发表机构
Korea Advanced Institute of Science and Technology (KAIST); Hanyang University（韩国科学技术院; 汉阳大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出突破秩1限制的随机映射强度光学神经网络（RMI-ONN），结合无序介质、强度掩模与功率求和实现全光两层计算，在单一光学配置下完成多模态分类，为可扩展光处理器奠定基础。
AI 中文摘要
自由空间光学神经网络因能直接对入射光场进行计算，在计算成像和机器视觉领域具有独特优势。然而，由级联线性光学组件构成的传统光学神经网络（ONN）受限于输入光场x与输出得分y之间的平方律检测，呈现通用线性输入-输出关系，即y=|Tx|²，其中传输矩阵T的每个复值元素至多可训练，这使得每个输出得分被限制为具有秩1决策矩阵A=t†t的二次型x†Ax。本文提出随机映射强度光学神经网络（RMI-ONN）作为全光两层计算网络，突破了该秩1限制。数值实验表明，无序介质的高维特征投影、可编程非负强度掩模与分段空间功率求和可通过更高秩二次决策边界的表达能力超越秩1上限。此外，利用无序介质的矢量相干波混特性，本文在MNIST、Fashion-MNIST和Quick Draw数据集上，于跨所有编码域的单一光学配置下，通过RMI-ONN实验验证了振幅、相位和偏振的多模态分类。这些结果为可扩展的直接光场处理器提供了实践和概念基础，该处理器能在统一的基于强度的推理框架内利用振幅、相位和偏振信息。
英文摘要
Free-space optical neural networks offer distinct advantages for computational imaging and machine vision because they can compute directly on incident optical fields. However, conventional ONNs composed of cascaded linear optical components are bound to a general linear input-output relation with square-law detection between the input field $\mathbf{x}$ and output score $\mathbf{y}$, $\mathbf{y}=|\mathbf{T}\mathbf{x}|^2$, where at best every complex-valued element of the transmission matrix $\mathbf{T}$ is trainable. This restricts each output score to a quadratic form $\mathbf{x}^\dagger\mathbf{A}\mathbf{x}$ with rank-one decision matrix $\mathbf{A}=\mathbf{t}^\dagger\mathbf{t}$. Here, we present a random-mapped intensity optical neural network (RMI-ONN) as an all-optical two-layer computational network that lifts this rank-one limit. We numerically demonstrate that a high-dimensional feature projection by a disordered medium, a programmable nonnegative intensity mask, and segmented spatial power summation together can surpass the rank-one ceiling through the expressivity of higher-rank quadratic decision boundaries. Furthermore, exploiting the vectorial coherent wave-mixing nature of the disordered medium, we experimentally validate multimodal classification of amplitude, phase, and polarization on MNIST, Fashion-MNIST, and Quick Draw with the RMI-ONN, under a single optical configuration across all encoding domains. These results provide a practical and conceptual basis for scalable direct-field optical processors capable of exploiting amplitude, phase, and polarization information within a unified intensity-based inference framework.
