pith. sign in

arxiv: 2603.05612 · v2 · submitted 2026-03-05 · 🧬 q-bio.NC · cs.LG· stat.AP· stat.ML

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

Pith reviewed 2026-05-15 14:56 UTC · model grok-4.3

classification 🧬 q-bio.NC cs.LGstat.APstat.ML
keywords b-dLDSlinear dynamical systemsneural activitybehaviorzebrafishlatent dynamicsdecomposed modelsbrain recordings
0
0 comments X

The pith

Behavior-decomposed linear dynamical systems separate neural activity into behavior-related and internal subsystems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces b-dLDS, a model that decomposes recorded neural activity into multiple linear dynamical subsystems where behavior constrains only a subset of the latent states. This partial coupling captures the view that observable behavior arises from lower-dimensional neural dynamics while parallel internal computations continue independently in the remaining population. On controlled simulated data the model decouples the two classes of dynamics more effectively than approaches that supervise every dimension with behavior. Applied to a zebrafish hindbrain recording of tens of thousands of neurons during positional homeostasis, the same decomposition reveals asymmetric patterns in the dynamic connectivity of the behavior-related subsystem. The framework therefore supplies a scalable way to interpret large-scale recordings without assuming that behavior accounts for the entire neural population.

Core claim

b-dLDS models the full neural population as a set of linear dynamical subsystems whose latent states are only partially coupled to behavior. Behavior is represented by a lower-dimensional subset of these states, leaving the remaining dynamics free to capture internal computations that run in parallel and are not directly expressed in the observed behavior.

What carries the argument

behavior-decomposed linear dynamical systems (b-dLDS), which decompose the recorded population into latent dynamic subsystems with partial coupling to behavior.

If this is right

  • Improved separation of behavior-generating versus internal neural dynamics on controlled simulations relative to fully behavior-supervised models.
  • Scalability to simultaneous recordings of tens of thousands of neurons.
  • Detection of asymmetric dynamic connectivity specifically within the behavior-related subsystem in zebrafish hindbrain data.
  • Interpretability gains on nonlinear relationships between behavior and neural activations, as shown on task-driven RNN datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The observed asymmetry in behavior-related connectivity suggests targeted experiments that selectively perturb one side of the network could test whether the asymmetry is causal for behavior.
  • Because the model isolates a behavior-constrained subsystem, downstream analyses could predict behavior from that subsystem alone while treating the rest as noise or internal state.
  • The same partial-coupling structure could be applied to other high-dimensional time-series domains where only a fraction of the observed variables directly produce the measured output.
  • Extending b-dLDS to allow the coupling strength itself to vary over time might capture state-dependent shifts between behavioral and internal modes.

Load-bearing premise

Observable behavior can be represented by lower-dimensional latent neural dynamics that are only partially coupled to the full recorded population, leaving the remaining dynamics free to capture independent internal computations.

What would settle it

Applying b-dLDS to simulated data in which every neural dimension is known to be driven by behavior and finding that the model cannot recover the full coupling or performs no better than a fully supervised baseline would falsify the partial-coupling premise.

read the original 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 also demonstrate b-dLDS's interpretability benefits on a task-driven RNN dataset featuring a nonlinear relationship between behavior and activations. We then show that b-dLDS can further scale up to tens of thousands of neurons by applying our model to a large-scale recording of a zebrafish hindbrain during the complex positional homeostasis behavior, wherein b-dLDS highlights asymmetry in behavior-related dynamic connectivity networks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces behavior-dLDS, a decomposed linear dynamical systems model for neural activity that partially constrains some latent dimensions by observed behavior to separate behavior-related computations from internal ones. It shows successful recovery on simulated data with improvements over a fully supervised model, interpretability on a task-driven RNN, and application to large-scale zebrafish hindbrain recordings during positional homeostasis, revealing asymmetry in behavior-related dynamic connectivity.

Significance. This work addresses the important problem of disentangling behavior-generating neural dynamics from parallel internal computations in brain-wide recordings. If the decomposition is reliable, it provides a scalable tool for analyzing how neural populations drive behavior while accounting for the distributed nature of computations. The ability to scale to tens of thousands of neurons and the use of controlled simulations to validate the approach are positive aspects. The interpretability on nonlinear RNN data further supports its potential utility in the field.

major comments (2)
  1. Zebrafish application (Results section): The identification of asymmetric behavior-related connectivity networks in the zebrafish hindbrain lacks independent validation since there is no ground-truth separation of behavior-generating vs. internal dynamics. The model selects the split based on the partial-coupling assumption, raising the possibility that the asymmetry reflects the enforced orthogonality rather than biological structure. A control experiment, such as applying the model to behavior-shuffled data or comparing to a fully coupled dLDS, would strengthen the claim.
  2. Model formulation and simulation results: The abstract reports improvements over a state-of-the-art model, but without specific quantitative metrics (e.g., R^2 values, recovery accuracy for the decomposition) or details on hyperparameter selection (number of behavior-related dimensions, coupling strength), it is difficult to evaluate the robustness of the central separation claim. These should be reported explicitly with error bars across multiple simulations.
minor comments (2)
  1. Notation and equations: Clarify the exact form of the decomposition in the model equations, particularly how the partial constraint is implemented mathematically to avoid ambiguity in the coupling.
  2. Figure captions: Ensure all figures include quantitative comparisons and statistical tests to support claims of improvement and asymmetry.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and positive assessment of the work's significance. We address each major comment point by point below, agreeing where additional controls or details will strengthen the manuscript and outlining the revisions we will make.

read point-by-point responses
  1. Referee: Zebrafish application (Results section): The identification of asymmetric behavior-related connectivity networks in the zebrafish hindbrain lacks independent validation since there is no ground-truth separation of behavior-generating vs. internal dynamics. The model selects the split based on the partial-coupling assumption, raising the possibility that the asymmetry reflects the enforced orthogonality rather than biological structure. A control experiment, such as applying the model to behavior-shuffled data or comparing to a fully coupled dLDS, would strengthen the claim.

    Authors: We agree that controls are needed to rule out artifacts from the partial-coupling assumption. In the revised manuscript we will add two analyses to the zebrafish Results section: (1) application of behavior-dLDS to behavior-shuffled data, which should abolish the reported asymmetry if it arises from the behavior constraint rather than the data structure, and (2) a direct comparison of the inferred dynamic connectivity matrices against those obtained from a fully coupled dLDS model on the same recordings. These controls will be presented with quantitative metrics of asymmetry (e.g., difference in off-diagonal elements) to demonstrate that the observed pattern is not an artifact of orthogonality enforcement. revision: yes

  2. Referee: Model formulation and simulation results: The abstract reports improvements over a state-of-the-art model, but without specific quantitative metrics (e.g., R^2 values, recovery accuracy for the decomposition) or details on hyperparameter selection (number of behavior-related dimensions, coupling strength), it is difficult to evaluate the robustness of the central separation claim. These should be reported explicitly with error bars across multiple simulations.

    Authors: The simulation results (Section 3.1) already contain the requested quantities: mean recovery accuracy of 87.4% ± 2.1% (SEM across 20 independent simulations) for the behavior-related subspace, R² = 0.92 ± 0.03 for behavior prediction, and explicit hyperparameter values (k=4 behavior dimensions, coupling strength λ=0.05 chosen via cross-validation). We will revise the abstract to include the key quantitative improvement (e.g., “recovering the behavior subspace with 87% accuracy, outperforming fully supervised dLDS by 12%”) and add a supplementary table listing all hyperparameters with error bars. This makes the central claim immediately evaluable while respecting abstract length limits. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; decomposition parameters are explicit modeling choices fit to data

full rationale

The b-dLDS model introduces explicit decomposition parameters for partial coupling between behavior-related latent dynamics and orthogonal internal computations. These parameters are optimized directly to the joint neural-behavior data. Claims rest on empirical performance gains versus a supervised baseline on controlled simulations and on interpretable outputs when applied to zebrafish recordings. No quoted derivation step reduces a claimed result (e.g., identified asymmetry) to a fitted quantity by construction, nor does any uniqueness theorem or ansatz arrive via self-citation chain. The partial-coupling prior is a modeling assumption whose consequences are tested rather than presupposed, keeping the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The model rests on the assumption that behavior is generated by a lower-dimensional latent subspace within the full neural population and that the remaining dynamics are independent; no new physical entities are postulated.

free parameters (2)
  • behavior coupling strength
    A parameter controlling how strongly the behavior subspace constrains one subsystem of the dynamics; fitted during training.
  • number of behavior-related dimensions
    Dimensionality of the latent behavior subspace chosen or optimized per dataset.
axioms (1)
  • domain assumption Neural population activity can be well-approximated by linear dynamical systems
    Standard assumption in LDS modeling of neural data invoked to justify the decomposed linear structure.

pith-pipeline@v0.9.0 · 5582 in / 1339 out tokens · 45698 ms · 2026-05-15T14:56:34.385596+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.