Title: State-Space Approaches to Blind Source Separation

Date: Thursday, 8 March 2001

Time: 11:00am - 12:00noon

Venue: Room 2404, HKUST

Abstract:
Blind source separation (BSS) is to recover a set of statistically independent source signals (shortly as sources) from observed mixtures of them while the mixing process is unknown. Due to its attractive applications in speech processing, wireless communications, time series analysis and so on, this problem with a linear instantaneous mixture has been formulated as independent component analysis (ICA), widely studied in the past decade. However, the typical existing ICA techniques require that: (1) each source is independently and identically distributed, and (2) observations are measured noiselessly. Otherwise, their performance will deteriorate. In this talk, some advanced works on BSS are presented with considering two cases: real-valued BSS and discrete-valued BSS. In the former, a general state-space approach called temporal factor analysis (TFA), developed from temporal Bayesian Ying-Yang (TBYY) learning system, is investigated. A connection between the TFA and a traditional filtering problem in control theory is set up, resulting in a linear-state-space TFA algorithm obtained for temporal source separation under noisy environment. Beside this, an alternative approach called dual multivariate auto-regressive (AR) modeling is also suggested when measurement noise is negligible. This method models both of sources and observations as a multivariate AR process, respectively. Consequently, the mixing process from temporal sources to observations is the same as that from the independent residuals of the source AR process to those of the observation AR (OAR) process. It is therefore that the source temporal effects in performing BSS can be avoided by learning the de-mixing process on the OAR residuals rather than the observations. A specific adaptive algorithm has been proposed to implement this approach. In the latter, we set up the connection between the state identification and the observation clustering features, whereby an approach based on Rival Penalized Competitive Learning rule is proposed to perform discrete-valued source recovery without any knowledge about the number of sources. The effectiveness of the algorithms mentioned above is all demonstrated by the experiments.

Biography:
Dr. Yiu-ming Cheung received his Ph.D. degree in Computer Science and Engineering from the Chinese University of Hong Kong in 2000. He is currently an assistant professor in the same institution. His research interests include blind signal separation, machine intelligence learning, pattern recognition, time series forecasting and automated trading system. He is a member of the program committee in the international conferences: IC-AI'2000, IC-AI'2001 and CISST'2001, and is a special session chair of IDEAL'2000. Also, he is a journal reviewer of Neurocomputing and IEEE Transactions on Neural Networks.