@inproceedings{oai:nagoya.repo.nii.ac.jp:02003937,
 author = {Ichikawa, Taisei and Masuda, Yutaka and Ishihara, Tohru and Shinya, Akihiko and Notomi, Masaya},
 book = {2022 IEEE Computer Society Annual Symposium on VLSI (ISVLSI)},
 month = {Oct},
 note = {Optoelectronic implementation of artificial neural networks (ANNs) has been attracting significant attention due to its potential for low-power computation at the speed of light. Among the ANNs, adopting recurrent neural network (RNN) is a promising solution since it provides sufficient inference accuracy with a more compact structure than other ANNs. This paper proposes a novel optoelectronic architecture of RNN. The key idea is to implement the vector-matrix multiplication optically to exploit the speed of light and implement the activation and feedback electronically to exploit the controllability of electronics. The electronics part is composed of an electrical feedback circuit with a dynamic latch to synchronize the recurrent loops with a clock signal. Using a commercial optoelectronic circuit simulator, we confirm the correct behavior of the optoelectronic RNN. Experimental results obtained using TensorFlow show that the proposed optoelectronic RNN achieves more than 98% inference accuracy in image classification with a minimal footprint without sacrificing low-power and high-speed nature of light., 2022 IEEE Computer Society Annual Symposium on VLSI (ISVLSI). 04-06 July 2022. Nicosia, Cyprus},
 publisher = {IEEE},
 title = {Optoelectronic Implementation of Compact and Power-efficient Recurrent Neural Networks},
 year = {2022}
}