SSP'05 -- Proceedings

	IEEE/SP 13th workshop on Statistical Signal Processing July, 17-20, 2005 - Bordeaux - France

Information regarding the paper

Title

Bayesian noisy ICA for source switching environments

Author(s)

Jun-ichiro Hirayama	Graduate School of Information Science, Nara Institute of Science and Technology
Shin-ichi Maeda	Graduate School of Information Science, Nara Institute of Science and Technology
Shin Ishii	Graduate School of Information Science, Nara Institute of Science and Technology

Get the paper in PDF format

To obtain Acrobat Reader (version 5 minimum required) necessary to his read.

Abstract

Most of the existing algorithms for blind source separation (BSS) assume that the number of sources is known and constant for all samples. Real situations, however, often have difficult non-stationarity such that each source signal abruptly switches to appear or disappear and hence the number of sources varies with time. In this article, we propose a noisy independent component analysis (ICA) algorithm that assumes unknown and varying number of sources. We employ Bayesian variable selection in combination with the hidden Markov model to automatically select and switch the set of sources which are temporally active in a certain period. We formulate our algorithm based on the Bayesian inference using the variational Bayes method. A simulation study using artificial data showed that our approach successfully recovered source signals even when the number of sources varied with time.