BLIND SOURCE SEPARATION AND DECONVOLUTION - THE DYNAMIC COMPONENT ANALYSIS ALGORITHM

We derive a novel family of unsupervised learning algorithms for blind separation of mixed and convolved sources. Our approach is based on f ormulating the separation problem as a learning task of a spatiotempor al generative model, whose parameters are adapted iteratively to minim ize suitable error functions, thus ensuring stability of the algorithm s. The resulting learning rules achieve separation by exploiting high- order spatiotemporal statistics of the mixture data. Different rules a re obtained by learning generative models in the frequency and time do mains, whereas a hybrid frequency-time model leads to the best perform ance. These algorithms generalize independent component analysis to th e case of convolutive mixtures and exhibit superior performance on ins tantaneous mixtures. An extension of the relative-gradient concept to the spatiotemporal case leads to fast and efficient learning rules wit h equivariant properties. Our approach can incorporate information abo ut the mixing situation when available, resulting in a ''semiblind'' s eparation method. The spatiotemporal redundancy reduction performed by our algorithms is shown to be equivalent to information-rate maximiza tion through a simple network. We illustrate the performance of these algorithms by successfully separating instantaneous and convolutive mi xtures of speech and noise signals.