Statistical method of pruning neural networks

Abstract

A statistical method of pruning multilayer perceptrons is proposed, which is expected to involve less computation than does `optimal brain surgeon.' The statistical method iteratively performs the steps of estimating the error co-variance of the weights, evaluating the z-statistics for the weights, pruning the weights selected by hypothesis testing, and retraining the neural network. An interesting relationship between the statistical method and `optimal brain surgeon' is discussed. The relationship provides some insight into these two methods.