DYNAMIC NEURAL NETWORK FOR TIME-SERIES PATTERN RECOGNITION

摘要

PURPOSE: To inexpensively attain a time series pattern recognition system for a voice or the like which can be highly accurately and quickly driven by the small number of learning patterns by time serially constituting a lower layer and dynamically determining optimum coupling relation on a part for transmitting the output of the lower layer to an upper layer. CONSTITUTION: An output group yik } from an intermediate layer is stored in a buffer 35. A coefficient group an jk } is stored in a load storing part 45. A microprocessor part 50 calculates outputs O(n) for respective output units (n) by using a table g(i, j) stored in a work memory 55. In the microcomputer constituted of a switching layer and an output layer, the lower layer can process a time series pattern to be an input pattern, and since the part for transmitting the output of the lower layer to the upper layer dynamically determines optimum coupling relation, the extension distortion of a time base existing in the lower layer can be removed. Consequently the time series pattern recognition system to be highly accurately and quickly driven by the small number of l-earning patterns can be inexpensively attained.