Spatial navigation and multiscale representation by hippocampal place cells

AbstractHippocampal lesions are known to impair success in navigation tasks. While such tasks could be solved by memorizing complete paths from a starting location to the goal, animals still perform successfully when placed in a novel starting position. We propose a navigation algorithm to solve the latter problem by exploiting two facts about hippocampal organization: (1) The size of the place fields of hippocampal place cells varies systematically along the dorsoventral axis, with dorsal place cells having smaller place fields than ventral (Kjelstrup et. al. 2008); and (2) the theta oscillation propagates as a traveling wave from dorsal to ventral hippocampus (Lubenov and Siapas, 2009). Taken together, these observations imply that the hippocampal representation of space progresses from fine- to coarse-grained within every theta cycle.The algorithm assumes that place cells can be activated by the animal's imagining a goal location, in addition to physically standing in the appropriate location. In the proposed algorithm, place cell activation propagates from small scale to large scale until place cells are found which respond strongly to both the physical location and the goal location. These place fields have their centers aligned roughly in the direction of the goal, providing a crude estimate of which direction the animal should step to approach the goal. Fine-grained directional information is contained in the smaller scale place fields within these large ones. Our algorithm therefore identifies a sequence of place cells, one from each scale, whose centers lie roughly along the line to the goal.Simulations reveal successful navigation to the goal, even around obstacles. By minimizing the number of steps the animal takes to reach the goal, we predict the organization of the optimal place field "map"; specifically the fraction of place cells which should be allocated to each spatial scale. This prediction is, in principle, experimentally testable.The set of place fields with centers lying along a line to the goal is used to compute a step direction by maximizing the probability that those cells will be active in the next time step, given that a particular step direction is chosen.The proposed algorithm handles navigation around obstacles by including "border cells" (Solstad et. al. 2008) which inhibit place cells in proportion to the degree of overlap between the place field and the obstacle. Furthermore, including firing rate adaptation of place cells prevents the animal from getting stuck in one spot.