Deep Learning by Optical Computing – New Development in Silicon Photonics –

Abstract

Artificial intelligence based on deep learning has developed drastically in the past decade, requiring more computing capability. However, the saturation in Moore’s law is expected in the next decade, promoting research　for new computing technologies which do not rely on Moore’s law. A deep learning accelerator based on a Si programmable photonic integrated circuit (PIC) is one of the promising computing technologies. After the first demonstration in 2017, many proposals of optical neural network using Si photonics has been reported. To program a programmable PIC, an optical phase shifter is a key device. Currently a thermo-optic phase shifter is used in a programmable PIC, while its power consumption is too high for large-scale integration. To resolve this issue, we have investigated a hybrid metal-oxide semiconductor (MOS) optical phase shifter enabling an ultralow-power programmable PIC. We have also proposed a Si microring crossbar array which enables on-chip inference and learning.