Multiple representations of space underlying behavior

Abstract

AbstractWe argue that a map is meaningless unless we have a process for using it. Thus, in this paper, we not only offer the world graph as a representation of relationships among situations the animal has encountered and may encounter again, but we also offer algorithms for how the information encoded in the world graph may be used by the animal in determining its behavior. Each node of the graph encodes a recognizable situation in the animal's world, but a given place may well be encoded in a number of different nodes. Nodes not only require algorithms for the recognition of the situation; they store information about drive reduction associated with the encoded situation. We note that the use of graphs as a basis for exploring some search space is well known in artificial intelligence (AI), but we stress the importance of the animal's exploration of its environment for growing the graph, as distinct from the mathematically described potential nodes frequent in AI search spaces. To explore a number of hypotheses about the way information in the world graph is used to guide the animal's movement, we recall a number of classical experiments on maze exploration by animals, and use them to argue for the nonlocal hypothesis (selection of a path does not depend only upon information about the immediate environment of the animal) and the competing nodes hypothesis (more than one drive may enter into the determination of the animal's behavior at any time). An important feature of the model is that it yields exploration and latent learning without the introduction of an exploratory drive. We also note that the performance exhibited by the model appears to be state-dependent when the animal operates under high drive levels. The drive-interaction matrix is offered as a subject for future research.We complement the presentation of the world graph model, its drive dynamics, and how these are constrained by experiments on maze behavior, with a brief analysis of maps in the brain. We distinguish egocentric maps–which we relate to the many visual systems–from allocentric maps. We offer a somewhat unconventional view of short-term and long-term memory. We examine cooperative computation in somatotopically organized networks, relating this to visually guided behavior in the frog, and to the interaction of colliculus and cortex in the control of eye movements. We examine, but do not advocate, the hypothesis that the hippocampus is a cognitive map. We do stress that if it is a cognitive map, it must be seen as a chart of the local neighborhood, rather than the whole atlas; and we note that the cognitive map hypothesis would lead one to expect the region to exhibit activation of place cells before the animal leaves the previous place.