Factors influencing learning by backpropagation

Abstract

The authors report on an investigation of learning by the backpropagation algorithm in a neural network. Computer simulations are used to predict the performance of a network in which noise is introduced into the interconnection weights, the network contains weights that are either analog or discrete, the maximum weight value is clamped, and the range of initial weight values is varied. The effect of these conditions is explored for the XOR problem. These simulations are a partial investigation of general factors which impact implementation designs. It is found that best results are achieved in a system with clamped analog weights and noise. However, surprisingly good performance is also obtained in a network with discretized weights as long as noise is present.<>