Memory properties of electromigrated Au nanogaps to realize reservoir computing

• Published in: Applied physics letters Vol. 119; no. 8
• Main Authors: Sakai, Keita, Yagi, Mamiko, Ito, Mitsuki, Shirakashi, Jun-ichi
• Format: Journal Article
• Language: English
• Published: Melville American Institute of Physics 23-08-2021
• Subjects: Applied physics; Artificial intelligence; Computation; Control methods; Electromigration; Memory tasks; Performance evaluation; Plastic properties; Synapses
• ISSN: 0003-6951; 1077-3118
• DOI: 10.1063/5.0055352

The reservoir computing (RC) scheme, which employs the inherent computational capabilities of dynamical systems, is a key technology to implement artificial intelligence systems physically. Ensuring the nonlinear expansion of input data through the dynamics of physical systems is a necessary aspect of RC. Previously, we developed artificial synapses of Au nanogaps by using the “activation” technique, which allowed the implementation of synaptic functions such as short-term plasticity, long-term plasticity, and spike-timing-dependent plasticity. The activation technique is an electromigration-based method to control the tunnel resistance of nanogaps. In this study, the memory property of the Au nanogap, using activation for RC, was evaluated via short-term memory (STM) and parity check (PC) tasks. More specifically, memory capacity was introduced to evaluate the performance of the Au nanogap, defined as the sum of squares of the correlation between the outputs of RC and the teacher for delay D = 1 to 6. By utilizing the simple dynamics of short-term plasticity, the memory capacities of the STM and PC tasks were found to be 1.07 and 0.90, respectively, when 10 virtual nodes were used. This demonstrates that the dynamic process of the activation technique enables the Au nanogap-based reservoir to process information directly in the temporal domain. The experimental results can facilitate the development of compact devices to realize physical RC.