Torus-cluster membership, not gridness, is causally load-bearing for path integration

measured in 1 paper

Schoyen et al. train a continuous-attractor path-integration RNN (4096 units) across up to 50 environments and cluster cells by autocorrelogram similarity via UMAP+DBSCAN [schoyen-etal-2023-toroidal-cells-path-integration] A 315-cell cluster whose point cloud is torus-shaped is distinct from the high-gridness-score subset; the two only partially overlap [schoyen-etal-2023-toroidal-cells-path-integration] The torus cluster remaps coherently across environments (shared spacing and orientation) but with a consistent nonzero phase shift, mirroring biological grid-module remapping [schoyen-etal-2023-toroidal-cells-path-integration] Pruning the 315 torus-cluster cells drives path-integration decoding error to the untrained baseline, while pruning equally many high-gridness or random cells barely matters, so torus membership rather than gridness carries the computation [schoyen-etal-2023-toroidal-cells-path-integration]

Structure

Torus

Context

toroidal cell cluster, grid-cell score dissociation, coherent phase remapping, cluster-based ablation, path-integration decoding error

Confirmed in models

CARNN trained on path integration across multiple (1-50) environments · CARNN (Continuous Attractor RNN, Sorscher-style path-integration model)

Method

UMAP, Geometric analysis, Causal interventions (steering)

Papers

Coherently Remapping Toroidal Cells But Not Grid Cells are Responsible for Path Integration in Virtual Agents — Schøyen, Vemund, Pettersen, Markus Borud, Holzhausen, Konstantin, Fyhn, Marianne, Malthe-Sørenssen, Anders, Lepperød, Mikkel Elle2023