Leveraging deep learning to control neural oscillators

Data-Driven Control of Neuronal Networks with Population-Level Measurement

“…Our control technique differs from the existing data-driven techniques for complex networks [25][26][27][29][30][31] in two important aspects. Firstly, we do not extract information from a rich pre-training data set, as in [25][26][27], and directly learn the appropriate control online from scratch. Online interactions with the system enable the control synthesis to effectively adapt to system variation.…”

“…Online interactions with the system enable the control synthesis to effectively adapt to system variation. Secondly, unlike Reinforcement Learning [25][26][27], we do not learn from a limited reward signal, which without a hand-crafted function requires many trials with a large amount of data. Instead, we utilize the rich local dynamics of the system along its current trajectory to step-by-step regulate the system toward the desired synchrony.…”

“…In this paper, we address this problem by designing control inputs that can synchronize or desynchronize a network of neurons within a pre-specified time frame. The advocated framework differs from the existing works [25][26][27][28][29][30][31] in that it does not require the availability of a rich training data set but instead leverages the system local dynamics to directly learn appropriate control with a small number of trails. We learn the control input online by iteratively perturbing the system and observing its response.…”

“…The controlled trajectories, designed input, ∥L∥ in each iteration, and the estimated neuron phases in the final cycle are shown in Figures 3(b)-(e). It is worth noting that to archive (de)synchronization, some existing data-driven neuronal control methods in fact require a rather complete set of pre-training data with a uniform distribution [25][26][27]. In practical applications, such complete data sets are often not readily available and must be carefully collected during experiments.…”