Research map / Probabilistic and scientific modeling / Dynamical systems
Flow-field inference from neural data using deep recurrent networks
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 FINDR, an unsupervised sequential VAE with a gated neural SDE drift that infers low-dimensional, nonlinear, stochastic flow fields from population spike trains, separating task-relevant latents from a task-irrelevant time-varying bias.
Problem
Estimating low-dimensional nonlinear neural population dynamics from spike trains is hard, and existing methods make restrictive assumptions (autonomous, linear, switching, or high-dimensional) and can give inconsistent dynamics across data splits.
Approach
A sequential VAE whose prior dynamics are a gated MLP drift inside an Euler-Maruyama SDE, with a fixed time-varying bias absorbing task-irrelevant activity so the low-dimensional latent captures task-relevant flow fields and attractors.
Key results
- FINDR's co-smoothing log-likelihood is consistently higher than SLDS, rSLDS, autoLFADS and GPFA when latent dimension L < 3.
- Across 5 CV folds, FINDR flow fields are consistent (<|r|>=0.94-0.99) while autoLFADS trajectories are not dynamically consistent (<|r|>=0.22 after alignment).
- On synthetic data, FINDR recovers an approximate 2D disk attractor, with log-likelihood saturating at L=2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FINDR instead of autoLFADS without a latent-space alignment | Rat frontal cortex decision-making dataset · Neural trajectories across folds 1–2 | trajectory consistency ⟨|r|⟩: 0.53 -> 0.99 | Helped |
| Linearly transform autoLFADS fold 1 to match fold 2 | autoLFADS · Rat frontal cortex decision-making dataset, folds 1–2 | trajectory consistency ⟨|r|⟩: 0.53 -> 0.99 | Helped |
| Affine transform fold-1 steady-state trajectories to match fold 2 | autoLFADS · Rat frontal cortex decision-making dataset, steady-state trajectories across folds | trajectory consistency ⟨|r|⟩: 0.99 -> 0.22 | Hurt |
| Use FINDR instead of autoLFADS for steady-state dynamics across folds | Rat frontal cortex decision-making dataset · Steady-state trajectories across folds | trajectory consistency ⟨|r|⟩: 0.22 -> 0.94 | Helped |
| Add a task-irrelevant time-varying component (ablation) | FINDR · Rat frontal cortex decision-making dataset | dynamical representation consistency and interpretability | Helped |
| Use FINDR with fewer than 3 latent dimensions instead of existing methods | FINDR versus SLDS, rSLDS, autoLFADS, and GPFA · Rat frontal cortex decision-making dataset | co-smoothing log-likelihood | Helped |
| Use FINDR to infer latent dynamics instead of autoLFADS | FINDR versus autoLFADS · Rat frontal cortex decision-making dataset | presence of two slow-point clusters across folds | Helped |
Surprising findings
- autoLFADS reaches topological trajectory consistency (<|r|>=0.99 after linear transform) yet its underlying dynamics do not show two clusters of steady states, indicating topological consistency does not imply dynamical consistency.
Limitations
- Performs best with many simultaneously recorded neurons and many repeated trials; performance may vary across datasets.
- Dynamical consistency across folds is not theoretically guaranteed and must be checked empirically.
- Learning task-irrelevant components separately is needed for consistent interpretable dynamics; without it results are less interpretable.
Experimental setup
Datasets: Rat dmFC/mPFC auditory decision-making recordings (Luo et al. 2023) Synthetic n-bit flip-flop style spike trains
Benchmarks: Rat dmFC/mPFC decision-making dataset (co-smoothing) Synthetic disk and rectangular attractor spike trains
Metrics: normalized log-likelihood (co-smoothing) evidence-sign conditioned PSTH R2 consistency <|r|> across folds PCA variance explained
Baselines: SLDS rSLDS autoLFADS GPFA CEBRA
Builds on: Kim et al. 2023 gated neural SDE LFADS beta-VAE GPFA rSLDS
Closest papers by meaning
- Inference of Neural Dynamics Using Switching Recurrent Neural Networks (NeurIPS 2024 · similarity 0.87)
- BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data (ICLR 2025 · similarity 0.86)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.86)
- Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics (NeurIPS 2024 · similarity 0.85)
- Inferring stochastic low-rank recurrent neural networks from neural data (NeurIPS 2024 · similarity 0.85)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.85)
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.