ABSTRACT
The architectures and mechanisms of learning are becoming more important to the design of machines that will work on more complex and open problems. Biological systems are being studied as the prior art for such designs, but several controversies still remain about the role of habituation in conditioning, the generation of responses that cannot be defined a priori, and the reality of a new computational rubric within actual nervous systems. As a bridge between recent discoveries in both computer and neural sciences, this work presents a quantitative model of habituation and sensitization, most notably found in the mollusc Aplysia, but as part of a larger dynamics analogous to connectionism’s computation of energy contours and their control by simulated annealing. Pavlovian conditioning is modeled as the control of propagation through laterally connected reflex arcs. These arcs pass through two layers: The alpha layer computes temporal contingencies according to activity-dependent sensitization and habituation. The beta layer stores an energy contour similar to a dominant focus. Feeding forward to the beta layer, the alpha layer controls both stimulus propagation through the contour and the formation of the contour itself by potentiation. Negative feedback from the beta layer provides the control function analogous to simulated annealing. This model demonstrated habituation, classical conditioning, extinction, and spontaneous recovery. Both alpha and beta conditioning were produced, but, in contrast to alpha conditioning, the beta response learning curve showed initial positive acceleration and much higher final asymptote. Extinction and spontaneous recovery were demonstrated as properties of habituation. The model’s limitations are considered for the future elaboration of a more complete and general mechanics.
