A Canonical Information Theoretic Algorithm for Blind Source Separation

C.-Y. Lee, D. Nam, and C.H. Park (Korea)


Blind Source Separation, Information Theory, MaximumEntropy, Minimum Mutual Information


Blind Source Separation (BSS) is based on statistical in dependence between the sources. Many algorithms have been developed among which the Minimum Mutual In formation (MMI) and the Maximum Entropy (ME) algo rithms are mostly used. These conventional algorithms have shown good results in some cases, but they have some drawbacks in several cases. In this paper, we will present some problems of conventional information theoretic BSS algorithms. To solve the problems, we propose a new infor mation theoretic BSS algorithm which can be used with any probability density estimation algorithm which produces differtiable probability density functions. Computer sim ulation with Hermite expansion and Gaussian mixture ap proximation shows that the proposed algorithm yields bet ter performance than the conventional ones.

Important Links:

Go Back