Recognition of temporal patterns using state transitions of neural networks (auditory application)

The author's propose a three-layer neural network model for the recognition of spatiotemporal patterns. The operation of this model is based on the voluntarily or externally leaded state transition of activity patterns on mutually connected neurons. The authors also demonstrate that the model has some interesting characteristics related to the ability to recognize time-warped sequences, sequences that have not been isolated (segmented) to words, sequences that include some dropping out of the constituent patterns, and sequences that include some noise patterns. In other words the model can be interpreted to have the primitive functions of segmentation, top-down processing, and selective attention.<>