A mechanistic model for the formation of globally consistent maps of space in complex environments

When animals are placed in multi-compartment or complex environments, entorhinal grid cells exhibit fragmentation of their highly regular spatial tuning curves, with a disjoint grid map for each compartment. However, as the animal explores and increases familiarity with the environment’s geometric and topological structure, the grid maps gradually approach a single globally consistent periodic map [1,2]. Here we show that Hippocampal-MESH — a novel neocortical-hippocampal-entorhinal recurrent circuit model that we introduce in parallel submitted abstracts — can, with the addition of slow plasticity in the hippocampal-entorhinal synapses, reproduce experimental findings of fragmented grid maps evolving to locally and then globally consistent maps at different timescales.
We consider environments with two chambers separated by a barrier, with distinct grid maps, then remove the barrier and model the dynamics of grid cell tuning in response to the merger of the two chambers into one. If there are landmarks signaling the removed boundary location, the learned discontinuity between the two maps initially leads to abrupt remapping when traversing across the removed boundary, with the original chambers' maps near the far walls of the space. Plasticity between the representations on both sides of the former boundary gradually leads, at intermediate times, to a smooth interpolation of the two maps. At longer timescales, the tendency to generate local consistency of grid maps cascades through the generated representations, driving shifts in the map which eventually evolve to a globally consistent grid map, with phases in between the original maps of the two environments before barrier removal.
These results encapsulate key findings of entorhinal map plasticity, thereby providing the first neocortical-hippocampal-entorhinal model that captures grid map fragmentation at short timescales (as described in a parallel submitted abstract), interpolation at intermediate timescales and map merger at long timescales all enshrined within the same architecture.