Harnessing the Intrinsic Dynamics of Biological Neural Networks for Reservoir Computing

Abstract

Living neuronal networks exhibit nonlinear, recurrent, and evolving dynamics that make them promising substrates for reservoir computing, yet their computational use is complicated by spatial heterogeneity, spontaneous state transitions, and biological nonstationarity. Here, we investigate whether the native dynamics of a neuronal culture can be characterized and harnessed as a living reservoir without deliberately modifying the underlying recurrent network. Using multielectrode-array recordings and electrical perturbations, we characterize spontaneous population dynamics, validate channel-adaptive spike detection, and evaluate reservoir properties including nonlinearity, fading memory, state-dependent processing, separability, scalability, and temporal robustness. The neuronal reservoir exhibits nonlinear transformation, achieves 95.83\% XOR accuracy at the selected operating point, and retains stimulation-induced state information with characteristic relaxation times of approximately 30--46~ms. Importantly, separability depends on the pre-stimulation network condition, with a near-chaotic regime supporting broader high-separability regions than synchronized activity. Distinct stimulation conditions remain discriminable as the input space increases from 10 to 40 classes, while separability persists despite multi-hour neural drift. Finally, the reservoir is incorporated into a closed perception--action loop for Mario Kart control using a fixed readout. These results establish neuronal cultures as state-dependent living reservoirs whose native dynamics support computation.