Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Place Cells and Grid Cells Refine Each Other: A Clone-Structured Cognitive Graphs Model

Yicong Zheng¹, Thomas Miconi¹, Zacharie Bugaud¹, Kevin L McKee¹, Miguel Lazaro-Gredilla¹, Dileep George²; ¹Astera Institute, ²Vicarious AI

Presenter: Yicong Zheng

Place cells and grid cells support complementary roles in spatial cognition, yet the computational mechanisms underlying their mutual refinement remain unclear. Using Clone-Structured Cognitive Graphs (CSCG), a cloned HMM of hippocampal-entorhinal interactions, we demonstrate three reciprocal effects. First, augmenting sensory observations with grid codes improves learned spatial graph fidelity to better reconstruct the room layout. Second, the learned place cell graph can stabilize corrupted grid codes: even when grid input is fully noisy, the model recovers grid phases with high accuracy, consistent with hippocampal necessity for grid cells. Third, when a model trained in room A is tested in room B (a deformed version of room A), the recovered grid pattern exhibits boundary-anchored phase shifts selectively along the deformed axis, mirroring boundary-tethered rescaling observed in rodents. These results provide a unified account of reciprocal refinement of place and grid cells.

Topic Area: Memory, Learning & Knowledge Structures