Spiking Reservoir Computing Based on Stochastic Diffusive Memristors

Abstract Reservoir computing (RC), a type of recurrent neural network, is particularly well‐suited for hardware implementation in edge computing. It is shown that RC hardware based on dynamic memristors potentially offers much lower power consumption and reduced computation times than digital electronics. However, challenges such as stochasticity and read noise in these devices can impair its performance. Furthermore, the external analog‐to‐digital (ADC) readout circuits may require substantial area and energy. In this work, it is experimentally demonstrated that a population of stochastic diffusive Ag:SiO x memristors can effectively construct a spiking reservoir computing system. This system demonstrates remarkable resilience to read noise and delivers exceptional performance across a range of computational tasks, achieving a 98% accuracy in waveform classification and a normalized root mean square error (NRMSE) of 0.154 in time‐series prediction. Further simulations reveal that a certain degree of device stochasticity actually enhances system performance. Without using ADC converters, a hybrid memristor‐CMOS spiking RC system is designed that demonstrates significantly lower power consumption compared to fully digital systems.