CREATING ARTIFICIAL NEURAL NETWORKS THAT GENERALIZE

We develop a technique to test the hypothesis that multilayered, feed-forward networks with few units on the first hidden layer generalize better than networks with many units in the first layer. Large networks are trained to perform a classification task and the redundant units are removed ("pruning") to produce the smallest network capable of performing the task. A technique for inserting layers where pruning has introduced linear inseparability is also described. Two tests of ability to generalize are used - the ability to classify training inputs corrupted by noise and the ability to classify new patterns from each class. The hypothesis is found to be false for networks trained with noisy inputs. Pruning to the minimum number of units in the first layer produces networks which correctly classify the training set but generalize poorly compared with larger networks.