A neural network model based on short‐term memories for the hierarchical recognition of temporal patterns

Abstract

AbstractA new self‐organizing neural network model is proposed and evaluated for the mechanism of temporal pattern recognition of the auditory system. The model is constructed based on the hypothesis that total recognition of temporal patterns approximately of the length of words, is carried out by hierarchical identification and integration of the temporal relations of the constituent features. The model has a hierarchical structure in which short‐term memories storing spatial patterns, the circuits extracting temporally transient components of the pattern, and the feature detection circuits to identify spatial patterns are iteratively cascaded. After the circuit is self‐organized by repetitive presentation of training patterns, the model can correctly identify the training patterns and their temporally compressed and stretched patterns. It is also indicated that the short‐term memory function at each layer of the model is essential to the acceptance of temporally deformed patterns. Studies to expand the model to top‐down processing are also discussed.