An echo state network with working memories for probabilistic language modeling

Abstract

In this paper, we propose an ESN having multiple timescale layer and working memories as a probabilistic language model. The reservoir of the proposed model is composed of three neuron groups each with an associated time constant, which enables the model to learn the hierarchical structure of language. We add working memories to enhance the effect of multiple timescale layers. As shown by the experiments, the proposed model can be trained efficiently and accurately to predict the next word from given words. In addition, we found that use of working memories is especially effective in learning grammatical structure. © 2013 Springer-Verlag Berlin Heidelberg.