The problem of overlearning in high-order ICA approaches: Analysis and solutions

Abstract

We consider a type of overlearning typical of independent component analysis algorithms. These can be seen to minimize the mutual information between source estimates. The overlearning causes spike-like signals if there are too few samples or there is a considerable amount of noise present. It is argued that if the data has flicker noise the problem is more severe and is better characterized by bumps instead of spikes. The problem is demonstrated using recorded magnetoencephalographic signals. Several methods are suggested that attempt to solve the overlearning problem or, at least, diminish/reduce its effects. © Springer-Verlag Berlin Heidelberg 2001.