Unsupervised Learning of Temporal Sequences by Neural Networks

Abstract

We propose to define a new model of formal neural network. This model extends existing Hopfield networks to process temporal data and achieve a non-supervised learning of them. We propose a learning law to adress in this context the sensitivity to input changes. A spatial representation of network’s temporal activity is given by which learnt sequences can be identified. An example of such a network is given and the results of the simulation are presented.