Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Place Cell Representations and Context-Dependent Remapping
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 similarity-based objective that maps spatial (and contextual) proximity to representational similarity, showing feedforward and recurrent networks learn place-like, remapping-like, and band-like representations, and that orthogonal transformations of learned maps yield new remapped maps without relearning.
Problem
How place-cell-like representations emerge in neural networks, how they encode context jointly with space, and how they remap across contexts, which remains poorly understood.
Approach
A minimal objective that matches Gaussian similarity of physical positions (and context scalars) to the Gaussian-like similarity of network population vectors with a lower bound beta and L2 activity regularization; the objective is invariant to orthogonal transformations, so rotating/permuting a learned map yields new distinct maps (remapping) without retraining.
Key results
- Feedforward networks trained on the objective learn place-like units with field size controlled by sigma; the linear decoder gives best mean error while population decoding has lowest maximum error, with best-case mean error of a few percent of arena size.
- Across sufficiently different contexts, median spatial correlation between active units tends to zero, consistent with global remapping.
- Best-fit orthogonal transformations achieve reconstruction errors substantially lower than a random orthogonal baseline across learned contexts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with the spatial similarity objective | Feedforward network · 2×2 square enclosure | place-like spatial tuning | Helped |
| Train with the joint space-context similarity objective | Feedforward network · Multiple contexts | spatial correlation across contexts | Helped |
| Train with the spatial similarity objective | Recurrent network · 2×2 square enclosure | spatial tuning | Helped |
| Use population decoding instead of a linear decoder | Recurrent network · Trajectories over time | position decoding error | Mixed |
| Train from velocity inputs without explicit positional information | Long-sequence recurrent network · 2×2 square enclosure | place-like spatial tuning | Helped |
| Increase scale parameter σ (ablation) | Feedforward network · Spatial encoding | place-field size | Helped |
| Remove activity regularization (λ = 0) (ablation) | Feedforward network · Spatial encoding | place-like tuning | Hurt |
| Set the similarity lower bound to zero (β = 0) (ablation) | Feedforward network · Spatial encoding | place-like tuning | Mixed |
| Set β = 0 and remove activity regularization (λ = 0) (ablation) | Feedforward network · Spatial encoding | place-like tuning | Hurt |
| Replace squared-distance Gaussian similarity with exponential Euclidean-distance similarity (ablation) | Feedforward network · Spatial encoding | place-like spatial tuning | Mixed |
| Increase the similarity lower bound β (ablation) | Feedforward network · Spatial encoding | number of place fields | Helped |
| Use nonzero β (ablation) | Joint space-context feedforward network · Multiple contexts | unit activity across contexts | Helped |
Surprising findings
- Place-like tuning requires either a nonzero similarity lower bound beta or activity regularization, but not both; with both removed, units change linearly across the arena.
- Permutations induce more strongly decorrelated representations than reflections or rotations.
Limitations
- Requires explicit location labels and preprocessed scalar context signals during training, which may be biologically implausible.
- Uses rate coding only; unclear extension to spiking networks.
- Relies on a specific Gaussian similarity measure and lower bound beta, not validated against experimental data.
Experimental setup
Datasets: simulated trajectories in 2x2 square enclosure
Metrics: training loss spatial correlation between ratemaps mean Euclidean position decoding error Procrustes fit error
Baselines: linear decoder Top n decoding
Builds on: Dorrell et al. 2022 normative spatial encoding Schaeffer et al. 2023 Sorscher et al. 2022 Kanerva hyperdimensional computing
Closest papers by meaning
- Trading Place for Space: Increasing Location Resolution Reduces Contextual Capacity in Hippocampal Codes (NeurIPS 2024 · similarity 0.93)
- Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.91)
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.90)
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.87)
- REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 2025 · similarity 0.87)
- Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.86)
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.