Generalization of back-propagation to recurrent neural networks

An adaptive neural network with asymmetric connections is proposed that is related to the Hopfield (1984) network with graded neurons. The present back-propagation algorithm uses a recurrent generalization of the delta rule of Rumelhart et al. (1986) to adaptively modify the synaptic weights. The network is architecturally simpler than the master/slave network of Lapedes and Farber (1986), and it vectorizes naturally because the units are homogeneous. (R.R.)