Chaotic Time Series Prediction in Biological Neural Network Reservoirs on Microelectrode Arrays

We encoded a chaotic time series produced by the logistic map as delays between adjacent stimulation pulses and used in vitro neural networks reservoirs combined with ridge regression to predict future time steps of 1-15 time step horizons. We control our results by replicating the training procedure on synthetic data containing stimulation events without spikes.Our results show that some networks outperformed the control experiment on both Mean Absolute Error (MAE), Median Absolute Error (MedAE) and R² score, but only for longer prediction horizons of 4, 5 and 6 time steps, where the target function reaches substantial complexity. Best MAE increase was 32 % at 5 time step and 27 % at 6 time step -prediction. The two best networks showed a low lower and upper bound as well as low mean spike counts during extracted epochs compared to the other networks.These results show that biological in vitro neural network reservoirs can be used for chaotic time series prediction. A number of challenges must however be solved to effectively utilize such neural reservoirs, as only two out of six networks showed substantial improvements over our control setup. Effective methods of consistently producing good performing neural reservoirs is therefore needed.