Multi-Criteria Optimized Learning Rule for Optical Implementation

Abstract

We address the problem of learning rules for optical processor considered as a feature extractor for electronic or digital Neural Networks (N.N.). This approach consists in limiting the optical processing to the first layer of a N.N and yields, for the mean term, more realistic and promising solutions than ‘all-optical’ architectures. We propose a supervised non-iterative learning rule (i.e. with explicit formula for the weight interconnections) well suited for optical implementation. This method includes not only discrimination but also signal processing abilities. These new abilities are obtained with multi-criteria optimization which overcome the problem of over-specialization.