|
| 面向原位训练的集成衍射神经网络 |
| Towards in-situ training integrated diffractive neural networks |
| 投稿时间:2025-02-10 |
| DOI:10.3969/j.issn.1005-5630.202502100020 |
| 中文关键词: 衍射神经网络 光计算 片上集成 补偿算法 原位训练 |
| 英文关键词:diffractive neural networks optical computing on-chip integration compensation algorithms in-situ training |
| 基金项目:上海市青年科技启明星计划(21QA1403600);上海市自然科学基金 (21ZR1443400) |
|
| 摘要点击次数: 972 |
| 全文下载次数: 1637 |
| 中文摘要: |
| 衍射神经网络利用光的波动特性实现高效信息处理,在低功耗、高效率的光计算应用中展现出巨大潜力。当前的研究主要集中于空间衍射神经网络和二维衍射神经网络两类。空间结构中各衍射层之间相互分离,容易因对准误差导致系统稳定性下降。通过片上集成可显著提升二维结构的紧凑性,但在输入输出信号的处理以及制造工艺中仍会存在误差,通常需要依赖补偿算法进行修正。基于原位训练的概念,提出了一种单方向参数更新策略,旨在统一网络训练与制造流程,并将其应用于光学神经网络中。通过不同复杂任务的分类准确率进行验证,结果表明,该策略在性能上与传统更新方法相当。该方案兼容了训练过程与制造工艺,为衍射神经网络的原位训练提供了高效可行的优化方案,并为光学神经网络的小型化与集成化奠定了基础。 |
| 英文摘要: |
| Diffractive neural networks leverage the wave properties of light to achieve efficient information processing, demonstrating significant potential for low-power, high-efficiency optical computing applications. Current research primarily focuses on two types of diffractive neural networks, spatial diffractive neural networks and two-dimensional diffractive neural networks. In spatial structures, the diffractive elements are spatially separated, rendering such system susceptible to stability issues induced by alignment errors. In contrast, two-dimensional structures achieve improved compactness through on-chip integration, yet errors still persist in signal input, signal output, and manufacturing precision, often requiring compensation algorithms for correction. Building on the concept of in-situ training, we propose a unidirectional parameter update strategy to unify the network training process with manufacturing workflows and apply it to optical neural networks. Tests on classification tasks with different complexity show that this strategy achieves performance comparable to standard update methods. This approach integrates the training process with manufacturing techniques, providing an efficient and feasible optimization scheme for enabling the in-situ training of diffractive neural networks. Moreover, it establishes a solid foundation for the miniaturization and integration of optical neural networks. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|