Autoassociators and independent components

Author: Erkki Oja (Helsinki University of Technology)

Abstract: It is known that a linear autoencoder network (Baldi and Hornik) can compute PCA. It is shown that using a nonlinearity in the hidden layer, but keeping the weight matrices in the hidden and output layers suitable constrained, will solve the problem of independent component analysis and blind source separation, provided that the input vector is a linear mixture of independent sources. The learning rule is a nonlinear extension of the basic PCA subspace rule. Several examples are shown.