CMU-CS-06-170
Computer Science Department
School of Computer Science, Carnegie Mellon University



CMU-CS-06-170

Space and Context in the Rodent
Hippocampal Region

Mark C. Fuhs

November 2006

Ph.D. Thesis

CMU-CS-06-170.pdf


Keywords: Entorhinal cortex, hippocampus, place cells, grid cells, path integration, attractor dynamics, model selection

Recently, cells in the dorsal medial entorhinal cortex (dMEC) were found have spatially modulated activity patterns, comprising multiple active regions organized in a hexagonal lattice across the environment. These "grid cells" show the same field spacing in any environment and are sometimes modulated by the rodent's speed and direction of travel, leading to the hypothesis that dMEC subserves a path integration system. Downstream, hippocampal "place cells" are typically active only within a single contiguous region of an environment. Unlike grid cells, changes either in sensory information or to the rodent's task can cause place cells to "remap," or radically change their activity patterns.

The first part of this thesis presents a neural network model of dMEC grid cells that provides both a cogent explanation of the firing properties of the grid cells and a mechanism by which they could satisfy the computational requirements for path integration. The efficiency and properties of a hippocampal spatial code derived from grid cells is also explored.

The second part considers place cell remapping as a method of encoding context and presents a Bayesian statistical model for context learning. Context learning is defined as the problem of decomposing the rodent's history of experiences into temporal windows within which the distribution of sensory and task-related hippocampal input is statistically stationary. Context learning can therefore be understood as a model selection problem: how many contexts make up the rodent's world? The theory provides an understanding of why remapping sometimes develops gradually over many days of experience, why the time course of reversal learning depends on the degree to which the reward contingencies were changed, and why overlapping sequence learning does not consistently result in "context-dependent" sequence representations.

The third part presents an analysis of a hippocampal physiology experiment, in collaboration with Bruce McNaughton (University of Arizona), that pitted sensory information against path integration information. Rats foraged in two identical, connected boxes with either the same or opposite orientations. The observed pattern of place cell responses suggest that a combination of linear and angular path integration eventually overrides sensory cues, even when linear path integration alone does not.

237 pages


Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu