A FastICA algorithm for non-negative independent component analysis

Abstract

The non-negative ICA problem is here defined by the constraint that the sources are non-negative with probability one. This case occurs in many practical applications like spectral or image analysis. It has then been shown by [10] that there is a straightforward way to find the sources: if one whitens the non-zero-mean observations and makes a rotation to positive factors, then these must be the original sources. A fast algorithm, resembling the FastICA method, is suggested here, rigorously analyzed, and experimented with in a simple image separation example. © Springer-Verlag 2004.