PURPOSE: To enable pattern recognition with small misrecognition by performing the feedback processing for a discrete period of time at every sampling time of an input pattern as continuous time processing instead of sampling, and thus easily learning between a recognition dictionary as coupling weights and the threshold value of an intermediate layer prescribing permissible variation ranges of respective recognition categories.

CONSTITUTION: A three-layered feedback associative storage model consists of differential addition state update units 12 and 15, sum of product computing elements 13, 14, and 17, a nonlinear processor 16, etc. Then while the dictionary is put a little closer to the input pattern indicated with the difference between the stored dictionary and feedback input pattern, an element to be fired is made easy to fire by lowering the threshold value, prescribing the permissible variation range of each recognition category, and an element not to be fired is made hard to fire by increasing the threshold value. Here, the state is updated not at intervals of sampling time, but continuously, so not only the dictionary, but also the threshold value can be learnt.