The effect of weight quantization on the performance of linear decision functions

Adaptively derived linear decision boundaries are used in the recognition of typewritten numeric characters from a variety of fonts. During the adaption process, the weights are allowed to have as many quantization levels as required by the adaptive algorithm. Once these weights and their recognition rate on a test sample are determined, the number of quantization levels is successively reduced, and the effect on recognition rate is tabulated. System performance is maintained surprisingly well as the number of weight levels is reduced.