Learning recognition of temporal sequences by coding temporal distance in neural networks

Abstract

Presents a neural network approach for recognition of temporal sequences. A dynamic-weight neural network (DNN) capable of explicitly extracting the temporal order of input sequences is introduced. The architecture of DNN employs a fully connected structure in that each neuron is linked to other neurons by a pair of long-term excitatory and short-term inhibitory weights. A two-pass training rule is developed to encode the temporal distance between two arbitrary occurrences. The overwriting problem is solved in the DNN by using the minimum requirement of hardware resources. We formally prove that a trained DNN ensures correct recognition of input training sequences and rejection of incorrect inputs.