红外与毫米波学报, 2020, 39 (1): 13, 网络出版: 2020-03-12
基于10.6微米全光深度神经网络衍射光栅的设计与实现
摘要
光子人工智能芯片以光速执行运算,且具有低功耗、延迟低、抗电磁干扰的优势。小型化与集成化是实现这一技术革新的关键步骤。本文将光刻技术运用于衍射光栅的制作,提出一种基于10.6微米激光的全光衍射深度学习神经网络光栅设计及实现方法。由于光源波长由毫米波向微米波进化,神经元的特征尺度缩小至20微米,与现有光衍射神经网络相比,深度学习神经网络特征尺寸缩小了80倍,为进一步实现光子计算芯片大规模集成奠定了基础。
Abstract
The photonic artificial intelligent chip performs calculations at the speed of light, and has the advantages of low power consumption, low delay, and anti-electromagnetic interference. Miniaturization and integration are the key steps to realize this technological innovation. In this paper, lithography is applied to the fabrication of diffraction gratings. A design and implementation method of all-optics diffraction deep learning neural network grating based on 10.6 micron laser is proposed. Since the wavelength of the light source evolved from the millimeter wave to micrometer wave, the characteristic scale of the neuron are reduced to 20 micrometers. Compared with the existing optical computing neural network, the feature size of the deep learning neural network is reduced by 80 times, which laid the foundation for further large-scale integration of photonic computing chips.
Hai-Sha NIU, Ming-Xin YU, Bo-Fei ZHU, Qi-Feng YAO, Qian-Kun ZHANG, Li-Dan LU, Guo-Shun ZHONG, Lian-Qing ZHU. 基于10.6微米全光深度神经网络衍射光栅的设计与实现[J]. 红外与毫米波学报, 2020, 39(1): 13. Hai-Sha NIU, Ming-Xin YU, Bo-Fei ZHU, Qi-Feng YAO, Qian-Kun ZHANG, Li-Dan LU, Guo-Shun ZHONG, Lian-Qing ZHU.