A general architecture for classical conditioning of perceptual-motor sequences

Abstract

A general neural network architecture, loosely modeled on the cerebral cortex, for the classical conditioning of perceptual-motor sequences is described. The utility of such an algorithm in robotics applications lies in its potential to adaptively order gross and fine motor actions under sensory control. The architecture includes (1) an input field of the observed pattern, (2) short-term storage of the input history, (3) an output field of the predicted pattern, (4) a comparator of the observed and predicted patterns, (5) an error field to store the differences, and (6) association matrices of the input history with the current input, the current error, and the next motor command. Applications of the model to machine vision and an artificial auditory system are discussed