Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior
| Authors | Eva Yezerets et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.05612 |
| Download | |
| Categories | cs.LG, stat.AP, stat.ML |
Abstract
Brain-wide recordings of large-scale networks of neurons now provide an unprecedented view into how the brain drives behavior. However, brain activity contains both information directly related to behavior as well as the potential for many internal computations. Moreover, observable behavior is executed not only by the brain, but also by the spinal cord and peripheral nervous system. Behavior is a coarse-grained product of neural activity, and we thus take the view that it can be best represented by lower-dimensional latent neural dynamics. Capturing this indirect relationship while disambiguating behavior-generating networks from internal computations running in parallel requires new modeling approaches that can embody the parallel and distributed nature of large-scale neural populations. We thus present behavior-decomposed linear dynamical systems (b-dLDS) to disentangle simultaneously recorded subsystems and identify how the latent neural subsystems relate to behavior. We demonstrate the ability of b-dLDS to decouple behavioral vs. internal computations on controlled, simulated data, showing improvements over a state-of-the-art model that uses behavior to supervise all dynamics based on behavior. We then show that b-dLDS can further scale up to tens of thousands of neurons by applying our model to large-scale recording of a zebrafish hindbrain during the complex positional homeostasis behavior, wherein b-dLDS highlights behavior-related dynamic connectivity networks.
Engineering Breakdown
Plain English
This paper addresses a fundamental challenge in computational neuroscience: how to extract behavior-driving neural dynamics from massive brain-wide recordings that contain both behavioral information and unrelated internal computations. The authors propose that behavior is best understood as a lower-dimensional latent projection of high-dimensional neural population activity, rather than a direct mapping. Their key insight is that the spinal cord and peripheral nervous system also contribute to executed behavior, so brain activity alone provides an incomplete view. The paper develops new modeling approaches designed to handle the parallel and distributed nature of large-scale neural populations, disambiguating which networks generate behavior versus which perform parallel internal processing.
Core Technical Contribution
The core technical novelty is a modeling framework that explicitly treats behavior as emerging from latent neural dynamics rather than direct neural-to-behavior mappings. Unlike prior approaches that assume a direct relationship between observed neural activity and behavior, this work introduces a dimensionality reduction strategy that captures the indirect, distributed nature of neural computation. The key architectural innovation is the ability to simultaneously model parallel neural computations—some behavior-relevant, some internal—within a single unified framework. This requires new inference and learning mechanisms that can disentangle which components of network activity drive behavior and which represent independent parallel processing, moving beyond classical linear decoding or end-to-end neural encoding models.
How It Works
The approach begins with high-dimensional brain-wide neural recordings as input, containing activity from millions of neurons recorded simultaneously. These recordings are processed through a latent dynamics model that projects the high-dimensional neural activity into a lower-dimensional latent space representing the true behavioral dynamics. The model simultaneously learns: (1) an encoder mapping observed neural activity to latent states, (2) latent dynamics equations governing state transitions, and (3) behavior decoders that extract task-relevant outputs from latent states. Critically, the architecture includes parallel pathways that allow some neural components to be decoded as behavior-driving latents while others are modeled as independent internal computations, without forcing all activity into a single behavioral channel. The model is trained to maximize both behavioral prediction accuracy and the interpretability of which neural populations correspond to which functional roles, effectively disentangling the mixed signals in raw neural recordings.
Production Impact
For engineers building brain-computer interfaces or neural recording analysis pipelines, this approach would enable more accurate and interpretable decoding of intended behavior from neural signals. Instead of training black-box decoders that map raw neural activity directly to behavior (which struggle with spurious correlations and internal noise), you could deploy latent variable models that explicitly separate behavior-driving dynamics from background neural processing. This improves robustness: if your recording setup introduces artifacts or captures neurons unrelated to behavior, the model's parallel pathway structure naturally isolates these without degrading behavioral predictions. The main production trade-offs are computational cost (fitting latent dynamics models is more expensive than linear decoders) and data requirements (you need longer, more diverse behavioral recordings to learn the latent structure reliably). Integration would require refactoring existing neural decoding pipelines to output uncertainty estimates over latent states rather than point predictions of behavior.
Limitations and When Not to Use This
The paper assumes that behavior can be meaningfully represented in a low-dimensional latent space, which may not hold for complex, multifaceted behaviors with many independent degrees of freedom. It also relies on the assumption that you have access to ground-truth behavior labels synchronized with neural recordings—in many real systems, behavior measurement is itself noisy, sparse, or only partially observable. The approach requires disambiguating behavior-driving vs. internal computations, but provides limited guidance on how to validate this disambiguation in the absence of ground truth about which neurons are truly behavior-relevant. Scalability to even larger neural populations or longer temporal sequences is unexplored, and the paper does not address how to handle non-stationary neural dynamics that may change across sessions or over time within a single session.
Research Context
This work builds on decades of neural dimensionality reduction research (from principal component analysis to modern variational autoencoders applied to neural data) while extending prior behavioral decoding methods that treated neural activity as a static feature matrix. It advances beyond classical approaches like linear discriminant analysis or Wiener filtering by explicitly modeling latent dynamics, similar to recent work on recurrent neural network models of neural populations (like those from Sussillo and colleagues). The paper contributes to the broader neuroscience shift toward treating the brain as a dynamical system rather than a static pattern classifier. This opens new research directions in understanding how parallel neural computations integrate, how to infer unobserved internal states from behavioral outputs, and how to design better brain-machine interfaces that account for the brain's intrinsic parallel processing architecture.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::