Nonlinear independent component analysis by learning generalized adalines

This work proposes the novel modeling of post-nonlinear mixtures of independent sources and a learning method for the reverse problem that addresses on concurrent estimation of model parameters and independent components subject to given multichannel observations. The proposed post-nonlinear mixture model is realized by a network of multiple generalized adalines(gadalines), where each weighted gadaline emulates a transmitting link that maps independent sources to single channel observations. Based on the post-nonlinear mixture assumption, learning multiple weighted gadalines for retrieving independent components is resolved by the leave-one-out approximation operated under the mean-field-annealing process. Each time for some selected single channel observations, the dominant independent component is refined to compensate for the error of approximating the selected channel by the remaining independent components. This work shows that interactive dynamics derived for gadaline optimization executed under the mean field annealing is accurate and reliable for blind separation of post-nonlinear mixtures of independent sources.