Two-stage Neural Network for Blind Sources Separation

Seungjin Choi Ruey-Wen Liu Laboratory for Image and Signal Analysis Department of Electrical Engineering University of Notre Dame Notre Dame, IN 46556 Phone: (219) 631{6999 Fax: (219) 631{4393 E-Mail: [email protected] ABSTRACT In this paper, an on-line implementation of the simultaneous diagonalization (SD) of two di erent symmetric matrices is addressed. A two-stage neural network which consists of self-normalizing decorrelation and extended Oja's rule, is presented for an on-line implementation of SD. The SD of the 2ndand 4th-order moment matrices is known as one solution to the blind sources separation problem. It will be shown that the two-stage network presented can recover the source signals from a linear mixture without the knowledge of the mixing matrix and the distribution of the source signals.