A biologically plausible implementation of error-backpropagation for classification tasks

Van Ooyen, A., and Roelfsema, P. R. (2003). In: Kaynak, O., Alpaydin, E., Oja, E., and Xu, L., eds. Supplementary Proceedings of the International Conference on Artificial Neural Networks - ICANN/ICONIP 2003, Istanbul, Turkey, June 2003, pp. 442-444. [Full text: PDF]

Abstract

Error-backpropagation is a powerful method to train neural networks, but its current implementations lack biological realism. Here we present a novel scheme for implementing error-backpropagation in a biologically plausible way. Our scheme, called attention-gated reinforcement learning (AGREL), uses an "attentional" feedback signal to gate the plasticity of connections to hidden units. We show that the average changes in connection weights in AGREL are the same as the changes in weights in error-backpropagation.