Simulation of a Classically Conditioned Response: Components of the Input Trace and a Cerebellar Neural Network Implementation of the Sutton-Barto-Desmond Model

摘要

Workers in the behavioral and neurosciences have developed a fruitful approach to modeling brain function by combining mathematics with neurophysiology and anatomy. This approach, sometimes referred to as computational neuroscience, provides a framework for integrating seemingly divergent areas of scientific inquiry. A particular example is the extension of a general mathematical model of learning to a specific instance of behavioral learning such as classical conditioning. Although most models represent the cumulative effects of conditioning without reference to motor output, the authors have shown how a template of the classically conditioned nictitating membrane response (NMR) of the rabbit can be incorporated into the neurally inspired model of classical conditioning proposed by Sutton and Barto. The original Sutton-Barto (SB) model was presented in the context of the extensive behavioral literature on NMR conditioning. In essence, the approach we used resulted in an implementation of the SB model that not only describes cumulative effects of training but also response topography. The Sutton-Barto-Desmond (SBD) implementation of the Sutton-Barto model of the connectionist learning describes many features of rabbit nictitating membrane response (NMR) conditioning. In addition to response topography, the model's performance was assessed in terms of rate and terminal levels of learning, simulated interstimulus interval functions, and other criteria. Simulation results indicate that the SBD model's ability to capture conditioned stimulus major features of NMR conditioning is highly constrained by parameters that shape the conditioned stimulus (CS).