Prof. Mitsuru Takenaka and his research group have proposed a new silicon optical circuit based optical neural network with a crossbar array of ring resonators and developed a deep learning accelerator that can accelerate learning as well as inference. They have demonstrated that this accelerator can accelerate learning computation by a factor of more than 1,000. As performance improvement through semiconductor miniaturization tends to slow down today, the computing power required for AI is demanded beyond its Moore’s Law. Optical neural networks on silicon circuits have been mainly limited to inference applications, but now, for the first time, they have succeeded in accelerating learning computation as well as inference. The paper was published in the online edition of “ACS Photonics” on July 22.
[Link]