Independent Component Analysis on the Basis of Helmholtz Machine
Masashi Ohata, Toshiharu Mukai, Kiyotoshi Matsuoka · 2003
This paper addresses an algorithm for independent component analysis on the basis of Helmholtz machine, which is an unsupervised learning machine. The algorithm is constructed by two terms. One is a term for obtaining a desired demixing process, which is the conventional rule on the base of the information theory. The other is a term for evaluating whether the inverse system and the set of obtained components are desirable or not. Due to this term, the algorithm effectively works in blind separation of overdetermined mixture with additive noise. We demonstrate the effectiveness by computer simulation.