Investigation on the 3D Memristor Array Architecture for 3D Reservoir Computing System Implementation

作者：Wenxuan Sun, Jie Yu, Danian Dong, Xu Zheng, Jinru Lai, Shaoyang Fan, Hongzhou Wang, Jianfeng Gao, Junfeng Liu, Xiaoxin Xu · 发表于：IEEE Electron Device Letters · 年份：2024 · DOI：10.1109/led.2024.3417691 · 被引用次数：6 · 研究领域：Advanced Memory and Neural Computing、Neural Networks and Reservoir Computing、Machine Learning and ELM

Three-dimensional reservoir computing (3D RC) system is an energy efficient recurrent neural network for achieving high area efficiency. Considering the physical implementation of 3D RC by memristor, 3D Vertical Memristor Array (3D-VMA) is proved a more economical and reliable approach than 3D Cross-point Memristor Array (3D-CMA). Due to the natural competitive parallel structure in the 3D-VMA, the cumulative effect of the devices is better, which improves the differentiation between reservoir states. Meanwhile, the system noise is reduced since the Cycle-to-Cycle (C-to-C) variation is well controlled. The Mackey-Glass (M-G) prediction shows that the prediction performance of RC system implemented by the 3D-VMA improves by 4.2 times compared with that by 3D-CMA. The highly robust 3D vertical architecture proves the novel approach to realize high-performance 3D RC system.