Proceedings of the Annual Meeting of the Cognitive Science Society

This article contributes to the theory of network learning procedures by identifyingand analyzing two problems with the backpropagation procedure of Rumelhart, Hinton,and Williams (1985) that may slow its learning. Both problems are due to backpropagation's being a gradient- or steepest-descent method in the weight space of the network. Thefirst problem is that steepest descent is a particularly poor descent procedure for surfacescontaining ravines—places which curve more sharply in some directions than others—andsuch ravines are common and pronounced in performance surfaces arising from networks.The second problem is that steepest descent results in a high level of interference betweenlearning with different patterns, because those units that have so far been found most useful are also those most likely to be changed to handle new patterns. The same problemsprobably also arise with the Boltzmann machine learning procedure (Ackley, Hinton andSejnowski, 1985) and with reinforcement learning procedures (Barto and Anderson, 1985),as these are also steepest-descent procedures. Finally, some directions in which to lookfor improvements to backpropagation based on alternative descent procedures are brieflyconsidered.