Research map / Probabilistic and scientific modeling / Dynamical systems
REMI: Reconstructing Episodic Memory During Internally Driven Path Planning
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 single-layer RNN model of the HC-MEC loop in which place cells autoassociate sensory (SMC) and grid cell patterns, enabling cue-triggered goal recall, grid-based path planning, and reconstruction of sensory experience along planned routes, validated in RatatouGym and Habitat-Sim.
Problem
How grid cells, place cells, and sensory-responsive MEC cells jointly support internally driven navigation: recalling goal locations from partial cues and planning novel routes, including shortcuts through unvisited locations.
Approach
Place cells autoassociate sensory inputs with grid cell patterns so sensory cues recall goal grid patterns; planning is performed on the grid cell manifold by a planner subnetwork driving speed and direction, with sensory experiences reconstructed from intermediate grid states via hippocampal pattern completion; per-module displacement decoding is averaged across scales.
Key results
- Recall: all nine masking-ratio models recalled the goal location from SMC cues with high accuracy; higher masking ratio increased HPC count and decoding accuracy but reduced sensory recall.
- Planning: a planner trained on only 128 fixed start-end pairs over 100 steps generalized to plan paths between locations reachable within 10 s, including shortcuts.
- Habitat-Sim: decoded intermediate SMC embeddings (1024-dim from BtnkMAE) were decoded into images closely resembling expected views along the planned route.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase masking ratio (ablation) | REMI HC-MEC recall model · Sensory-cue recall | GC location decoding accuracy | Helped |
| Increase masking ratio (ablation) | REMI HC-MEC recall model · Sensory-cue recall | sensory observation recall | Hurt |
| Train the planner with 128 rather than 48 grid cells per module (ablation) | REMI GC-only planner · Long-range path planning | planning accuracy | Helped |
| Use HC-MEC auto-associative dynamics with fewer grid cells (ablation) | REMI HC-MEC planner · Planned trajectories | trajectory smoothness | Helped |
| Use HC-MEC auto-associations to reconstruct SMC states from intermediate GC states | REMI HC-MEC planner · Planned routes | sensory-state reconstruction | Helped |
| Replace simulated SMC rate maps with Habitat visual feature embeddings (ablation) | REMI · Habitat-Sim navigation | reconstructed-view correspondence | Helped |
| Vary direction-cell tuning width and magnitude (ablation) | REMI path-integration and planning model · Planning | planning performance | No clear effect |
Surprising findings
- Grid cells emerged after place cells during training even with partial supervision, and place fields became more refined once grid patterns stabilized.
- Place-like units emerged without a firing-rate constraint on HPCs, unlike prior work.
Limitations
- Grid cells were partially supervised rather than emerging on their own, to avoid simultaneous emergence complexity.
- No boundary avoidance; the framework assumes environment scale similar to the largest grid scale.
- Long-range planning may require other brain regions beyond grid cells.
Experimental setup
Datasets: ImageNet-1k Habitat Synthetic Scene Dataset (HSSD)
Benchmarks: RatatouGym Habitat-Sim
Metrics: L2 distance between decoded and ground-truth position decoded trajectory alignment with planned GC path reconstructed image quality along planned route
Baselines: GC-only planning decoding (prior grid-cell vector models) SR-based planning
Builds on: Masked Autoencoder Successor representation Path integration RNN (Banino et al. / Sorscher et al.)
Closest papers by meaning
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.92)
- Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.90)
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.89)
- Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.89)
- Trading Place for Space: Increasing Location Resolution Reduces Contextual Capacity in Hippocampal Codes (NeurIPS 2024 · similarity 0.89)
- A Multi-Region Brain Model to Elucidate the Role of Hippocampus in Spatially Embedded Decision-Making (ICML 2025 · similarity 0.87)
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.