Discriminant component pruning: Regularization and interpretation of multilayered backpropagation networks

Neural networks are often employed as tools in classification tasks. The us e of large networks increases the likelihood of the task's being learned, a lthough it may also lead to increased complexity. Pruning is an effective w ay of reducing the complexity of large networks. We present discriminant co mponents pruning (DCP), a method of pruning matrices of summed contribution s between layers of a neural network. Attempting to interpret the underlyin g functions learned by the network can be aided by pruning the network. Gen eralization performance should be maintained at its optimal level following pruning. We demonstrate DCP's effectiveness at maintaining generalization performance, applicability to a wider range of problems, and the usefulness of such pruning for network interpretation. Possible enhancements are disc ussed for the identification of the optimal reduced rank and inclusion of n onlinear neural activation functions in the pruning algorithm.