In vitro living neurons in a reservoir computing framework

The Reservoir Computing (RC) paradigm has been devised for the processing of dynamic, real-time, and multimodal input signals. The core mechanism of RC relies on decoding output signals from the reservoir in a supervised manner, while the reservoir itself serves the purpose of “digesting” inputs recurrently. This leads to computationally efficient machinery that does not require expensive backpropagation weight updates, as in classical recurrent neural networks (e.g. RNN/LSTM). Nonetheless, the choice of the reservoir is critical, and a good reservoir requires the ability to express complex dynamics with sufficient non-linearity and separation properties. In this work, we target live in vitro neuron cultures, known for their scalable hypernetworks and complex spiking dynamics, making them potentially suitable for constructing quality reservoirs. Additionally, the self-organizing and malleable nature of neuronal synapses may allow the biological neural network to adapt to problems continuously and more quickly.

To build a living reservoir from neuron cultures, we integrate optogenetic stem cell-derived neurons with micro-electrode arrays (MEA). Electrical and optical stimulation are employed to feed encoded information into the biological neural networks (BNN), and the reservoir state is represented by spiking activities. We exploit the reservoir capability of BNNs and demonstrate them through benchmark problems.