Towards a Computational Theory of Rat Navigation

A century of behavioral studies has generated an abundance of proposals for how animals represent and navigate through space. Recently, neurophysiological recording in freely-behaving animals has begun to reveal cellular correlates of these cognitive processes, such as the existence of place cells in hippocampus and head direction cells in postsubiculum and parietal cortex. We propose computational mechanisms to explain these phenomena. A variety of computer models have demonstrated place cell-like responses, given inputs that encode distance and/or bearing to one or more landmarks. These models utilize machine learning algorithms such as competitive learning [Sharp 1991], recurrent backpropagation/Elman nets [Shapiro & Hetherington 1993, Hetherington & Shapiro 1993], genetic algorithms[Treves et al. 1992], competitive learning with radial basis units [Burgess et al. 1993], and specialized architectures employing a combination of delta rule and radial basis units [Schmajuk & Blair 1993]. The problem with all of these models is that their processing is mainly a function of visual input. The experimental literature clearly shows that hippocampal processing is not that simple. Specifically, although place fields are sensitive to visual input (they rotate in agreement with rotation of distal visual cues), place cells remain active when the lights are turned out, and place fields can form when the animal explores novel environments in the dark. Place cells also continue to fire when distal landmarks are removed, but permutation of landmarks causes the animal to behave as if it were in an unfamiliar environment. Finally, place cell firing may be dependent on head direction, at least under certain conditions. An acceptable model of place memory must allow the “current place” to be updated by non-visual means such as motor feedback, and must be both sensitive to visual cues and robust in their absence. We propose a computational theory of the core of rat navigation abilities, based on coupled mechanisms for path integration, place recognition, and maintenance of head direction. We assume the rat has a path integration system (see [Etienne 1987, Mittelstaedt & Mittelstaedt 1980]) that is able to keep track of its current position relative to selected reference points. We postulate that hippocampal pyramidal cells form place descriptions by learning correlations between perceptual inputs and the rat’s internal states, which include the output of the path integrator. Place codes are associated with landmark bearings, so that visual cues can recall previously stored directional information in a manner similar to McNaughton’s local view hypothesis [McNaughton 1989]. We describe a connectionist implementation of this theory. Our model reproduces a variety of experimental observations, including reset of head direction in response to visual input, persistence of place fields in the absence of visual input, and modulation of place cell directional sensitivity. We compare our theory with other theories of hippocampal function and offer some predictions based on the model.