Research map / Probabilistic and scientific modeling / Dynamical systems
Inference of Neural Dynamics Using Switching Recurrent Neural 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.
Proposes Switching Recurrent Neural Networks (SRNNs), RNN-based switching state-space models trained by amortized variational inference, to infer discrete behavioral states and nonlinear neural dynamics from monkey reaching and mouse widefield calcium imaging data.
Problem
Linear switching models (HMM, SLDS, rSLDS) may not capture the nonlinear dynamics of neural population activity that change across behavioral stages; existing nonlinear switching models lack RNN-based dynamics constrained to reconstruct and predict neural data.
Approach
Switching RNNs whose recurrent weights switch via a Markov discrete latent state (transition dependent on previous continuous state), trained end-to-end with amortized variational inference using a bidirectional RNN inference network and HMM-based initialization to avoid collapse to one state.
Key results
- On curved and direct monkey reaching, SRNNs recover behaviorally relevant discrete states with lower error than SLDS, rSLDS, SNLDS and mrSDS across latent dimensionalities (P values), and K=5 chosen as the number of discrete states.
- On WFCI decision-making, SRNN prediction far outperforms linear switching models, and SRNN reconstruction is slightly better than other models.
- On WFCI lever pull, SRNN recovers the inhibitory dynamics ~3 s before behavior onset that SLDS and rSLDS miss, though its reconstruction is slightly worse than other methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SRNN (switching RNN) instead of SLDS, rSLDS, SNLDS, mrSDS for recovering behaviorally-relevant states | Curved reaching monkey dataset · Monkey curved reaching (electrophysiology) | error between behavioral states and recovered neural states (1-recall average) | Helped |
| Use SRNN instead of linear switching models (SLDS, rSLDS) for future-neural-activity prediction | Monkey reaching and mouse decision-making WFCI · Neural activity prediction forward in time | prediction MSE | Helped |
| Use SRNN instead of SLDS/rSLDS/SNLDS/mrSDS for reconstruction of neural activity | Monkey curved reaching · Neural reconstruction MSE | MSE | Mixed |
| Use SRNN instead of other models for reconstruction in WFCI lever-pull | Mouse self-initiated lever pull WFCI · Neural reconstruction MSE | MSE | Hurt |
| Increase number of discrete states K from 4 to 5 (vs 6) for reaching (ablation) | SRNN on curved reaching · Behavioral-state variability across conditions | variability of recovered states: 0.098 (K=5) vs 0.384 (K=4), 0.282 (K=6) | Helped |
| Initialize SRNN training using HMM states with penalty encouraging use of all discrete states | SRNN training · Training stability (qualitative) | whether model gets stuck in single state | Helped |
| Constrain SRNN to lever-pull region-specific activity (modular SRNN) | Mouse self-initiated lever pull WFCI · Region-to-region communication analysis | qualitative recurrent weight maps | No clear effect |
Surprising findings
- On the lever-pull dataset SRNN reconstruction is slightly worse than competing methods, despite better prediction and state recovery.
- Setting variance sigma of the inference network to a constant (0.0001) rather than learning it gave no apparent improvement.
Limitations
- Maximum number of discrete states K is set manually from behavioral observations and hyperparameter sweeps.
- Requires good initialization (HMM-based) to avoid collapse into a single discrete state.
- Training is substantially slower than SLDS and rSLDS.
Experimental setup
Datasets: Lorenz attractor simulated data Churchland et al. 2012 monkey reaching dataset Musall et al. 2019 mouse decision-making WFCI dataset Mitelut et al. 2022 mouse self-initiated lever pull WFCI dataset
Benchmarks: Churchland 2012 monkey reaching Musall 2019 WFCI decision-making Mitelut 2022 WFCI lever pull
Metrics: behavioral state recovery error (average 1-recall) reconstruction MSE forward prediction MSE co-smoothing MSE variability of recovered states
Baselines: SLDS rSLDS SNLDS mrSDS LFADS
Builds on: SLDS rSLDS SNLDS LFADS CURBD
Closest papers by meaning
- Modeling Latent Neural Dynamics with Gaussian Process Switching Linear Dynamical Systems (NeurIPS 2024 · similarity 0.91)
- Mechanistic Interpretability of RNNs emulating Hidden Markov Models (NeurIPS 2025 · similarity 0.91)
- Learning Dynamics of RNNs in Closed-Loop Environments (NeurIPS 2025 · similarity 0.90)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.89)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.89)
- Inferring stochastic low-rank recurrent neural networks from neural data (NeurIPS 2024 · 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.