A master course student, Yuki Kume, made a presentation in ASPDAC2020 held at China National Convention Center, Beijing, China.
Echo State Network (ESN), which is a kind of recurrent neural network (RNN), has recently attracted many attentions. The weights of the input and middle layers of the ESN are randomly determined constants that need not to go through a training process. ESN has low calculation cost compared with LSTM, an other type of RNN, and thus it is suitable for hardware implementation. In this presentation, we proposed a method of expressing random weights by a MOSFET crossbar array circuit, a method of controlling the feedback parameters to satisfy convergence requirements, and an architecture of ESN to enhance accuracy under the nonidealities due to the hardware implementation. Experiments and evaluations has shown that the proposed “Dual-MOS-ESN” achieves equal accuracy as the software ESN.
- Yuki Kume, Song Bian, and Takashi Sato, “A tuning-free hardware reservoir based on MOSFET crossbar array for practical echo state network implementation,” in Proc. ACM/IEEE Asia and South Pacific Design Automation Conference (ASPDAC), pp.458-463, January 2020.
