Experiments on learning in recursive neural networks

Abstract

Summary form only given. An examination is made of training a network with recursive connections. The author is able to train recursive networks reliably using the generalized delta rule of error backpropagation of Rumelhart, Hinton, and Williams on the stationary states of the recursive network. This method for training recursive networks is a truncated form of the recursive error backpropgation algorithm developed by Pineda and Almeida.