AI 中文总结
本文以随机激光网络为物理视觉系统，通过图论指标建立物理与性能的关联，用进化算法优化网络拓扑，实现3000倍加速，提升了图像分类准确率，可迁移至其他物理学习系统。
AI 中文摘要
物理神经网络通过物质的本征非线性动力学实现学习，其设计优化面临重大挑战：复杂的多体物理可提供强大计算能力，但模拟成本高昂，大规模实验优化运行的制备也不切实际，因此无法有效直接搜索可能网络拓扑的高维空间。本文展示，该搜索可在抽象图空间中高效完成，其探索成本大幅降低。以随机激光网络（由互联的纳米波导构成，承载强耦合激光模式）作为典型物理视觉系统，建立了三层定量关联：简单图论指标可预测非线性激光物理，而后者又可预测视觉性能。在通过物理模拟验证该关系后，本文利用其驱动基于图指标的进化算法，生成的网络拓扑性能优于随机设计，计算成本仅为后者的极小部分（与物理模拟相比速度提升3000倍）。在模拟图像分类任务中，经图优化的网络大幅提升了分类准确率。由于本文框架作用于网络拓扑而非特定基底物理，预计其可迁移至其他基于网络的物理学习系统，为复杂强相互作用物理神经网络的定向设计与优化提供高效途径。
英文摘要
Physical neural networks perform learning through the intrinsic nonlinear dynamics of matter. Optimising their design presents a considerable challenge: complex many-body physics can provide powerful computation, but are expensive to simulate and large experimental optimisation runs are impractical to fabricate. Hence, the high-dimensional space of possible network topologies cannot be effectively directly searched. Here, we show that this search can be efficiently performed in an abstract graph space that is vastly cheaper to explore. Using random lasing networks -- composed of interconnected nanoscale waveguides and hosting strongly coupled lasing modes -- as an exemplar physical vision system, we establish a quantitative three-layer link: simple graph-theoretic metrics predict the nonlinear lasing physics, which in turn predicts vision performance. After validating this relationship using physical simulations, we exploit it to drive an evolutionary algorithm using graph metrics, producing network topologies that outperform random designs at a fraction of the computational cost (3000$\times$ speed-up compared to physical simulation). On simulated image-classification tasks, graph-optimised networks substantially improve classification accuracy. As our framework operates on network topology rather than substrate-specific physics, we anticipate it can transfer to other network-based physical learning systems, providing an efficient route for the directed design and optimisation of complex, strongly-interacting physical neural networks.
