Platform Photonics Research Center, National Institute of Advanced Industrial Science and Technology (AIST), 16-1, Onogawa, Tsukuba, Ibaraki, 305-8569, Japan.
NTT Basic Research labs., 3-1, Morinosato Wakamiya, Atsugi-shi, Kanagawa, 243-0198, Japan.
Nat Commun. 2022 Jun 30;13(1):3261. doi: 10.1038/s41467-022-30906-3.
On-chip training remains a challenging issue for photonic devices to implement machine learning algorithms. Most demonstrations only implement inference in photonics for offline-trained neural network models. On the other hand, artificial neural networks are one of the most deployed algorithms, while other machine learning algorithms such as supporting vector machine (SVM) remain unexplored in photonics. Here, inspired by SVM, we propose to implement projection-based classification principle by constructing nonlinear mapping functions in silicon photonic circuits and experimentally demonstrate on-chip bacterial foraging training for this principle to realize single Boolean logics, combinational Boolean logics, and Iris classification with ~96.7 - 98.3 per cent accuracy. This approach can offer comparable performances to artificial neural networks for various benchmarks even with smaller scales and without leveraging traditional activation functions, showing scalability advantage. Natural-intelligence-inspired bacterial foraging offers efficient and robust on-chip training, and this work paves a way for photonic circuits to perform nonlinear classification.
在片上训练仍然是光子器件实现机器学习算法的一个具有挑战性的问题。大多数演示仅在光子学中实现用于离线训练的神经网络模型的推断。另一方面，人工神经网络是部署最多的算法之一，而其他机器学习算法（如支持向量机（SVM））在光子学中尚未得到探索。在这里，受 SVM 的启发，我们通过在硅光子电路中构建非线性映射函数，提出了实现基于投影的分类原理，并通过实验演示了该原理的细菌觅食训练，从而实现了单布尔逻辑、组合布尔逻辑和 Iris 分类，准确率约为 96.7% - 98.3%。与传统的激活函数相比，这种方法即使在规模较小的情况下，也能为各种基准提供与人工神经网络相当的性能，显示出可扩展性优势。受自然智能启发的细菌觅食提供了高效和强大的片上训练，这项工作为光子电路执行非线性分类铺平了道路。
