A Geometric Explanation of Hippocampal Remapping

The spatial navigation system of rodents spans multiple brain regions and hosts a zoo of different neuron types. In particular, both grid cells in the MEC and place cells in the Hippocampus have responses correlated to location in space. The coding strategies associated with the grid system are provably efficient for encoding location, while for path-integration the place system appears redundant. However, Hippocampal place cells remap between different contexts, and place cell firing fields can vary drastically from room to room. This has been hypothesized to have a range of functions, from error correction to assistance in learning. We postulate that this remapping is informative, and examine this phenomenon via the geometry of neural codes. We find that a place like coding strategy allows one to locally decode both context and location robustly. Further, we estimate an upper bound on the capacity of the place network for storing contexts, and demonstrate the plausibility of our framework by decoding context from numerical simulations of place fields.