Fully Flash-Based Reservoir Computing Network With Low Power and Rich States
Yang Feng, Mingfeng Tang, Zhaohui H. Sun, Yueran Qi, Xuepeng Zhan, Jing Liu, Junyu Zhang, Jixuan Wu, Jiezhi Chen · IEEE Transactions on Electron Devices · 2023
A flash-based dynamic reservoir computing (RC) system is proposed. It is demonstrated that after applying a high bias voltage, the flash memory shows nonlinear transform and short-term memory properties with low current, which can be utilized as the reservoir node. Meanwhile, the cell’s nonlinear characteristic determined by the stressing time can be controlled to generate rich reservoir states. Also, the low current of flash memory greatly reduces the power consumption to 0.8 pJ per input. When executing time-series prediction tasks, high accuracy is achieved with a low normalized root mean square error (NRMSE) of 0.0096. The RC system based on flash technology exhibits significant potential as a large-scale neural network with exceptionally low power consumption. Our findings suggest that flash-based RC systems could be an attractive option for neuromorphic computing applications, where energy efficiency is a critical consideration.