🧠 neuroscience

CLEAR-FORCE: A reservoir computing framework for closed-loop brain stimulation

The paper introduces CLEAR-FORCE, a novel reservoir computing framework that enables real-time, adaptive closed-loop brain stimulation to successfully induce or suppress neural oscillations, offering a promising solution for personalized treatment of neurological disorders on low-powered hardware.

Original authors: Moreno-Gris, A., Godin, C., Thivierge, J.-P.

Published 2026-10-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Moreno-Gris, A., Godin, C., Thivierge, J.-P.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

For decades, doctors have used electricity to calm the stormy brains of people with conditions like Parkinson's disease, essential tremor, and epilepsy. By delivering mild electrical pulses to specific areas, they can often relieve symptoms that medication cannot touch. However, the current way these devices work is somewhat rigid. They operate on a fixed schedule, sending out the same pattern of pulses regardless of what the brain is actually doing at that moment. This is like a thermostat that turns the heat on and off at the same times every day, ignoring whether the house is already warm or cold. Because brain activity changes constantly—shifting from minute to minute or even second to second—a fixed approach often misses the mark. It cannot adapt to the patient's changing state, leading to treatments that are either too weak to help or too strong to be comfortable.

To solve this, scientists are developing "closed-loop" systems, which act more like a smart thermostat. These devices constantly listen to the brain's electrical signals and adjust the stimulation in real time. If the brain starts to show signs of a seizure or a tremor, the device reacts immediately. The challenge has been creating a controller that is fast enough to keep up with the brain's chaotic, shifting rhythms without needing a massive computer to do the math. In a new study, researchers at the University of Ottawa have proposed a solution called CLEAR-FORCE. It is a type of artificial intelligence designed to learn on the fly, capable of both turning on helpful brain rhythms and turning off harmful ones, all while running on very little power.

The researchers tested their idea using a computer simulation of a brain network. They built a virtual brain made of one hundred tiny, interconnected units that fire electrical signals, much like real neurons. They then created a digital controller, a small network of one hundred units that could send signals to this virtual brain and listen to the results. The key innovation was how the controller learned. Instead of just guessing what to do, it watched the brain's reaction to its own signals and adjusted its strategy instantly. If the brain started firing in a chaotic, synchronized burst—a pattern often seen in disease—the controller learned to send a counter-signal to calm it down. If the brain was too quiet, the controller learned to inject a specific rhythm to wake it up.

In their simulations, the system proved remarkably effective. When the researchers wanted to create a steady, healthy rhythm in the virtual brain, they set a target frequency of 30 pulses per second. With the controller turned off, the brain remained quiet and irregular. But as soon as the controller was switched on, it quickly learned to drive the network into a steady, synchronized rhythm at that exact speed. The system worked even when the brain was already trying to make its own rhythm, successfully boosting the existing pattern. Conversely, when the virtual brain began to exhibit a dangerous, high-energy burst of activity, the controller learned to suppress it. It did this by detecting the surge and immediately delivering a signal that dampened the excitement, bringing the network back to a calm, steady state. This happened in real time, with the controller adapting to the brain's behavior as it unfolded.

The study also looked at how well this system could handle different scenarios. The researchers found that the controller worked best when it paid close attention to the brain's feedback. If the system ignored the brain's response and just followed a preset plan, it was less effective. However, when the controller used the brain's own signals to guide its learning, it could induce or suppress rhythms with high precision across a wide range of speeds. It even worked when the brain was made of different types of cells or when the controller only had access to a portion of the network. The system was robust, meaning it didn't break down when conditions changed slightly.

Perhaps most importantly for the future of medical devices, the researchers calculated how much energy this system would need to run on a real implant. Because the controller is designed to be efficient, it requires very little power. The study suggests that a device with a few thousand units could manage a brain network of millions of neurons while consuming only a few milliwatts of power. This is a level of efficiency that fits within the limits of battery-powered implants that patients can carry for years. The researchers emphasize that while these results come from computer simulations, they demonstrate a path forward for creating brain implants that are not just smart, but also energy-efficient and capable of adapting to the unique, changing needs of every individual patient.

Technical Summary of "CLEAR-FORCE: A reservoir computing framework for closed-loop brain stimulation"

Problem Statement
Electrical stimulation is a proven therapeutic intervention for neurological disorders such as Parkinson's disease, Tourette syndrome, and epilepsy. However, current clinical approaches largely rely on open-loop stimulation with manually tuned parameters (amplitude, frequency, wave shape). This static approach is inefficient because brain dynamics are non-stationary; patient states fluctuate over timescales ranging from minutes to months, requiring frequent recalibration. While closed-loop systems that adapt to real-time neural activity offer a solution, existing machine learning (ML) controllers often fail to operate in real-time or lack the capacity to simultaneously enhance and suppress complex, nonlinear neural oscillations. The core challenge lies in the "organized complexity" of neural systems, where outcomes depend on nonlinear interactions that are difficult to model or predict for precise control.

Methodology
The authors propose CLEAR-FORCE (Closed-Loop Error-based Adaptive Regulation, based on First-Order, Reduced and Controlled Error), a novel framework utilizing reservoir computing to train a real-time neural controller. The system consists of two primary components:

Simulated Brain Network: A population of $N=100$ spiking neurons (80% excitatory, 20% inhibitory) modeled using the Izhikevich equations. These neurons receive external current driven by the controller.
Neural Controller: A recurrent reservoir network ( $N=100$ units) that projects activity to the brain network. Unlike standard FORCE learning, which relies solely on the error between the reservoir output and a target signal, CLEAR-FORCE employs a hybrid error formulation.

The learning rule combines two error terms to update the readout weights ( $w_{out}$ ) via recursive least squares:

$e_{force}$ : The difference between the reservoir output and a desired target signal (e.g., a sine wave).
$e_{IZ}$ : The difference between the average membrane potential of the spiking network (serving as a proxy for Local Field Potentials) and the target signal.

The total error $e(t)$ is a weighted sum of these terms ( $e(t) = \lambda_{force}e_{force}(t) + \lambda_{IZ}e_{IZ}(t)$ ). This allows the controller to adapt to the non-stationary responses of the target network without requiring an explicit model of the brain dynamics. The framework was tested in two distinct scenarios: (1) inducing specific neural oscillations (synchronization) and (2) suppressing pathological bursting activity (desynchronization).

Key Results

Inducing Oscillations: The controller successfully induced sustained oscillations in a sparse, non-oscillatory spiking network. When trained with a 30 Hz sine wave target, the system generated strong 30 Hz activity. Comparative analysis showed that closed-loop training ( $\lambda_{IZ} > 0$ ) significantly reduced error compared to open-loop training ( $\lambda_{IZ} = 0$ ) across frequencies ranging from 5 Hz to 30 Hz. Optimal performance was achieved with $\lambda_{IZ}$ values between 3.1 and 4.7. The system remained effective even when only a subset of reservoir units or target neurons were active.
Suppressing Pathological Oscillations: The framework successfully suppressed transient, high-synchrony bursts in a network driven by tonic current. By subtracting the reservoir output from the tonic drive, the controller quenched bursts while preserving background activity. In real-time simulations, the controller prevented bursts from emerging when external drive was increased, whereas the uncontrolled network exhibited bursting. This suppression was robust regardless of whether the bursts originated from excitatory or inhibitory neurons and regardless of which neuronal subtypes the controller targeted.
Power Consumption and Scalability: The authors estimated the power requirements for neuromorphic hardware implementation. Results indicated that error scales linearly with the number of target spiking neurons ( $N_{spk}$ ) but decreases exponentially with the number of reservoir units ( $N_{res}$ ). This suggests a logarithmic relationship ( $N_{res} = \mathcal{O}(\log N_{spk})$ ) is sufficient to control large populations (e.g., human striatum). For a reservoir of ~2,000 units, the estimated dynamic computing power is in the range of a few to tens of milliwatts, compatible with low-power implantable hardware.

Significance and Claims
The paper positions CLEAR-FORCE as a promising computational framework for adaptive, low-power neural control. Its primary significance lies in bridging the gap between advanced reservoir computing theory and the practical constraints of implantable neuromodulation hardware.

Model-Free Adaptation: Unlike Bayesian optimization or deep learning approaches that may require offline training or heavy computational resources, CLEAR-FORCE offers a model-free, continuous real-time adaptation capability.
Dual-Functionality: The framework demonstrates the ability to both synchronize and desynchronize neural populations, addressing the bidirectional needs of different therapeutic contexts (e.g., enhancing gamma power for depression vs. suppressing beta bursts for Parkinson's).
Hardware Feasibility: By leveraging the efficiency of reservoir computing, the authors argue that personalized, closed-loop brain stimulation can be implemented on low-powered, portable devices, moving beyond the limitations of current open-loop or computationally intensive closed-loop systems.

The authors maintain a modest stance, noting that the current work relies on fixed target frequencies and a single neuronal parameter regime. They identify future directions such as dynamic target adaptation, personalized Izhikevich regimes, and the incorporation of synaptic plasticity (STDP) to induce long-term therapeutic changes, but do not claim these are currently implemented.

More like this