Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior

This paper introduces behavior-decomposed linear dynamical systems (b-dLDS), a novel modeling approach that disentangles behavior-related neural dynamics from internal computations in large-scale brain recordings, demonstrating superior performance over existing supervised models and successfully scaling to tens of thousands of neurons in zebrafish hindbrain data.

The Gist

Imagine your brain as a massive, bustling orchestra playing a symphony 24/7. Sometimes, the music is loud and obvious, like a trumpet blast that tells you, "The animal is swimming!" But most of the time, the orchestra is playing a complex, layered background score: some musicians are tuning their instruments, others are rehearsing a different song, and some are just chatting about the weather (internal thoughts, hunger, or planning).

For a long time, scientists trying to understand how the brain controls behavior have been like conductors trying to figure out which specific violinist is making the "swimming" sound. They often assumed that every musician in the orchestra was playing the same song at the same time, or that the sheet music (the brain's internal state) was directly tied to the audience's reaction (the behavior) at every single second.

The Problem:
The reality is messier. The brain is doing a million things at once. Some parts are driving the behavior (swimming), while other parts are doing internal math (deciding how hard to push against a current) or just maintaining the body's balance. If you try to listen to the whole orchestra to find the "swimming" note, you get lost in the noise.

The Solution: "Behavior-dLDS"
The authors of this paper introduced a new tool called Behavior-dLDS (behavior-decomposed Linear Dynamical Systems). Think of this tool as a super-smart audio mixer with a "Magic Filter."

Here is how it works, using simple analogies:

1. The "Dictionary of Moves"

Imagine the brain doesn't just have one way to move. Instead, it has a "dictionary" of about 15 or 20 basic "moves" or "patterns" (called Dynamics Operators).

• Move A: A pattern that looks like "swimming forward."
• Move B: A pattern that looks like "calibrating balance."
• Move C: A pattern that looks like "daydreaming."

At any given moment, the brain mixes these moves together, like a DJ mixing tracks. Sometimes it's 90% "swimming" and 10% "balance." Sometimes it's 100% "balance" because the fish is just sitting still.

2. The "Magic Filter" (The Breakthrough)

Previous tools tried to guess the mix by looking at the behavior first. They said, "The fish is swimming, so the brain must be doing Move A." This is like assuming that because you hear a trumpet, the whole orchestra must be playing a jazz song. It misses the fact that the bassist might be playing a totally different genre in the background.

Behavior-dLDS flips the script. It says:

"Let's look at the brain's 'moves' first. Then, let's ask: Which of these moves actually caused the swimming?"

It uses a special filter (called $\Psi$) that acts like a spotlight. It shines the spotlight only on the specific moves that are actually driving the behavior.

• The Spotlight: It finds that "Move A" is the one making the fish swim.
• The Dark Room: It realizes "Move B" and "Move C" are happening in the dark—they are real brain activity, but they have nothing to do with swimming right now. They are internal computations.

3. Why This Matters (The Zebrafish Example)

The researchers tested this on a tiny zebrafish with 13,000 neurons recorded at once. That's like trying to listen to 13,000 people talking in a stadium all at once.

• Old Methods: Tried to force all 13,000 neurons to explain the swimming. They got confused and crashed (literally ran out of computer memory).
• Behavior-dLDS: Successfully separated the noise. It found that while the fish was swimming, a specific group of neurons was doing the "swimming math," while another group was doing "balance math," and a third group was just "chilling."

It even discovered that the "swimming" neurons changed their connections over time, like a team changing their strategy mid-game.

The Big Picture Takeaway

Think of the brain not as a single machine that does one thing, but as a Swiss Army Knife with many tools.

• Old models tried to say, "The whole knife is cutting the steak."
• Behavior-dLDS says, "Ah, the blade is cutting the steak, but the screwdriver and the scissors are also open and doing their own jobs in the background."

This new model allows scientists to finally stop guessing and start seeing exactly which parts of the brain are driving a specific behavior, and which parts are just doing their own internal work. It's a huge step toward understanding how our complex brains manage to do so many things at once without falling apart.

Technical Summary

Here is a detailed technical summary of the paper "Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior."

1. Problem Statement

Modern neuroscience faces a critical challenge in interpreting large-scale, brain-wide neural recordings. While these recordings capture vast amounts of data, the relationship between neural activity and observable behavior is complex, indirect, and often non-stationary.

• The Core Issue: Neural activity contains both information directly driving behavior and numerous parallel internal computations (e.g., decision-making, homeostasis, sensory processing) that are not immediately reflected in behavior.
• Limitations of Existing Models:
  • Unsupervised Models (e.g., dLDS): Can identify latent dynamics but do not explicitly link them to behavior, making it difficult to distinguish behavior-generating networks from internal noise.
  • Fully Supervised Models (e.g., CLDS, DCM): Condition dynamics entirely on external variables (behavior). This forces the model to assume all neural dynamics are driven by behavior, failing to capture independent internal computations.
  • Shared Latent Space Models (e.g., PSID, CEBRA): Assume a single latent state generates both neural activity and behavior simultaneously. This fails when behavior is a coarse-grained product of only a subset of the latent dynamics or when the relationship is hierarchical rather than direct.

The authors propose a need for a "middle ground" model that can disentangle behavior-related dynamics from internal computations while scaling to tens of thousands of neurons.

2. Methodology: Behavior-dLDS (b-dLDS)

The authors introduce behavior-dLDS, a model that decomposes neural dynamics into a linear combination of basis operators (Dynamics Operators or DOs) but explicitly constrains only a subset of these dynamics to behavior.

Model Architecture

The model operates on three main components:

1. Latent Neural Dynamics: Observed neural data $y_t$ is modeled as a linear projection of a lower-dimensional latent state $x_t$:
   $$y_t = D x_t + \epsilon_y$$
   The latent state evolves via a time-varying linear dynamical system:
   $$x_t = F_t x_{t-1} + \epsilon_x$$
   Crucially, the transition matrix $F_t$ is decomposed into a sum of fixed basis operators ($f_m$) weighted by time-varying dynamics coefficients ($c_{mt}$):
   $$F_t = \sum_{m=1}^{M} f_m c_{mt}$$

2. Behavioral Generation: Instead of assuming the latent state $x_t$ directly generates behavior, b-dLDS posits that the dynamics coefficients $c_t$ generate behavior. This acknowledges that behavior is a coarse-grained output of specific dynamic processes.
   $$b_t = \Psi c_t + \epsilon_b$$
   Here, $\Psi$ is a mapping matrix. The model assumes $\Psi$ is sparse, meaning only a subset of the dynamics coefficients (those in set $\Gamma$) actually drive the observed behavior $b_t$, while others drive internal processes.

3. Inference and Regularization:

   • Objective: The model uses an Expectation-Maximization (EM) algorithm.
   • E-Step: Infers latent states $x_t$ and coefficients $c_t$ using a dynamic filtering approach (BPDN-DF) that minimizes a cost function including reconstruction error, sparsity on $x_t$ and $c_t$, and a behavior reconstruction term ($\|b_t - \Psi c_t\|^2$).
   • M-Step: Updates parameters ($D, \{f_m\}, \Psi$) via stochastic projected gradient descent.
   • Key Innovation: A Frobenius norm regularization on $\Psi$ is introduced. This biases the model to learn that the mapping from coefficients to behavior is sparse, effectively forcing the model to identify which specific DOs are behavior-relevant and which are not.

3. Key Contributions

• Partial Constraint Framework: Unlike previous models that either ignore behavior or condition all dynamics on it, b-dLDS explicitly models a sparse mapping between latent dynamics and behavior. This allows the simultaneous identification of behavior-generating subsystems and independent internal computation subsystems.
• Scalability: The model is designed to handle large-scale data (tens of thousands of neurons) by avoiding the calculation of high-dimensional covariance matrices, leveraging efficient sparse-state filtering algorithms.
• Hierarchical Interpretation: It reframes the "brain state" not as an instantaneous firing rate, but as the configuration of active Dynamics Operators over time, providing a more interpretable link between neural circuitry and behavior.

4. Results

The authors validated b-dLDS through simulations and real-world biological data.

A. Simulation Studies

• Disentanglement: In simulations with two independent neural systems (one driving behavior, one internal), b-dLDS successfully identified the specific dynamics coefficients driving behavior while ignoring the independent system.
• Comparison with CLDS: When compared to Conditionally Linear Dynamical Systems (CLDS), b-dLDS demonstrated superior accuracy. CLDS, being fully conditioned on behavior, failed to recover the intrinsic dynamics of the non-behavioral system, whereas b-dLDS correctly separated them.
• Comparison with PSID: b-dLDS outperformed Preferential Subspace Identification (PSID) in scenarios where neural activity and behavior had different temporal structures (e.g., oscillatory vs. smooth states), as PSID's shared latent space assumption failed to capture the hierarchy.

B. Zebrafish Hindbrain Application

• Dataset: Applied to a large-scale calcium imaging dataset (~13,000 neurons) from a zebrafish hindbrain during positional homeostasis (swimming in a current).
• Findings:
  • b-dLDS identified specific Dynamics Operators (DOs) whose coefficients were strongly correlated with motor output (swimming).
  • It simultaneously identified other DOs that were active but uncorrelated with motor output, likely representing internal computations (e.g., self-location encoding).
  • Connectivity Maps: The model generated dynamic connectivity maps showing that behavior-related DOs (e.g., DO 8) exhibited global hindbrain connectivity correlated with movement, while non-behavioral DOs showed more localized or distinct temporal patterns.
  • Validation: The identified behavior-related networks aligned with known biological structures (SLO-MO neurons in the medulla oblongata) previously characterized in the literature.

5. Significance

• Conceptual Shift: b-dLDS offers a new paradigm for analyzing brain-behavior relationships, moving away from "all-or-nothing" assumptions toward a modular view where specific neural subsystems drive behavior while others run parallel internal processes.
• Scalability for Big Data: The ability to process 13,000+ neurons makes b-dLDS one of the few models capable of analyzing whole-brain imaging data without dimensionality reduction that loses critical information.
• Future Applications: The framework enables researchers to dissect complex behaviors by isolating the specific neural circuits responsible for execution versus those responsible for internal state maintenance, decision-making, or sensory integration. The modular nature of the model also suggests potential for extending it to hierarchical models involving sensory inputs and higher-order cognitive variables.