正文：
近日，清华大学的光电智能技术交叉创新团队在智能光计算领域取得重大突破，相关研究成果发表在国际顶级学术期刊《Nature》上。该团队由戴琼海院士和方璐教授领衔，与薛智威、周天贶等研究人员共同完成。论文共同通讯作者还包括清华电子系徐智昊教授和之江实验室虞绍良研究员。
该研究提出了全前向模式（FFM）训练方法，这是智能光计算领域的一项创新性进展。传统的光计算训练依赖于基于数据的模型建立，这在实际操作中存在难以修正实验误差和难以完美建模光学系统的瓶颈。而全前向模式训练方法摒弃了传统的模型构建步骤，直接在物理光学系统上设计和调整光学参数，利用梯度下降算法有效地得出最终的模型参数。这种方法不仅提高了机器学习操作的并行性，还大大减轻了AI对光学系统建模的限制。
全前向模式训练方法利用光子传播的对称性，将神经网络训练中的前向与反向传播等效为光的前向传播。这种方法使得光神经网络训练可以在没有分析模型的情况下自动搜索非厄米异常点，并且能够支持通过散射介质进行全光学聚焦，分辨率达到衍射极限。同时，该方法还能够以超过千赫兹的帧率平行成像隐藏在视线外的物体，并在室温下进行光强弱至每像素亚光子的全光处理。
这项研究不仅有助于将学习过程提高几个数量级，还能够推动深度神经网络、超敏感感知和拓扑光学等应用和理论领域的发展。清华大学的研究团队通过实验验证了全前向模式训练方法的有效性，并在MNIST和Fashion-MNIST数据集上取得了显著的实验成果。
全前向模式训练方法的提出，为智能光计算领域的发展提供了新的思路和技术支撑，有望成为未来计算架构的有力竞争方案之一。随着技术的不断进步，智能光计算有望在人工智能、自动驾驶、生物医学成像等领域发挥重要作用。
英语如下：
Title: Tsinghua University Breakthrough in Nature: Revolutionary All-Forward Intelligent Optical Computing Architecture
Keywords: Tsinghua University Nature, Optical Computing, Innovative Architecture
Content:
Recently, a cross-disciplinary innovation team from Tsinghua University in the field of optoelectronic intelligent technology has achieved a significant breakthrough in the field of intelligent optical computing. Their research findings have been published in the internationally renowned academic journal Nature. Led by Academician Dai Qionghua and Professor Fang Lu, the team also includes researchers such as Xue Zhiwei and Zhou Tiankuan. The paper also has co-corresponding authors from the School of Electronic Engineering at Tsinghua University, Xu Zhihao, and researcher Yu Shao’el from the Zhejiang Institute of Advanced Technology.
This research proposes an all-forward mode (FFM) training method, which is an innovative advancement in the field of intelligent optical computing. Traditional optical computing training relies on the establishment of data-based models, which in practical operation suffers from limitations such as difficulty in correcting experimental errors and the inability to perfectly model optical systems. The all-forward mode training method abandons the traditional model construction step, directly designing and adjusting optical parameters on the physical optical system using gradient descent algorithms to effectively obtain the final model parameters. This method not only improves the parallelism of machine learning operations but also significantly reduces the constraints of AI on optical system modeling.
The all-forward mode training method utilizes the symmetry of photon propagation to equate forward and backward propagation in neural network training with forward propagation of light. This method enables the training of optical neural networks to automatically search for non-Hermitian exceptional points without analyzing the model, and it supports full-optical focusing through scattering media, with resolution reaching the diffraction limit. Additionally, the method can also perform parallel imaging with a frame rate exceeding 1 kHz of objects hidden from sight, and conduct full-optical processing with light intensities as weak as sub-photon per pixel at room temperature.
This research not only helps to raise the learning process by several orders of magnitude but also pushes the development of applications and theoretical fields such as deep neural networks, ultra-sensitive perception, and topological optics. The Tsinghua University research team has experimentally verified the effectiveness of the all-forward mode training method and achieved significant experimental results on the MNIST and Fashion-MNIST datasets.
The introduction of the all-forward mode training method provides a new idea and technical support for the development of intelligent optical computing, and it is expected to become one of the powerful competition solutions for future computing architectures. As technology continues to advance, intelligent optical computing is poised to play a significant role in fields such as artificial intelligence, autonomous driving, and biomedical imaging.
【来源】https://www.jiqizhixin.com/articles/2024-08-08-5
Views: 1