Neural Network Models of Cognitive Maps

Abstract

Neural networks - parallel systems consisting of highly interconnected simple neuron-like processing elements - have become a subject of intense interest to scientists spanning a broad range of disciplines including psychology, physics, mathematics, computer science, biology and neurobiology. A variety of connectionist or neural network models have been proposed, ranging from simplified models to more realistic models of learning, associative memory, and sensori-motor development. Neural networks hold a great promise in the field of spatial cognition since they are capable (in principle) of approximating any real valued function mapping, and have been used to solve complex problems in allied fields such as visual pattern analysis and robotic control. In addition, neural networks have biological relevance, a problem that has plagued the field of symbol processing Artificial Intelligence (AI) systems. Neural networks simulated based on known physiological and anatomical properties of the brain may reveal the process by which groups of neurons interacting according to some local rales undergo self-organization. This paper will examine specific problems in spatial cognition where neural networks have been used and have produced plausible models of spatial behavior. Neural networks have helped in understanding the types of computations that might be performed by the place cells in the hippocampus when the animal moves about and constructs an internal spatial representation. Other problems in spatial cognition such as recognizing places and locating goals further demonstrate the success of neural networks in this domain. A neural network model of route learning is proposed that can learn different routes in an environment and locate a goal given the route information.