Realistic model on the blind separation of convolved information sources

The difficult problem of separating multiple sources from mixed recordings in "real" enviroment situations (common in many medical signal recording applications) is addressed in this paper. Although other researchers have also used infomax to learn the demixing filters, one limiting assumption exists in all of those solutions: All sources except one reach each recording site delayed! Our feedback model allows for zero-delayed versions of all sources to be present in each one of the recorded mixtures, a much better approximation of "real" situations, but at the same time a much more difficult one because of the recursiveness it introduces. The update rules are derived in matrix form with special care to keep the diagonals zero, so to avoid "temporal whitening" at the output. Examples of successful separation of artificially mixed information signals are presented.