Performance improvement of a Ag-ion controlled molecular-gap atomic switch by reducing a switching area for applying to a deep learning system

Abstract In today’s advanced information society, hardware-based neuromorphic systems attract much attention for achieving more efficient information processing. Hardware-based neuromorphic systems need devices that change their resistance in an analog manner like biological synapses. A molecular-gap atomic switch exhibits analog resistance change over a wider range compared to other non-volatile memory devices. However, several issues remain with the device, such as in cyclic endurance and retention. In this study, we fabricated a molecular-gap atomic switch with a reduced switching area. We expected that the reduction would limit the number of Ag + cations that contribute to a switching phenomenon and solve the remaining issues. The fabricated devices endured 1000 switching cycles and exhibited stable analog resistance change. Deep learning was successfully demonstrated using 293 fabricated devices as synapses, which resulted in the accuracy of 93.65% in 26th epoch in a 5 × 5 pixel image classification task.