Research map / Probabilistic and scientific modeling / Molecules and proteins
On Conformal Isometry of Grid Cells: Learning Distance-Preserving Position Embedding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that learning a conformal isometry (distance-preserving) position embedding with a transformation RNN in a minimal single-module grid-cell setting yields hexagonal grid patterns, and proves the hexagon flat torus is maximally distance-preserving.
Problem
Explaining the mathematical principle behind the emergence of hexagonal periodic grid firing patterns in grid cells, which prior RNN path-integration models only obtain inconsistently and with extra assumptions about place cells.
Approach
Isolate a single grid-cell module with an explicit metric and train a position embedding so that embedded distances scale linearly with physical distance (conformal isometry loss L1) plus a one-step transformation loss (L2) with a generic RNN/linear transformation; theoretically show the hexagonal flat torus minimizes the fourth-order deviation from local isometry.
Key results
- Learned linear model achieves gridness 1.70 with 100% valid rate, versus 0.18/25.2% to 0.90/73.1% for prior learning-based baselines.
- Estimated grid scale is proportional to 1/s (s=5: 0.82, s=10: 0.41, s=15: 0.27).
- Theory shows the hexagon flat torus is the unique isotropic minimizer of the fourth-order deviation from local flatness among flat tori.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add conformal isometry loss (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | Helped |
| Remove non-negativity constraint (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | No clear effect |
| Add the transformation assumption (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | Helped |
| Add normalization constraint (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | Helped |
| Use Leaky ReLU activation (ablation) | Nonlinear grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | No clear effect |
| Use Swish activation (ablation) | Nonlinear grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | No clear effect |
| Allow small deviations from unit norm (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | Mixed |
| Vary the conformal-training displacement range (ablation) | Single-module grid-cell model · Grid-cell response maps | Hexagonal pattern emergence | No clear effect |
| Increase the scaling factor s | Single-module linear model · Single-module linear models | Estimated scale: 0.82 -> 0.41 -> 0.27 | Hurt |
Surprising findings
- Non-negativity of grid activities is not necessary for hexagonal patterns to emerge, per ablation.
- Hexagon patterns emerge across linear and nonlinear transformation models regardless of the specific RNN form.
Limitations
- Minimal setting omits place cells and multi-module interactions in the main analysis
- Conformal isometry holds only approximately for larger displacements (s*|Δx| beyond about 0.8)
- Neural data analysis uses a single rodent module with noisy recordings
Experimental setup
Datasets: Neuropixels recordings of rodent grid cells (rat R, day 1, module 1) from prior study
Benchmarks: gridness score on learned response maps neural recording data linearity analysis
Metrics: gridness score valid rate estimated scale vs scaling factor s conformal isometry deviation
Baselines: Banino et al. (2018) grid cells RNN Sorscher et al. (2019) Gao et al. (2021) conformal isometry model
Builds on: Adam conformal isometry hypothesis (Gao et al.) continuous attractor neural networks
Closest papers by meaning
- Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.87)
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.87)
- Flexible mapping of abstract domains by grid cells via self-supervised extraction and projection of generalized velocity signals (NeurIPS 2024 · similarity 0.87)
- Not so griddy: Internal representations of RNNs path integrating more than one agent (NeurIPS 2024 · similarity 0.85)
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.85)
- Hitting Time Isomorphism for Multi-Stage Planning with Foundation Policies (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.