Frequency multiplexed raster neural networks 1: Theory

Abstract

In the first part of this paper we present frequency multiplexed raster (FMR) optical implementation of neural networks. A hidden difficulty for hardware (optical and electronic) implementation is that the dimension of the synaptic matrix is twice that of the input and output matrices or vectors. For 2-D images, which is we believe one of the greatest potentialities of neural networks, the synaptic matrix is 4-D and cannot be directly implemented in optics. We propose FMR as a method to fold this matrix into a 2-D format. In the second part of this paper we describe the system built in our laboratory showing the feasibility of FMR optical neural networks. The system is built from an optical input module, a fixed synaptic matrix coded on a transparency, a CCD camera, and a microcomputer which performs the thresholding and feedback operations. In a later stage the fixed matrix will be replaced by a programmable matrix.