EXTRACTING ALGORITHMS FROM PATTERN-CLASSIFICATION NEURAL NETWORKS

It has been shown that a pattern classification neural network can be synthesized in three layers (one hidden layer) if each class is separated into a single convex region by hyperplanes and in four layers if not. Based upon this, extraction of a pattern classification algorithm from a network is discussed, namely, first separation hyperplanes are extracted from a network trained by the backpropagation algorithm, and then the weights corresponding to the separation hyperplanes and those connected to the output neurons are successively tuned to improve the generalization ability of the network. Classification is made according to which sides of the hyperplanes a test datum (or the second layer output of a test datum for four layers) is on. Finally, weights of the neural networks, generated by the backpropagation algorithm for a number recognition system, are tuned and classification algorithms extracted.