Fast Fixed-point Algorithms for Bayesian Blind Source Separation

In this paper, various fast algorithms for linear blind source separation (BSS) are developed. The fast xed-point algorithm for independent component analysis 8, 6] is interpreted as an EM-algorithm with a modiication for accelerating the convergence rate which can otherwise be slow particularly in a low noise environment. The new point of view opens way for developing several new fast xed-point algorithms for extracting signals with various properties. A Bayesian version of the ordinary independent component analysis and a version which takes into account both the time-domain behavior and non-Gaussianity of the source signals are studied in more depth. Both extensions have practical importance. The Bayesian version can be used for optimizing model structure and comparing diierent hypotheses. Natural signals have typically non-Gaussian distributions and time-dependencies. There has, therefore, been a demand for an algorithm which can utilize both types of information.