Abstract

Pseudoinverse is a linear mapping algorithm that produces minimum-squared errors. Pseudoinverse neural networks (PNNs) implement this algorithm by using neuralnetwork components in a two-layer structure. The first layer with N neurons is trained with a Kittler-Young transform and a pseudoinverse transform by using a set of training images from all image classes. As a result, the outputs of the first layer are the matching scores between input images and the basis images of different classes. A MAXNET of M neurons is built in the second layer of the PNN to pickup the maximum matching score. In the MAXNET, the interconnections between neurons are inhibitory with a value of -a, where a <1/M and the connections of the neurons to themselves are 1. High-space-bandwidth-product operation is possible because the number of interconnections is only M × N in the PNN. In our experiment, a optical PNN with 16 384 neurons is built, and pattern recognitions of 128 × 128 images are performed. Furthermore, the PNN does not suffer from spurious outputs, which can produce no-match results.

© 1990 Optical Society of America