Biomemristor Reservoir Computing With Multi-Value Mask for Improving Recognition Performance

In this letter, we report an efficient and affordable RC system where the masking process in input and reservoir layers are implemented by nonvolatile and volatile memristors, respectively. The nonvolatile memristors with a TiN/HfOx/TaOx/TiN structure show multilevel programmable resistance states, serving as multi-value mask matrices to increase the richness of reservoir state. Meanwhile, Ag/guar gum-WTe2 nanohybrid/Au biomemristor synapses exhibit short-term memory and nonlinear dynamics, fulfilling the implementation of physical reservoirs. The performance of the RC system was further evaluated by the tasks of speech digit recognition and electrocardiogram waveform classification. By implementing the multi-value mask, the accuracy rates of these two tasks were improved to 99.6% and 99.2%, respectively. Our findings suggest that the multi-value RC system shows a great potential for processing complex temporal signals in a low-cost and energy-efficient manner.