Harnessing the Intrinsic Dynamics of Biological Neural Networks for Reservoir Computing

Neuronal networks possess dynamic behaviors that can be utilized for reservoir computing, though their application is hindered by factors such as spatial variability, spontaneous transitions, and biological non-stationarity. This study examines whether the inherent dynamics of a neuronal culture can be recognized and utilized as a living reservoir without altering the existing network.

Through multielectrode-array recordings and electrical injections, researchers characterize spontaneous population dynamics, confirm spike detection, and assess reservoir characteristics like nonlinearity, fading memory, and state-dependent processing. The neuronal reservoir demonstrates nonlinear behavior, achieving a 95.83% accuracy for the XOR function at a specific operating point.

Moreover, it retains stimulation-induced state information, with relaxation times of about 30-46 milliseconds. Notably, the separability of the reservoir depends on its condition prior to stimulation, with a near-chaotic regime offering a wider range of high-separability regions compared to synchronized activity. The ability to distinguish distinct stimulation conditions as the input space expands from 10 to 40 classes, while preserving separability even amidst multi-hour neural drift, highlights the reservoir's robustness.

Incorporating this living reservoir into a closed perception-action loop allows it to control gameplay in Mario Kart using a fixed readout. These findings substantiate neuronal cultures as state-dependent living reservoirs that naturally support computation.