Connectionist networks for binary bit-string multiplication

The connectionist learning technique of backpropagation is utilized to train a three-layer network to simulate the operation of a binary bit-string multiplier. It is shown that the network develops internal representations of a high quality, that allow it to generalize correctly to novel input patterns. The learning algorithm of N. Littlestohe (1987) is also used to train a network with a problem-specific architecture to realize a multiplier. The advantages and disadvantages of such an approach are discussed and a brief analysis of the results is presented