Integration and multiplexing of positional and contextual information by the hippocampal network
Fig 4
CANN model for interplay between path integrator, external stimuli, and memory.
A. Phenomenology of the CANN model. The model is composed of a recurrent hippocampal network that has memorized two cognitive maps (place fields dispositions) denominated A and B, a path integrator input (PI), and a visual input (V). In the left panel both PI and V are activating place cells whose place-field centers (shown by blue dots) correspond to the position of the rodent (red cross X) in the cognitive map A. The activity is said to be localized in a bump around X in t map A, while it appears as sparse and uninformative if interpreted with respect to the place-field locations in map B. A feedback projection (blue arrows) from the hippocampal state (bump) to the path-integrator state (purple) maintains the stability of the system by enforcing that the retrieved hippocampal map and the PI state agree. After the light conditions have been switched (teleportation, red line), V projects on place cells encoding position X in map B. The two hippocampal cognitive maps are therefore in conflict, and the bump of activity is alternatively localized in A (center-top) or in B (center- bottom). When the hippocampal activity is localized in the cognitive map B, the feedback projection tries to realign the internal state of PI along the corresponding map. Once the realignment has succeeded (green line), both inputs are back to a coherent state, and stability is reached in the cognitive map relative to the post-teleportation external light conditions. B. Representation of the effective model for the activity bump and effects of parameters. The input strengths, γPI and γV, contribute to push the hippocampal activity towards the corresponding cognitive states. Increasing the strength of recurrent connections, γJ, results in an effective barrier separating the two collective hippocampal states, giving rise to well formed bumps in either map A (left) or in map B (right). Due to this effective trap the bump state remains localized in either map for more than a single theta bin. C. Temporal correlation of flickering events (theta bins with cognitive state opposite to external light conditions) decay over c.a. 7 theta bins in simulated data. Same model parameters as in Fig 3. D. Time trace of the log-likelihood difference (Eq (13) in Methods) in a simulated teleportation session; flickers can be observed during the conflicting phase following teleportation (shaded region). Prior to teleportation, and after the PI is realigned, inputs are coherent, and the system is stable: the sign of
is constant, mirroring the localization of the bump in one map. Screenshots of the activity projected on the two cognitive maps are shown for five different times. From left to right: bump localized in map A, bump localized in map A during the conflicting phase, mixed state during the conflicting phase, bump localized in map B during the conflicting phase, bump localized in map B. The video of the simulation can be found in SI.