Research map / Probabilistic and scientific modeling / Dynamical systems
POCO: Scalable Neural Forecasting through Population Conditioning
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.
Introduces POCO, a population-conditioned neural forecaster that pairs a univariate MLP with a POYO-based population encoder via FiLM, trained across five calcium imaging datasets from zebrafish, mice, and C. elegans.
Problem
Forecasting future cellular-resolution neural activity during spontaneous behavior across multiple sessions and individuals, which prior population-dynamics models (mostly interpretability- or decoding-focused, single-session, trial-based) have not addressed.
Approach
A lightweight univariate MLP forecaster whose hidden activations are modulated by FiLM parameters generated from a POYO-style Perceiver-IO population encoder; unit embeddings are reused as queries in the final attention layer so the population drives each neuron's prediction, and tokens span T_C=16 steps to limit token count.
Key results
- Multi-session POCO achieves the highest prediction score on four of five datasets at 512 PCs (e.g., 0.525 on Deisseroth zebrafish vs 0.419 for TexFilter), and 0.364 on Zimmer C. elegans.
- Fine-tuning only unit embeddings of pre-trained POCO reaches reasonable performance within tens of steps, with 200 embedding-tuning steps taking under 15 seconds and 3.5 ms inference.
- Unit embeddings cluster by brain region in mice without anatomical labels; on Zapbench at C=4 POCO matches UNet, but underperforms at C=256.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-session POCO instead of the strongest multi-session baseline | neural forecasting on five datasets · Zebrafish (Deisseroth), 512 PCs | prediction score: 0.440 -> 0.525 | Helped |
| use multi-session POCO instead of the strongest multi-session baseline | neural forecasting on five datasets · Zebrafish (Ahrens), 512 PCs | prediction score: 0.370 -> 0.440 | Helped |
| use multi-session POCO instead of the strongest multi-session baseline | neural forecasting on five datasets · Mice, 512 PCs | prediction score: 0.409 -> 0.420 | Helped |
| use multi-session POCO instead of the strongest multi-session baseline | neural forecasting on five datasets · C. elegans (Zimmer), 512 PCs | prediction score: 0.348 -> 0.364 | Helped |
| use multi-session POCO instead of the strongest multi-session baseline | neural forecasting on five datasets · C. elegans (Flavell), 512 PCs | prediction score: 0.274 -> 0.213 | Hurt |
| replace population-conditioned MLP forecasting with direct POYO forecasting (ablation) | POCO · Zebrafish (Deisseroth), 512 PCs | prediction score: -0.971 -> 0.525 | Helped |
| remove population conditioning from the MLP forecaster (ablation) | POCO · Zebrafish (Deisseroth), 512 PCs | prediction score: 0.417 -> 0.525 | Helped |
| replace the population encoder with a per-neuron univariate Transformer (ablation) | POCO · Zebrafish (Deisseroth), 512 PCs | prediction score: 0.463 -> 0.525 | Helped |
| fine-tune embeddings only rather than the full model (ablation) | pre-trained POCO · Zebrafish (Deisseroth), 512 PCs | prediction score: 0.532 -> 0.539 | No clear effect |
| pre-train on another zebrafish dataset rather than the target dataset (ablation) | POCO fine-tuning · Zebrafish (Deisseroth), 512 PCs | prediction score: 0.497 -> 0.464 | Hurt |
| pre-train on another zebrafish dataset rather than the target dataset (ablation) | POCO fine-tuning · Zebrafish (Ahrens), 512 PCs | prediction score: 0.424 -> 0.349 | Hurt |
| train on multiple species rather than sessions within the same species (ablation) | POCO · Mice, 512 PCs | prediction score: 0.420 -> 0.403 | Hurt |
Surprising findings
- Multi-species training does not consistently help and pre-training on mismatched datasets hurts performance.
- Single-session simulated data favored PLRNN, AR Transformer and TSMixer over POCO, despite the opposite on real data.
Limitations
- Systematic understanding of how calcium indicator, preprocessing, and species affect performance is lacking.
- Multi-dataset and multi-species training does not reliably improve performance.
- Unclear whether unit embeddings are comparable across species or generalize to unseen brain regions.
Experimental setup
Datasets: Zebrafish Deisseroth Zebrafish Ahrens Mice Harvey C. elegans Zimmer C. elegans Flavell Zapbench
Benchmarks: Zebrafish Deisseroth Zebrafish Ahrens Mice Harvey C. elegans Zimmer C. elegans Flavell Zapbench
Metrics: prediction score (relative improvement over copy baseline) MSE MAE
Baselines: MLP NLinear DLinear Latent_PLRNN TSMixer TexFilter AR_Transformer NetFormer ModernTCN UNet
Builds on: POYO POYO+ Perceiver-IO FiLM NLinear
Closest papers by meaning
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.88)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.88)
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.87)
- Population Transformer: Learning Population-level Representations of Neural Activity (ICLR 2025 · similarity 0.86)
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.86)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · 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.