Closed-loop brain stimulation stops chasing a fixed waveform: CLEAR-FORCE folds brain feedback into the error, inducing and suppressing oscillations in simulation
Synopsis
Closed-loop brain-stimulation controllers are usually trained against a preset target and struggle to follow real-time fluctuations in brain activity. CLEAR-FORCE combines the reservoir network's output error with a feedback error taken from the population membrane potential of a simulated spiking network, updating readout weights online while stimulation runs. In two spiking-neuron simulations it induced 30 Hz oscillations and suppressed burst events; error was lower than open loop when feedback strength was above zero, but weaker feedback did worse than open loop.
Interpretation
A reservoir-computing controller can learn while stimulation is being delivered, rather than separating a training phase from a stimulation phase. Earlier closed-loop controllers largely relied on offline training or preset parameters and could not track non-stationary changes in brain activity in real time. On a simulated brain network of 100 Izhikevich spiking neurons, switching the controller on produced sustained 30 Hz oscillations, while with it off spiking was sparse with no dominant band.
The controller is a recurrent reservoir of 100 units whose readout weights are updated online by recursive least squares. Standard FORCE learning uses only the error between output and target signal, whereas CLEAR-FORCE adds a second error term derived from the controlled spiking network. Total error is a weighted sum of the reservoir output error and the population-mean membrane-potential error, with weights λ_force and λ_IZ; setting λ_IZ to zero reduces it to open-loop FORCE.
The benefit of closed-loop feedback depends on its strength and is not monotonic. Across target oscillations from 5 to 30 Hz, error was generally lower with λ_IZ above zero than in the open-loop case at λ_IZ equal to zero. The optimal λ_IZ ranged from 3.1 to 4.7, and a low λ_IZ gave higher error than open loop, meaning open-loop stimulation outperformed closed loop when the feedback signal was too weak.
The same controller suppresses pathological bursts regardless of which neuron type originates them. With the controller off, raising external drive during a 5000–6000 ms window induced a burst; with it on, the same drive did not produce a burst. Whether bursts arose from excitatory or inhibitory neurons and whichever subtype the controller targeted, firing rates of both excitatory and inhibitory neurons stayed below 2 Hz, while rates were consistently higher with the controller off.
Perspective
The demonstrated scope is a simulated brain network of Izhikevich spiking neurons, and the controller target stays fixed in the oscillation-inducing simulations; the power estimate counts only controller computation, excluding sensing, stimulation output and communication.
In the oscillation-inducing case the target signal is given in advance, and the authors place the question of choosing optimal stimulation parameters from symptoms outside the current work; the spiking network uses a single Izhikevich parameter regime; the reservoir is not constrained to deliver the minimum effective stimulation amplitude.