Research map / Probabilistic and scientific modeling / Dynamical systems
Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences
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 place-like firing fields emerge in a recurrent autoencoder (a CA3 model) trained to reconstruct temporally continuous, partially occluded sensory experiences from simulated agents traversing rooms, without explicit sparsity constraints.
Problem
How place cells, with remapping, reversion, orthogonal room representations and drift, could arise from CA3 episodic pattern completion of sensory experiences during spatial traversal.
Approach
Model CA3 as a continuous-time recurrent autoencoder trained to pattern-complete temporally continuous, weakly spatially modulated sensory experience vectors with an MSE reconstruction loss plus a firing-rate penalty; place fields emerge in the hidden layer, and remapping/reversion is driven by reorganization of recurrent connections.
Key results
- About 80% of active hidden units have spatial information content above 5 bits, roughly 20x the SIC of the WSM inputs (0.29).
- Across rooms, mean population-vector correlation is 0.164 for different rooms versus 0.710 for the same room, qualitatively matching Alme et al.'s reported pattern.
- Recurrent weights reorganize much more than input weights during remapping (s=0.564 for W^rc vs 0.169 for W^in), and reversion correlation rises from 0.35 in cycle 3 to 0.92 in cycle 30.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train an RAE to complete partial, noisy sensory experiences with a firing-rate constraint | RAE hidden layer · Single-room simulated traversal | Spatial information content (SIC): WSM inputs: 0.29 ± 0.01; ~80% of active hidden units have SIC greater than 5 | Helped |
| Train across repeated visits to multiple rooms | RAE · 20-room representations, cycles 2–3 | Pearson correlation between same-room versus different-room representations: different rooms: 0.164 ± 0.029; same rooms: 0.710 ± 0.097 | Mixed |
| Continue training over more room cycles | RAE · Familiar room re-entry | Pearson correlation with previous firing patterns: cycle 3: r=0.35; cycle 30: r=0.92 | Helped |
| Use recurrent connections to support remapping and reversion | RAE · Room transitions T1R1–T2R2 and T2R2–T3R1 | Reorganization score: Wrc: s = 0.564 and s = 0.430; Win: s = 0.169 and s= 0.141 | Helped |
| Increase episodic memory segment duration (ablation) | RAE · Single-room place-field emergence | Number of place cells | Hurt |
| Increase trial duration (ablation) | RAE · Single-room place-field emergence | Number of place cells | Hurt |
| Increase the firing-rate loss coefficient (ablation) | RAE · Single-room place-field emergence | Place-cell proportion among active cells | Helped |
| Increase recall length (ablation) | RAE · Single-room place-field emergence | Number of place cells and active cells | Helped |
| Vary maximum WSM firing rate or WSM smoothing sigma (ablation) | RAE · Single-room place-field emergence | Place-field emergence | No clear effect |
| Reduce the number of WSM signals to 100 or fewer (ablation) | RAE · Single-room place-field emergence | Place-field emergence | Hurt |
| Vary the number of hidden units and WSM signals (ablation) | RAE · Single-room place-field emergence | Place-field emergence | Mixed |
| Vary the sensory masking ratio (ablation) | RAE · Single-room place-field emergence | Place-field emergence | No clear effect |
Surprising findings
- Place fields emerge without explicit sparsity constraints, unlike the sparse autoencoder of Benna and Fusi.
- Place fields only emerged when the number of WSM signals exceeded 100 in the parameter sweeps.
Limitations
- Results are from simulated agents and environments; no direct comparison statistics with rodent place cell data.
- Place field count decreases with longer trials as the network overfits a single room with unrealistic firing rates.
- Number of hidden units needed for place field emergence grows with the dimension of the experience vector; optimal input-to-hidden ratio left open.
Experimental setup
Datasets: simulated weakly spatially modulated sensory experience maps (Gaussian random fields) in simulated rooms
Metrics: spatial information content Pearson correlation of population vectors place field reversion correlation reorganization score of weights percentage of place cells
Baselines: Benna and Fusi (2016) sparse autoencoder
Builds on: Benna and Fusi sparse autoencoder place cells continuous-time RNN grid cell emergence
Closest papers by meaning
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.92)
- Trading Place for Space: Increasing Location Resolution Reduces Contextual Capacity in Hippocampal Codes (NeurIPS 2024 · similarity 0.91)
- Learning Place Cell Representations and Context-Dependent Remapping (NeurIPS 2024 · similarity 0.91)
- REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 2025 · similarity 0.90)
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.90)
- A Model of Place Field Reorganization During Reward Maximization (ICML 2025 · similarity 0.88)
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.