Feature extraction in the Neocognitron

Abstract

The authors present theoretical and numerical developments in the understanding of feature extraction in the Neocognitron. First, they show that the feature extraction process is equivalent to a generalized nonlinear discriminant. Second, they show that the operation of the feature-extraction process can be linked to the eigenvectors and eigenvalues of a matrix comprised of the excitatory and inhibitory convolution masks. Third, the authors show how the choice of parameters for the feature extraction and learning process affects the feature extraction capabilities of the machine. >