🧠 neuroscience

One Circuit, Many Flow Fields: Mechanistic Models of Single-Trial Neural Dynamics

This paper proposes a low-rank recurrent network model that reframes single-trial neural variability as continuous deformations of a shared circuit's flow field driven by trial-specific control parameters, successfully recovering ground-truth dynamics and revealing interpretable behavioral engagement and task structure in motor cortex data.

Original authors: Kaminitz, S., Levin, M., Pereira-Obilinovic, U., Darshan, R.

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kaminitz, S., Levin, M., Pereira-Obilinovic, U., Darshan, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The brain is often imagined as a vast, static machine, where the same electrical circuits fire in the same predictable patterns to produce the same movements. For decades, scientists have described neural activity using the language of dynamical systems, viewing the brain as a landscape of hills and valleys that guide thoughts and actions along specific paths. In this traditional view, the shape of the landscape is fixed; if a person reaches for a cup, their brain activity follows a set route, and any differences from one attempt to the next are simply random noise, like static on a radio line. However, recent observations have challenged this picture. When animals perform the same task repeatedly, their brain activity does not just jitter around a single path; it shifts in organized, meaningful ways. These shifts often correlate with the animal's internal state, such as how alert or engaged it is, suggesting that the brain's very rules for processing information might be flexible, changing from moment to moment to suit the situation.

A new study from researchers at Tel Aviv University and the Allen Institute proposes a radical shift in how we understand these fluctuations. Instead of treating trial-to-trial differences as mere errors to be averaged out, the team suggests that a single neural circuit can generate a continuous family of different operating modes. They argue that the brain does not just wander around a fixed map; it actually redraws the map itself. To test this, the researchers developed a mathematical model that treats each moment of brain activity as a unique configuration of a shared circuit. They found that by inferring a few hidden control parameters for each attempt, they could explain how the brain's internal landscape deforms, shifting from one stable state to another without changing the underlying wiring.

The researchers tested this idea first in a controlled simulation, creating a synthetic brain circuit with known rules and then asking their model to figure out those rules just by watching the activity. The model succeeded in recovering the hidden structure, identifying exactly how small changes in a control signal could turn a system with two stable states into one with only a single state. This proved that the method could distinguish between random noise and genuine, structured changes in how the circuit operates. The team then applied the same approach to real data recorded from the brains of mice performing a delayed movement task. In these experiments, mice had to wait for a signal before licking a spout to get a reward. The researchers noticed that in some trials, the mice simply ignored the signal and did not move. By analyzing the brain activity, the model identified a specific direction in the control space that corresponded to this disengagement. When the researchers mathematically shifted the activity of an engaged mouse along this direction, the model's simulated brain activity transformed to look exactly like the disengaged state, effectively switching the circuit from "working" to "ignoring" without altering the physical connections.

The study also looked at recordings from the motor cortex of monkeys performing reaching tasks. Here, the goal was to see if the model could find control directions related to the specific task at hand, rather than just internal states like attention. The researchers found that the model organized the different directions the monkey could reach into a smooth, two-dimensional plane within its control space. Moving a trial's activity along this plane in the simulation caused the decoded movement to rotate predictably, matching the new target direction. This demonstrated that the model had uncovered a continuous control mechanism where the brain smoothly adjusts its dynamics to steer movement, rather than switching between a few discrete options. Furthermore, the researchers showed that the statistical patterns of these control parameters learned from one monkey could be transferred to another monkey performing the same task. This suggests that the way the brain reshapes its dynamics is not a random quirk of a single animal but a shared, fundamental property of how these circuits operate.

The significance of this work lies in its ability to turn a methodological problem into a scientific discovery. By reframing trial-to-trial variability as a movement through a low-dimensional control space, the researchers provide a way to read the brain's internal settings directly from its activity. The model is not just a descriptive tool; it is mechanistic, meaning it identifies specific input patterns that could be tested in a living animal. The researchers suggest that by using targeted light stimulation to mimic these inferred control signals, scientists could experimentally push a brain circuit from one dynamical regime to another, such as forcing a disengaged animal to re-engage or changing the direction of a reach. This approach moves beyond simply describing what the brain does to explaining how it flexibly reuses the same hardware to produce a wide range of behaviors, offering a new lens through which to view the dynamic, adaptable nature of the mind.

Technical Summary: One Circuit, Many Flow Fields: Mechanistic Models of Single-Trial Neural Dynamics

Problem Statement
Neural circuits are traditionally modeled as dynamical systems where population activity evolves through a fixed flow field, with trial-to-trial variability treated as noise (stochastic fluctuations, initial condition differences) around this fixed landscape. However, large-scale recordings reveal that variability is often structured, low-dimensional, and linked to slowly varying internal states (e.g., arousal, engagement, strategy). Standard models either treat this as noise around a fixed system or as discrete switches between a small number of regimes. Neither approach captures how continuous internal-state variables can gradually deform the circuit's vector field. The central challenge is to move beyond fitting single-trial activity as noise and instead infer the low-dimensional control parameters that reshape the underlying dynamical system across trials.

Methodology
The authors propose a framework that reframes single-trial fitting as the inference of low-dimensional control parameters that deform a shared recurrent circuit's flow field.

Model Architecture: The core is a low-rank recurrent neural network (RNN) where connectivity $J$ is shared across all trials and constrained to be low-rank ( $J = UAU^\top$ ). Each recorded neuron corresponds to a unit in the network.
Trial-Specific Control: Trial-to-trial variability is modeled via a static, low-dimensional input bias vector $b_k$ for each trial $k$ . This bias enters the dynamics as a constant input $I_k = Ub_k$ within the trial.
Mechanism: Crucially, $b_k$ is constant over time within a trial, acting not as a time-varying driver but as a bifurcation parameter. It deforms the flow field of the shared circuit, shifting fixed points, altering stability, and inducing bifurcations without directly driving the temporal evolution of activity.
Fitting Procedure: The model is fit to recorded single-trial activity by jointly optimizing shared parameters (connectivity $U, A$ , neuron maxima $r_{max}$ , and task inputs $I(t)$ ) and the per-trial biases $\{b_k\}$ .
Generative Extension: A generative version treats biases as random variables sampled from a learned distribution $b_k \sim \mathcal{N}(0, \Sigma)$ , allowing the model to capture the distribution of trial-to-trial variability and test transferability across sessions and subjects.

Key Contributions

Control-Space Model: Introduction of a low-rank RNN where continuous, trial-specific inputs deform the flow field of a shared circuit, with the ability to infer this control space directly from activity.
Ground-Truth Validation: Demonstration on synthetic data that the method recovers trial-dependent flow fields and their bifurcation structures (control spaces) from activity alone.
Internal-State Control (Mouse): Application to mouse anterior lateral motor cortex (ALM) data, identifying a control axis for "disengagement" that systematically reorganizes dynamics between engaged and disengaged regimes.
Task-Variable Control (Monkey): Application to monkey motor cortex data, identifying a "reach-direction control plane" that generalizes to held-out directions; moving within this plane predictably rotates the decoded reach.
Shared Dynamics: A generative extension showing that the latent dynamics and variability structure transfer across different subjects performing the same task.

Results

Synthetic Validation: In a teacher-student setting with known bistable and limit-cycle dynamics, the model successfully recovered the ground-truth vector fields, fixed-point locations, and bifurcation structures. The pairwise distances in the inferred bias space strongly correlated with distances in the vector-field space ( $r=0.99$ ), confirming the model captures the geometry of variability.
Mouse Motor Cortex (ALM):
- The model achieved high validation performance ( $R^2 \approx 0.64$ ) in reproducing single-trial activity.
- It identified a specific "disengagement axis" in the bias space. Trials with higher projections onto this axis corresponded to "ignore" trials (no response).
- Causal Manipulation: Shifting the inferred biases of "hit" trials along the disengagement axis systematically deformed the latent dynamics toward the "ignore" regime, reducing choice and response projections. Conversely, shifting "ignore" trials in the opposite direction moved them toward the engaged regime.
Monkey Motor Cortex:
- Without explicit task inputs, the model inferred a two-dimensional "reach-direction plane" in the bias space.
- Biases projected onto this plane formed a ring ordered by target angle.
- Generalization & Control: The plane generalized to held-out target angles. Shifting a trial's bias within this plane by a specific angle $\Delta$ resulted in a decoded reach rotation of approximately $\Delta$ , demonstrating that the inferred control space directly maps to the dynamical steering of the circuit.
Cross-Subject Transfer: When latent dynamics ( $A$ ) and variability distributions ( $\Sigma$ ) learned from one mouse session were transferred to another, the model maintained performance close to a baseline trained directly on the target session, significantly outperforming random controls. This suggests a shared low-dimensional control space across animals.

Significance and Claims
The paper claims to reframe a methodological problem (fitting single-trial variability) as a scientific opportunity: reading off the control parameters of the underlying dynamics. By modeling trial-to-trial variability as movement through a low-dimensional control space that deforms a shared flow field, the approach connects data-driven inference to mechanistic theories of flexible behavior.

The authors emphasize that their model is mechanistic rather than merely descriptive. Because the inferred biases correspond to specific input patterns onto identified neurons, the model predicts how to experimentally steer the circuit (e.g., via holographic photostimulation) to move it between dynamical regimes. The framework suggests that a single neural circuit can generate a continuum of related dynamical systems ("one circuit, many flow fields") by reusing its dynamics through continuous modulation of internal states or task demands, rather than switching between discrete regimes or relying solely on noise.

Limitations
The authors note that the framework currently treats trial-specific biases as independent across trials. It does not yet account for scenarios where biases on a given trial depend nonlinearly on the history of preceding trials, which would require additional dynamics for the biases themselves.

Technical Summary: One Circuit, Many Flow Fields: Mechanistic Models of Single-Trial Neural Dynamics

More like this