A functional-neural network for post-nonlinear independent component analysis

Abstract

In this paper a hybrid approach, based on a functional net- work and a neural netw ork, for post-nonlinear independeit component analysis is presented. In order to obtain the independence among the outputs, it was used as cost function a measure based on Renvi's quadratic entrop y and Caulay-Sc h w artz inequalyit Also, the Kernel method was iLsed for nonparametric estimation of the probability dcnsit v function. The experimental results corroborated the soundness of the approach and a comparative study with a neural nctw ork sho ed its superior performance. © Springer-Verlag Berlin Heidelberg 2001.