講演情報
[8p-N205-3]Investigation on Photonic-Electronic Hybrid Convolutional Neural Network using Multi-Layer Metasurfaces and Photonic Integrated Circuit
〇(D)Chiyun Li1, Chun Ren1, Takahiro Suganuma1, Takuo Tanemura1 (1.The Univ. of Tokyo)
キーワード：
Optical Neural Networks、Metasurfaces、Photonic Intergrated Circuits
With the rapid expansion of the scale and application scope of deep neural networks (DNNs), the number of model parameters, the amount of computation, especially the inference calculation time and power consumption, have been increasing dramatically. As an alternative to traditional GPU-based inference systems, the potential of optical computing is being explored. Among them, the metasurface (MS) with a high-density integration of subwavelength fine scattering structures can achieve large-scale convolutional operations without power consumption and at ultra-high speed just by light transmission, so it has attracted rapid attention in recent years. So far, photonic-electronic hybrid neural networks (NNs) such as connecting the MS to an image sensor have been proposed and their effectiveness has been demonstrated. On the other hand, since these approaches use a large number of pixels on the image sensor, there are limitations in terms of speed up and power saving. In this study, by using multiple such metasurfaces (MSs) and further integrating programmable photonic integrated circuits (PICs), a photonic-electronic hybrid convolutional neural network (CNN) is proposed, which significantly reduces the processing of the photonic-electronic conversion part and the digital NN. Through the use of both variable controllability by multiple convolutional layers composed of multi-layer MSs and PICs, it has been clarified by numerical verification that a high classification accuracy can be maintained while notably reducing the scale of the digital NN at the electrical stage.