Learning temporal sequences using dual-weight neurons

Abstract

This paper considers the use of neural networks (NN 's) in learning temporal sequence recognition and reproduction for which the sequence degree is unknown. This approach uses the output ambiguity to train the network without the need to assume or construct a separate model fur the input sequence degree. First we introduce a primitive network called the DNN, comprising a plurality of dual-weight (DN) neurons. Each neuron is linked to other neurons by a long-term excitatory weight and a short-term inhibitory weight. Fast learning is made possible by employing a two-pass training rule to encode the temporal distance between two arbitrary pattern occurrences. The resulting DNN is then extended into a more generalized model, namely the DNN2. By incorporating the two-pass rule and a self-organizing algorithm, the DNN2 can achieve autonomous temporal sequence recognition acid reproduction. Using training efficiency and hardware complexity criteria, the DNN networks are also contrasted with the work of Wang and Yuwono (1995).