| name | behavior-decomposed-lds |
| description | Behavior-dLDS: decomposed linear dynamical systems model for neural activity partially constrained by behavior. Disentangles behavior-related neural dynamics from internal computations in large-scale neural recordings. Scales to tens of thousands of neurons. Use when modeling neural population dynamics, decomposing brain activity into behavioral vs. internal subsystems, or analyzing brain-wide recordings with behavioral correlates. Activation: behavior-dLDS, decomposed linear dynamical systems, neural dynamics decomposition, behavior-constrained neural modeling, brain-wide recordings, latent neural dynamics, zebrafish neural dynamics, positional homeostasis
|
Behavior-dLDS: Decomposed Linear Dynamical Systems
Based on arXiv:2603.05612 (Yezerets et al., May 2026).
Paper Overview
Title: Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior
Authors: Eva Yezerets, En Yang, Misha B. Ahrens, Adam S. Charles
Published: Submitted Mar 2026, revised May 4, 2026 (v2)
Categories: q-bio.NC, cs.LG, stat.AP, stat.ML
Core Problem
Brain-wide recordings contain both behavior-related information and internal computations.
Observable behavior is a coarse-grained product of neural activity executed by brain + spinal cord + PNS.
Existing models use behavior to supervise all dynamics, conflating behavioral and internal processes.
Need: disentangle behavior-generating networks from parallel internal computations.
Key Contributions
1. Behavior-Decomposed Linear Dynamical Systems (b-dLDS)
- Decomposes neural activity into behavior-related and behavior-independent subsystems
- Models indirect relationship between behavior and neural dynamics
- Embodies parallel and distributed nature of large-scale neural populations
- Represents behavior via lower-dimensional latent neural dynamics
2. Subsystem Disentanglement
- Identifies how latent neural subsystems relate to behavior
- Separates behavior-generating networks from internal computations
- Outperforms models that use behavior to supervise all dynamics
3. Scalability
- Demonstrated on simulated data with controlled ground truth
- Applied to task-driven RNN dataset with nonlinear behavior-activation relationships
- Scaled to tens of thousands of neurons on zebrafish hindbrain recordings
- Revealed asymmetry in behavior-related dynamic connectivity networks
Mathematical Framework
Neural State x(t) decomposed into:
├── x_behavior(t): behavior-related latent dynamics
│ └── Directly constrained by observed behavior
├── x_internal(t): behavior-independent latent dynamics
│ └── Captures parallel internal computations
└── Coupling: A_behavior→internal, A_internal→behavior
Linear dynamics:
x(t+1) = A * x(t) + noise
where A is block-structured to enable decomposition
Validation Results
| Dataset | Key Finding |
|---|
| Simulated data | b-dLDS outperforms behavior-supervised baselines |
| Task-driven RNN | Interpretability benefits on nonlinear behavior relationships |
| Zebrafish hindbrain | Asymmetry in behavior-related dynamic connectivity |
Application Domains
- Brain-wide calcium imaging analysis
- Electrophysiology population recordings
- Neural decoding with behavioral correlates
- Distinguishing motor vs. cognitive neural activity
- Studying internal states not directly observable in behavior
Key Insights
- Behavior is a coarse-grained proxy for neural dynamics
- Internal computations run in parallel with behavior generation
- Lower-dimensional latent dynamics best represent behavioral signals
- Decomposition reveals asymmetry in neural connectivity patterns
- Model scales to tens of thousands of recorded neurons
Activation Keywords
- behavior-dLDS
- decomposed linear dynamical systems
- neural dynamics decomposition
- behavior-constrained modeling
- brain-wide recordings
- latent neural dynamics
- zebrafish neural analysis
- positional homeostasis
- neural subsystem disentanglement
- internal vs behavioral computation
Related Skills
- neural-population-dynamics
- neural-dynamics-universal-translator
- brain-digital-twins-execution-semantics
- heteroclinic-cognitive-state-modeling
