Amiri Samaneh, Miri Mehdi
Department of Communications and Electronics, School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran.
Sci Rep. 2025 Jul 1;15(1):22055. doi: 10.1038/s41598-025-06259-4.
This paper presents a design for an integrated all-optical convolution neural network in which all three network layers i.e. convolution, max-pooling and fully connected can be implemented in silicon photonics platform with the use of GST-based active waveguides. In the convolution layer the need for the ReLU is mitigated by using positive kernel values. Also, the designed max-pooling layer is completely based on silicon photonics elements and requires no electrical or electro-optical logical element. These significantly reduce the overall network complexity. The individual optical elements, network layers and the overall convolution network are simulated using finite-difference time-domain method, coupled mode theory and Python programming, respectively. The network performance is evaluated in MNIST data classification and also signal modulation identification (RML2016.10a dataset) which showed accuracies of 91.90% and 80%, respectively. These are comparable to the highest-reported values for the convolution networks while the proposed network has a less complex design and can be completely implemented with integrated optical elements.
本文提出了一种集成全光卷积神经网络的设计方案，其中卷积、最大池化和全连接这三个网络层都可以在硅光子平台上利用基于 GST 的有源波导来实现。在卷积层中，通过使用正核值减轻了对 ReLU 的需求。此外，所设计的最大池化层完全基于硅光子元件，不需要电或电光逻辑元件。这些显著降低了整体网络复杂度。分别使用时域有限差分法、耦合模理论和 Python 编程对各个光学元件、网络层和整体卷积网络进行了模拟。在 MNIST 数据分类和信号调制识别（RML2016.10a 数据集）中对网络性能进行了评估，结果分别显示准确率为 91.90%和 80%。这些与卷积网络报告的最高值相当，而所提出的网络设计更简单，并且可以完全用集成光学元件实现。
