Stochastic learning in neural network models of categorization

A number of neural network models of categorization have been proposed. The models differ notably in the types of internal representation used (e.g. exemplars vs. prototypes; hyperplane vs. hypersphere activation regions). However, many of these NN models of categorization (e.g., ALCOVE) use some form of gradient method for learning. These methods have been successful in reproducing group learning curves, but tend to underpredict variability in individual-level data, for both accuracy and attention allocation measures (Matsuka, 2002). Here, we show that use of a different learning algorithm with a given model can result in different learning trajectories and more realistic variability in individual learning curves, especially for attention allocation. Our proposed algorithm is a form of constrained simulated annealing (Ingber, 1989). Initial parameter sets (dimensional attention weights and network connection weights) are randomly selected. At the beginning of each training epoch, a hypothetical “move” in the parameter space is computed by adjusting each parameter by an independently sampled term. These adjustment terms are drawn from a prespecified distribution (e.g., a Cauchy distribution). The move (i.e., the set of new parameter values) are accepted or rejected, based on the computed relative fit of the new values. Specifically, if the new parameter values result in better fit, they are accepted. If they result in worse fit, they are accepted with some probability P. The adjustment in parameters is very rapid initially, and it gradually decreases over learning blocks.