Separation of statistical dependent sources using a measure of non-Gaussianity

A new Blind Source Separation (BSS) algorithm for the case of dependent or independent sources is proposed. This is called MaxNG algorithm and is based on the Maximization of Non-Gaussianity of source estimates, which is equivalent to minimize the Shannon-entropy. In order to measure non-Gaussianity, the Parzen window non-parametric density estimation technique and the L_2-Euclidean distance in the space of density functions are proposed. In this presentation, after a brief review of some existing BSS/ICA methods, the main characteristics of MaxNG are explained and some results, comparing MaxNG against a commonly used strategy based on the minimization of Mutual Information (MinMI) are shown. It is shown that, for uncorrelated sources both strategies reach similar solutions but when sources are correlated, much better results are obtained using MaxNG. Also results of the application to real-world data of some known algorithms like AMUSE, EVD2, SOBI, JADE, FPICA are presented and compared with MaxNG.