[16p-A25-1~16] 3.14 Silicon photonics and integrated photonics

3:45 PM - 4:00 PM JST(6:45 AM - 7:00 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[16p-A25-9] Demonstration of 32-input optical neural network using multi-plane light conversion

〇(M2)Chun Ren¹, Ryota Tanomura¹, Kazuki Ichinose¹, Yoshiaki Nakano¹, Takuo Tanemura¹ (1.Univ. of Tokyo)

Keywords:

optical neural network,photonic integrated circuit,multi-plane light conversion

Photonic integrated circuits can reduce power consumption of matrix-vector multiplication required in deep learning, while the conventional method using Mach-Zehnder interferometers has not achieved large scale due to a lack of scalability. In this research, multi-plane light conversion is employed with multiport directional couplers and approximately 1/10 of phase shifters to realize the world’s first 32-input optical neural network on a single photonic chip, experimentally demonstrating 2-class image recognition extracted from MNIST at 90.3% test accuracy.