Reservoir Computing on Atomic Switch Arrays with High Precision and Excellent Memory Characteristics

Abstract

Reservoir computing (RC) refers to an artificial neural network framework that exhibits temporal dynamic behavior, and its computational structure can be implemented in physical systems. However, the network performance of conventional RC is limited in terms of power scale, readability, spatial and temporal scalabilities, mass producibility, and precision. Therefore, we propose an RC with atomic switches. An atomic switch is a new type of nanodevice that exhibits superior power scale, readability, spatial and temporal scalabilities, and mass producibility. In the proposed RC architecture, we arranged atomic switches sequentially in a ring formation and used time-division multiplexing. The resistance of the atomic switches in the proposed architecture changes nonlinearly with a change in input and memorizes the input; therefore, this RC architecture is expected to have high precision and a large memory capacity (MC). In this study, we simulated this architecture and compared it with the conventional architecture in terms of precision and MC. The results showed that the proposed RC architecture had higher precision and greater MC.