[17a-WL2_101-1~9] 3.2 Information photonics and image engineering

11:15 AM - 11:30 AM JST(2:15 AM - 2:30 AM UTC)Presentation by Applicant for JSAP Young Scientists Presentation Award

[17a-WL2_101-7] Enhancing Generalizability in Dielectric Metasurface-basedHybrid Optical Neural Networks via Supervised Contrastive Learning

〇(B)Deyu Cao¹, Chun Ren², Chiyun Li², Andre Rohm³, Ryoichi Horisaki³, Takuo Tanemura^1,2 (1.Faculty of Engineering, the Univ. of Tokyo, 2.Graduate School of Engineering, the Univ. of Tokyo, 3.Graduate School of Information Science and Technology, the Univ. of Tokyo)

Keywords:

metasurface,Optical neural network,Image Recognition

We introduce supervised contrastive learning to dielectric metasurface-based hybrid optical neural networks to extract robust, task-agnostic features in the passive optical domain. Our simulations demonstrate that an optical encoder pretrained on CIFAR-100 achieves 69.4% classification accuracy — matching an all-digital AlexNet with only 1/24 of the digital MAC operations — on CIFAR-10 and consistently exhibits superior generalization performance on other out-of-distribution datasets compared to standard training methods.