Time-Course of Processing in Feed-Forward Connectionist Networks: An Analysis of the Exclusive-Or Problem

Abstract

Although most of the work based on feed-forward connectionist networks has emphasized accuracy, time-course of processing issues can also be considered by processing the inputs in cascade. In this Monte Carlo simulation study, two functional classes of solutions of the exclusive-or problem that arise in a feed-forward network comprised of two input units. two hidden units, and one ouput unit are described. These two classes of solutions are distinguished by the time-course of activation of the output unit in response to one particular input: for one class, the activation of the output unit is virtually unchanged over time, but for the other class, the activation initially increases before it subsequently decreases. After examining the basis for this distinction, networks with different architectures are considered to determine the effect of increasing as well as decreasing the degrees of freedom. In the discussion, the implications of a time-course of activation that changes nonmonotonically in feed-forward networks without, as well as with, lateral inhibitory connections for simulating fast errors in choice-response time experiments (e.g., noise/compatibility and word naming) are examined. The implications or functionally distinct solutions in simulating individual differences are also considered.