Fig. 4: 8 × 8 eRNR prototypes for Mackey–Glass time series prediction.
From: Rotating neurons for all-analog implementation of cyclic reservoir computing
a An eRNR prototype consisting of eight 8-neuron eRNRs (i.e., M = 8 and N = 8). b NRMSE result for multistep-ahead Mackey-Glass time series prediction. The state matrix used in this experiment was obtained from the parallel output channels of the eRNR hardware. c, d. Two cases of one-step-ahead prediction with the Mackey–Glass time series result compared with the ground truth using c one eRNR (NRMSE = 0.17) and d eight parallel eRNRs (NRMSE = 0.03). e, f Phase space of the prediction compared with the ground truth using (c) one eRNR and (d) eight parallel eRNRs. The phase diagram was created by plotting the predicted and ground truth series y(t) for the x-axis and y(t − τ) for the y-axis.
