Yang Zhan, He Jiajing, Yan Zhouyuan, Hu Yibiao, Li Xiaobo, Dong Ningning, Wang Jun
Opt Express. 2024 Sep 9;32(19):34001-34014. doi: 10.1364/OE.531679.
Optical neural networks (ONNs) have been considered as an alternative solution to overcome the arithmetic and energy bottlenecks of electronic neural networks. However, the widespread implementation of ONNs is hindered by their lack of optical nonlinearity. In this work, three ultra-compact all-optical nonlinear activators are inverse-designed by combining the adjoint method and Kerr nonlinearity. The nonlinear response is mainly generated by the Kerr and the thermo-optic (TO) effect associated with the nonlinear refractive index. Transmission-as-computation and structure-as-function are realized, with a minimum activation threshold of 2.34 mW. In addition, we validated the feasibility and capability of the proposed method against benchmark machine learning tasks, in which the addition of nonlinear activation functions significantly improved the expressive power of the ONN, increasing the testing accuracy obtained from the Modified National Institute of Standards and Technology (MNIST) task from 88.15% to 93.25%. The proposed ONN framework with our nonlinear activators exhibited good robustness against phase errors in the network topology. We believe that this study contributes to the future development of large-scale chip-level ONNs.
光学神经网络（ONNs）被认为是克服电子神经网络算术和能量瓶颈的一种替代解决方案。然而，ONNs缺乏光学非线性阻碍了其广泛应用。在这项工作中，通过结合伴随方法和克尔非线性，反向设计了三种超紧凑型全光非线性激活器。非线性响应主要由与非线性折射率相关的克尔效应和热光（TO）效应产生。实现了传输即计算和结构即功能，最小激活阈值为2.34 mW。此外，我们针对基准机器学习任务验证了所提方法的可行性和能力，其中添加非线性激活函数显著提高了ONN的表达能力，将从改进的美国国家标准与技术研究院（MNIST）任务中获得的测试准确率从88.15%提高到了93.25%。所提出的带有我们的非线性激活器的ONN框架在网络拓扑中对相位误差表现出良好的鲁棒性。我们相信这项研究有助于大规模芯片级ONNs的未来发展。
