Weight discretization due to optical constraints and its influence on the generalization abilities of a simple perceptron

Abstract

Motivated from the optical implementation of NN which can be realized by storing the weights in holograms with a limited number of gray values, we focus our investigation on the dependence of the generalization and training errors of a simple perceptron with discrete weights, on the number of allowed discrete values (there are 2P allowed values for a bit precision of p) and on the training set size. Our starting point is the teacher pupil paradigm. The teacher is defined by fixing its continuous weights to random values. The pupil network that is only allowed to have discrete values was trained to learn the rule produced by the teacher with simulated annealing. For α αs, where a encodes the training set size, weight configurations exist so that, the training set can be reproduced without error whereas the generalization error is nonzero. For a > as there is no weight configuration of the pupil which can reproduce the training set without error and for a→ ∞ both training and generalization errors asymptotically converge to anЄmin. We found that between a precision of 5 bit and 8 bit there was no remarkable improvement in the generalization ablitity of the pupil perceptron. This result is very useful for the optical implementation since optical constraints for storing weights in holograms restrict precision to a maximum value of 6 bit.