A complexity estimation approach for estimating neural net size

A complexity estimation technique is developed for predicting the number of hidden units a multilayer perceptron (MLP) requires to reach a given performance. First, a distance measure is developed which rejects useless inputs. An approximately optimal clustering algorithm for designing nearest neighbor estimators (NNEs) is presented which uses the distance measure. New formulas are given for the numbers of hidden units required to store a given number of patterns. Examples are given which illustrate the usefulness of this technique.