A mutual information minimization approach for a class of nonlinear recurrent separating systems

Leonardo Tomazeli Duarte (1) , Christian Jutten (1)

1 GIPSA-SIGMAPHY - GIPSA - Signal Images Physique

In this work, we deal with nonlinear blind source separation. Our contribution is the derivation of a learning strategy that minimizes the mutual information between the outputs of a class of nonlinear recurrent separating systems. By using the concept of the differential of the mutual information, we obtain an algorithm that does not need a precise knowledge of the source distributions, in contrast to the one obtained by a direct derivation of the minimum mutual information framework, or equally the maximum likelihood approach, for the considered model. The validity of our approach is supported by simulations.

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Leonardo Tomazeli Duarte, Christian Jutten. A mutual information minimization approach for a class of nonlinear recurrent separating systems. MLSP 2007 - IEEE 17th International Workshop on Machine Learning for Signal Processing, Aug 2007, Thessaloniki, Greece. ⟨hal-00169526⟩