Comparison of digital, photonic, and memristor-based reservoir computing
Nataniel Holtzman, Patrick Edwards, Matthew J. Marinella, George C. Valley · 2025
Reservoir computing (RC) has been demonstrated in many media over the past 20 years, and the analytical formulation for conventional RC is now well-established. Here we compare the digital neural network RC to photonic and memristor RCs, considering properties such as feedback, signal bias, nonlinearities, internal weights, masking, leaking rate, and output weights. Even though photonic and memristor RCs each lack many of these properties, they are still able to provide excellent RC performance. We illustrate the comparison of the digital, photonic, and memristor RCs using three standard test problems: (1) Mackey-Glass (MG) waveform classification, (2) MG free-running prediction, and (3) nonlinear channel equalization, also known as the symbol error rate (SER) problem.