "true" probabilistic distribution of non-diagonal elements of approximately-diagonalized cumulant matrices in JAD is Gaussian with a fixed variance. Here, the distribution of the diagonal elements is also approximated as Gaussian where the variance is an adjustable parameter. Then, a new objective function is defined as the likelihood of the distribution. Numerical experiments verify that the new objective function is effective when the sample size is small.