Research map / Probabilistic and scientific modeling / Dynamical systems

A simple model of co-emergence of grid and place fields

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Synaptic plasticity learning rules · 0 citations (OpenAlex)

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.

Proposes a Dale's-Law-constrained recurrent network trained to predict the next egocentric-driven sensory observation from masked prior observations, showing grid and place cells co-emerge in separate subpopulations without spatial supervision.

Problem

How grid cells and place cells arise and reinforce each other during development (the chicken-and-egg problem), given that existing models derive one cell type from the other or model only one type in isolation.

Approach

A single recurrent network with Dale's Law, split into input-driven and recurrently driven neurons, trained with one next-sensory-observation prediction objective from masked observations and egocentric motion; grid-like units emerge in the recurrently driven population and place-like units in the input-driven population, with no spatial supervision.

Key results

• Across 10 random seeds the model yields 30.0 ± 19.2 grid-like units and 299.4 ± 4.9 place-like units at the same training step (grid score > 0.3).
• Removing Dale's Law reduces recurrently driven grid cells from 58 to 3 under grid-score threshold 0.3.
• Place cells emerge before grid cells during training, and grid patterns merge into a unified lattice after wall removal while place fields stay stable.

Surprising findings

• Making the bias learnable suppresses all grid-like cells passing the criteria, while place cells still emerge in input-driven neurons.
• Grid-like units and place-like units emerge without any motion input at all (or with allocentric motion), suggesting explicit motion input is not required.

Limitations

• Model produces only a single grid module with similar periods, unlike the discrete modular hierarchy in MEC.
• Grid cells in 3D are only locally ordered and global lattice is not formed; connected-room global alignment is partial.
• Grid classification is sensitive; seed-to-seed variability in grid-like counts is large (s.d. 19.2).

Experimental setup

Datasets: simulated sensory observations from boundary-aware Gaussian random fields simulated smooth random-walk trajectories

Benchmarks: hairpin maze grid fragmentation (replication of prior rodent experiment) two-room wall removal connected rooms 3D volumetric traversal (bat-like grid fields) developmental order of place vs grid emergence

Metrics: grid score spatial information content place-field count and mass locality autocorrelogram symmetry cell counts

Baselines: direct composition of place and grid RNNs unconstrained recurrent weights (no Dale's Law) learnable bias variant

Builds on: path-integration RNN grid cell models place cell emergence via sensory denoising successor representation continuous-attractor models

Closest papers by meaning

• Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.92)
• REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 2025 · similarity 0.92)
• Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.91)
• Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.89)
• Trading Place for Space: Increasing Location Resolution Reduces Contextual Capacity in Hippocampal Codes (NeurIPS 2024 · similarity 0.88)
• Flexible mapping of abstract domains by grid cells via self-supervised extraction and projection of generalized velocity signals (NeurIPS 2024 · similarity 0.88)