Fig. 8: Nonlinear autoregressive moving average (NARMA) task.
From: Self-organizing neuromorphic nanowire networks as stochastic dynamical systems
Simulation results of the NARMA-2 task in terms of NMSE as a function of a the bias voltage, and b steady state \(\widetilde{g}\) by considering deterministic and stochastic dynamics and for various amplitudes of the input signal. c NMSE as a function of the input signal amplitude for deterministic and stochastic dynamics, by considering polarization voltages of 3.6 V (\(\widetilde{g}\) ~ 0.5) and 5 V (\(\widetilde{g}\) ~ 0.99). Predictions of NARMA-2 relative to polarization voltages of 3.6 V and 5 V obtained with optimal parameters through deterministic dynamics in d, f, respectively ([\(N,\varTheta\)] in d, f are [9,1] and [2,7], respectively). Colormaps showing task performances as a function of \(N\) and \(\varTheta\) parameters for polarization voltages of 3.6 V and 5 V in e, g, respectively. Predictions of NARMA-2 relative to polarization voltages of 3.6 V and 5 V obtained with optimal parameters through stochastic dynamics in h, j, respectively ([\(N,\varTheta\)] in d, f are [3,1] and [3,5], respectively). Colormaps showing task performances as a function of \(N\) and \(\theta\) parameters for polarization voltages of 3.6 V and 5 V in i, k, respectively. Colormaps and predictions in d–k refer to results obtained by stimulating the network with an input with an amplitude of 50 mV.
