🧠 neuroscience

Feasible closed-loop stimulation for neural latent dynamics control

Through high-throughput simulations of biophysically grounded spiking neural networks, this study demonstrates that feasible closed-loop control of neural latent dynamics is achievable with low-dimensional actuation (~20% of channels) and balanced allocation of resources between stimulation and recording, providing critical design principles for future in vivo experiments.

Original authors: Johnsen, K. A., Lai, Y., Vivekanandasarma, L., Berman, G. J., Pandarinath, C., Rozell, C. J.

Published 2026-11-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Johnsen, K. A., Lai, Y., Vivekanandasarma, L., Berman, G. J., Pandarinath, C., Rozell, C. J.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 human brain is a vast, noisy network of billions of neurons firing in complex patterns. For decades, scientists have tried to understand how these chaotic signals translate into smooth, purposeful actions like reaching for a cup or speaking a word. A powerful idea in modern neuroscience suggests that beneath the surface noise, the brain operates on a simpler, hidden layer of activity. Think of this hidden layer as a set of invisible levers or "factors" that drive behavior. If researchers could identify these levers and learn to pull them, they might be able to understand exactly how the brain causes specific actions, rather than just watching the activity happen. However, a major hurdle remains: while scientists have become very good at recording these hidden patterns, they have struggled to manipulate them directly. Traditional methods often involve poking the brain with broad, untargeted signals or trying to control individual neurons one by one, a task that is incredibly difficult and often misses the bigger picture. The question has been whether it is possible to steer these hidden brain states in real time to guide behavior, and if so, what kind of tools would be needed to do it.

A team of researchers at Georgia Tech and Emory University set out to answer this question by building a detailed virtual laboratory. Instead of working with live animals immediately, which is expensive and technically demanding, they created a sophisticated computer simulation of a brain region involved in movement. They trained this digital network to mimic the activity of a monkey reaching for targets, allowing the computer to learn the hidden "factors" that drive the movement. Once the simulation was running, the researchers tested a new approach called closed-loop control. In this setup, the computer constantly watches the brain's hidden state, predicts where it is going, and instantly sends light signals to nudge it back on course if it starts to drift. This happens in a continuous loop, with the system reacting to changes as they occur, rather than just sending a pre-planned signal. The goal was to see if this method could successfully steer the brain's hidden dynamics to produce specific movements, and to figure out exactly how many light sources and recording sensors would be needed to make it work.

The researchers first tackled a fundamental question about precision: does controlling the brain require manipulating every single neuron, or can a smaller number of light sources do the job? In their simulations, they varied the number of independent light channels from just one, which would illuminate the whole area uniformly, up to over a thousand, where each neuron could be controlled individually. They found that the brain did not need to be micromanaged. Even with a relatively small number of light sources—about 200 out of the 1,024 available channels—the system could achieve roughly 85% of the maximum possible control performance. This suggests that the idea of needing to control every single neuron to influence behavior is likely unnecessary. The brain's hidden levers are robust enough that a moderate amount of targeted stimulation is sufficient to guide them, challenging the current trend of investing heavily in ultra-high-resolution tools that control neurons one by one.

The second part of the study addressed a practical engineering problem: how should a limited number of channels be split between recording what the brain is doing and stimulating it to make it do something? In many current experiments, scientists use many recording electrodes to listen to the brain but only a single light source to stimulate it, assuming that listening is more important than speaking. The researchers tested this assumption by running simulations with a fixed total of 100 channels, splitting them in different ways between recording and stimulation. They discovered that the most effective strategy was to split the resources almost evenly. When the system had roughly equal numbers of recording sensors and light sources, it performed the best at guiding the simulated arm to the correct target. The common practice of using many sensors and only one light source performed poorly, offering very little improvement over doing nothing at all. This finding suggests that to truly control brain activity, researchers need to invest just as much in the ability to send signals as they do in the ability to listen.

Throughout these simulations, the researchers tested their system under many different conditions, changing the density of the neurons, the width of the light beams, and the amount of data used to train the control system. The results remained consistent: low-dimensional stimulation works, and a balanced mix of sensing and acting is optimal. Even with standard control methods and without extreme optimization, the system reduced tracking errors by more than 60% and successfully guided the simulated movement in over 90% of trials. The study also highlighted the importance of real-time feedback. When the researchers tested a version of the system that planned the entire movement in advance without adjusting to changes as they happened, it performed significantly worse, often failing to guide the movement correctly. This confirms that the ability to react instantly to the brain's changing state is crucial for success.

While these results come from computer simulations and not live animals, they provide a clear roadmap for future experiments. The findings suggest that building a device to control brain dynamics does not require impossibly complex technology that targets every single cell. Instead, a system with a moderate number of independent light sources, balanced with an equal number of recording sensors, could be enough to steer the brain's hidden states and test causal theories about how behavior is generated. If these simulations hold true in living tissue, it could open the door to a new era of neuroscience where researchers do not just observe the brain's activity but actively engage with it to understand the fundamental rules of how we move and think. The path forward involves building these tools and testing them in the real world, but the simulations offer a strong indication that the goal is within reach.

Technical Summary: Feasible Closed-Loop Stimulation for Neural Latent Dynamics Control

Problem Statement
Modern neuroscience utilizes latent dynamical systems models to explain high-dimensional neural population activity with low-dimensional factors that predict behavior. However, these inferred latent factors are rarely targeted explicitly via experimental perturbations, leaving causal relationships between latent dynamics and behavior untested. While closed-loop optogenetic control (CLOC) offers a solution to steer neural trajectories in real-time, its feasibility for controlling latent factors remains unproven. Critical technical unknowns hinder experimental design:

Stimulation Dimensionality: Is fine-grained, single-neuron resolution necessary to navigate low-dimensional latent manifolds, or is low-dimensional actuation sufficient?
Resource Allocation: Given fixed constraints on implant volume and bandwidth, how should a limited channel budget be optimally allocated between recording (sensing) and stimulation (actuation) to maximize behavioral control?

Addressing these questions via in vivo experiments is prohibitively expensive and technically challenging due to the need for large parameter sweeps.

Methodology
The authors conducted a high-throughput in silico survey using the Cleo simulator, which integrates biophysically grounded spiking neural networks with models of multi-electrode recording, optogenetic stimulation, and closed-loop control.

Network Model: The study employed factor-based spiking neural networks (fSNNs). These networks use low-rank recurrent connectivity trained to reproduce low-dimensional latent factors (derived from a rate-based RNN) and behavioral outputs (approximating macaque electromyography during a center-out reaching task). The networks lacked distinct cell types or laminar structures to isolate the role of spatial actuation granularity.
Experimental Pipelines: Two primary experimental paradigms were simulated:
1. Stimulation Dimensionality: A fixed grid of 1,024 light sources was grouped into $k$ independently controlled channels ( $k \in [1, 1024]$ ), ranging from uniform illumination to individual source control. The controller used ground-truth latent factors to track reference trajectories.
2. Stimulation–Recording Balance: A fixed budget of 100 channels was partitioned between $k$ stimulation sources and $m$ recording electrodes ( $k+m=100$ ). In this setup, the controller had to infer latent states from electrophysiological recordings (MUA and/or LFP) using a Kalman filter to control behavioral outputs.
Control Framework: The system identification phase involved random photostimulation to fit Linear Dynamical System (LDS) models. Control experiments utilized Model Predictive Control (MPC) to minimize tracking error and input power over a receding horizon. Open-loop (OL) control was also tested for comparison.

Key Contributions and Results

Low-Dimensional Actuation is Sufficient:
- The study found that fine-grained, single-neuron stimulation is unnecessary for steering low-dimensional latent dynamics.
- In the reference configuration, approximately 20% of the stimulation channels (specifically, a "knee point" at $k_{knee} = 200$ channels for a 1,024-neuron network) achieved ~85% of the peak control performance.
- This diminishing returns profile was robust across variations in neuronal density, beam width, and training data, though $k_{knee}$ was sensitive to the number of latent factors and the amount of system identification data.
Optimal Resource Allocation is Balanced:
- For a fixed 100-channel budget, optimal behavioral control was achieved when stimulation and recording channels were allocated roughly equally ( $\alpha^*_{stim} \approx 0.525$ ).
- This contrasts sharply with prevailing experimental practices that pair many recording electrodes with a single optical fiber (low-actuation regimes), which performed poorly (near baseline) in these simulations.
- The optimal balance shifted slightly toward more stimulation channels when more system identification data was available, but toward more recording channels when using spike-only signals or wider illumination beams.
Feasibility of Linear Control:
- Even without heavy optimization of the model or control parameters, the linear MPC approach achieved ~66% tracking error reduction and 93.6% reach direction classification accuracy at optimal channel counts.
- Closed-loop (MPC) significantly outperformed open-loop (OL) control. OL control offered limited gains at low channel counts and performed below baseline at high channel counts, underscoring the necessity of intra-trial feedback for latent factor control.
Robustness:
- These findings held across extensive parameter variations, including network area scale, neuron density, beam width, and the inclusion of task inputs in the model.

Significance and Claims
The paper claims to establish the feasibility of closed-loop latent factor control using accessible channel counts, challenging the prevailing assumption that single-neuron resolution is required to manipulate population dynamics. The results provide concrete design principles for future in vivo experiments and tool development:

Design Principle 1: High-density stimulation arrays do not need to match neuron counts; a fraction of channels can capture the majority of control utility.
Design Principle 2: Experimental interfaces should allocate channel budgets roughly equally between sensing and actuation, rather than prioritizing recording over stimulation.
Design Principle 3: Standard linear control methods are sufficient baselines for latent dynamics control, provided the channel allocation is optimized.

The authors posit that if these findings translate to in vivo preparations, the field could enter a new phase of causal characterization of latent dynamics, leveraging goal-driven modeling and optimal control to test hypotheses about neural population function. However, the paper modestly notes that these are predictions based on simplified models (rate-coding dynamics, lack of cell-type specificity) and that quantitative channel requirements may differ in biological systems.

Technical Summary: Feasible Closed-Loop Stimulation for Neural Latent Dynamics Control

More like this