Weight discretization paradigm for optical neural networks

TLDR: In this paper a weight discretization paradigm is presented for back(ward error) propagation neural networks which can work with a very limited number of discretized levels.

Abstract: Neural networks are a primary candidate architecture for optical computing. One of the major problems in using neural networks for optical computers is that the information holders: the interconnection strengths (or weights) are normally real valued (continuous), whereas optics (light) is only capable of representing a few distinguishable intensity levels (discrete). In this paper a weight discretization paradigm is presented for back(ward error) propagation neural networks which can work with a very limited number of discretization levels. The number of interconnections in a (fully connected) neural network grows quadratically with the number of neurons of the network. Optics can handle a large number of interconnections because of the fact that light beams do not interfere with each other. A vast amount of light beams can therefore be used per unit of area. However the number of different values one can represent in a light beam is very limited. A flexible, portable (machine independent) neural network software package which is capable of weight discretization, is presented. The development of the software and some experiments have been done on personal computers. The major part of the testing, which requires a lot of computation, has been done using a CRAY X-MP/24 super computer.