Independent Component Analysis for Mixed Sub-Gaussian and Super-Gaussian Sources

An extension of the infomax algorithm of Bell & Sejnowski (1995) is presented that is able to separate the mixed suband super-Gaussian source distributions. The same learning rule has been derived by Giiolami & Fyfe (1997) from the negentropy perspective for projection pursuit. Using a Laplacian prior we also propose a learning rule that is especially convenient to realize in hardware. The natural gradient extension is presented fiom different perspectives and the use of preprocessing steps is proposed to further speed up the convergence. Simulation results show that the algorithm is able to separate 20 source with a variety of source distributions. On real data, Jung et al. (1997) and McKeown et al. (1997) demonstrate the successful use of the extended ICA algorithm to analyze EEG and fMRI recordings.