Neural Network Models of Cognitive Conflict Paradigms

Abstract

In recent years, some areas of cognitive psychology have proposed formal models in the form of computer simulations, using Back-Propagation Artificial Neural Networks (BP-ANNs). Such models represent an improvement in plausibility, and they allow quantitative results to be compared with empirical data.--- After learning, BP-ANN's cells exhibit a fixed input/output behavior. Using the black box method, this is shown to be a fundamental problem. Cells without an internal state to represent short-time memory cannot account for the sequence of stimuli, nor for the time elapsed between stimuli. As a consequence, BP-ANNs -as well as other neural networks without state-dependant input/output- are inadequate as models of some important cognitive processes. These include classical conditioning, operant conditioning, and sequence effects in cognitive control.--- Another major methodological problem is the use of free parameters. BP-ANNs cognitive models frequently use arbitrary amounts of cells, amounts of layers, connection structure, learning parameter value and other characteristics, without giving a theoretical justification.--- Three methods are proposed here to solve these problems: first, the black box method is used to produce a cell's input/output behavior more similar to that of neurons. Second, the reverse engineering method is used to simulate as many neural features as possible. And, third, a genetic algorithm is used to eliminate arbitrary free parameters.--- The use of these methods is illustrated through a series of spiking neural network models, implementing state-dependant input/output, spikes, refractory period, temporal summation, axon delay and synchronization of neuron groups. A genetic algorithm is used to choose the parameter values in another series of models.--- Finally, the feasibility of following this research strategy using parallel computer hardware is discussed.