| name | topological-grid-cell-decoding-codes |
| description | Topological decoding of grid cell activity via path lifting to covering spaces. Uses TDA to extract toroidal coordinates from grid cell populations and reconstructs spatial trajectories from a single module without training data or external position info. Activation: grid cells, topological data analysis, toroidal manifold, path lifting, spatial navigation, neural manifolds, continuous attractor network, entorhinal cortex, path integration, covering space |
| metadata | {"arxiv_id":"2510.16216","published":"2025-10-17","updated":"2026-07-08","authors":"Yuxing Jared Yao, Iris H. R. Yoon","tags":["grid cells","topological data analysis","toroidal manifold","path lifting","spatial navigation","neural manifolds","entorhinal cortex","covering spaces"]} |
Topological Decoding of Grid Cell Activity via Path Lifting to Covering Spaces
Core Concept
Grid cells in the medial entorhinal cortex encode spatial position on a toroidal manifold due to their periodic firing patterns. This paper introduces a purely topological framework to decode spatial trajectories from grid cell population activity without training data, external position labels, or supervised learning — using only the intrinsic topological structure of the neural manifold.
Key Innovations
1. Toroidal Coordinate Extraction via TDA
- Use topological data analysis (persistent homology) to identify the toroidal structure in high-dimensional grid cell population activity
- Extract toroidal coordinates (angular parameters on the torus) from the neural manifold
- Works on both simulated continuous attractor networks (CANs) and experimental recordings
2. Path Lifting to Covering Spaces
- The toroidal manifold is periodic — a single point on the torus corresponds to infinitely many physical positions
- Path lifting to the covering space (Euclidean plane ℝ²) resolves this ambiguity
- Reconstructed trajectories differ from ground truth only by an affine transformation (rotation, scaling, translation)
- Demonstrates that a single grid cell module contains sufficient information for path integration
3. Training-Free Decoding
- No neural network training, no calibration, no external position reference
- Entirely based on the topological structure of the population code
- Validated on both CAN simulations and real experimental grid cell recordings
Methodology
Step 1: Extract Neural Manifold
- Collect grid cell population activity vectors (firing rates across cells at each time point)
- Apply dimensionality reduction (e.g., PCA, UMAP) to identify low-dimensional embedding
- Verify toroidal topology via persistent homology (two independent 1-cycles = torus)
Step 2: Extract Toroidal Coordinates
- Map neural activity to angular coordinates (θ₁, θ₂) on the torus
- Use circular statistics to track the position on each cycle of the torus
- Handle wrapping/discontinuities at torus boundaries
Step 3: Path Lifting
- Starting from an initial position, integrate angular changes over time
- Lift the path from the torus T² to the covering space ℝ²
- Track unwrapping events when the path crosses torus boundaries
- Result: reconstructed trajectory in physical space (up to affine transform)
Step 4: Validation
- Compare reconstructed trajectory to ground truth via Procrustes analysis
- Measure correlation, angular error, and path similarity
- Validated on: (a) CAN simulations with known ground truth, (b) experimental recordings
Key Findings
- Single-module sufficiency: One grid cell module contains enough information for reliable path reconstruction
- Affine equivalence: Reconstructed paths match ground truth up to rotation + scaling + translation
- No training needed: Purely topological, no supervised learning or calibration
- Robust to noise: Works on experimental data with biological variability
- Path integration mechanism: Suggests how the brain performs path integration from grid cell activity alone
Applications
- Spatial navigation research: Decoding animal trajectories from neural recordings
- BCI for navigation: Brain-computer interfaces for spatial state estimation
- Neuroscience theory: Understanding how grid cell population codes represent space
- Robotics: Bio-inspired navigation systems using topological representations
Implementation Considerations
Data Requirements
- Grid cell population recordings (≥ ~50 cells recommended for reliable TDA)
- Sufficient spatial coverage (animal should explore environment thoroughly)
- Single module identification (cells with similar grid spacing/orientation)
TDA Parameters
- Persistence threshold: filter out topological noise
- Window size for sliding window analysis
- Choice of distance metric for point cloud construction
Pitfalls
- Multiple modules: If multiple grid modules are mixed, the manifold structure becomes more complex (higher-dimensional torus)
- Boundary effects: At environment boundaries, grid cell firing patterns may distort
- Temporal resolution: Path lifting requires sufficient temporal sampling to resolve wrapping events
- Ambiguity: Path lifting recovers trajectory up to affine transform — absolute position and orientation cannot be determined without additional information
Related Concepts
- Continuous attractor networks (CANs) for grid cells
- Persistent homology / topological data analysis
- Covering spaces and path lifting in algebraic topology
- Path integration in the entorhinal-hippocampal system
- Neural manifold analysis
References
- Paper: arXiv:2510.16216 (October 2025, updated July 2026)
- Authors: Yuxing Jared Yao, Iris H. R. Yoon
- Categories: q-bio.NC, math.AT