Hardware–Software Co-design Computational Framework and Hardware-Aware Training for Photonic Spiking Convolutional Networks with DFB-SA Laser

Abstract

Neuromorphic computing stands out as a highly competitive computing paradigm capable of overcoming the bottlenecks inherent in von Neumann architectures. The spiking convolutional neural network (SCNN) is a prominent type of model within the realm of neuromorphic computing. Adapting SCNNs to photonic neuromorphic hardware holds great promise for significantly increasing the computation speed and fully leveraging its low energy consumption. In this paper, we develop an end-to-end design framework of photonic SCNN. We present the design of a SCNN tailored for a photonic platform utilizing distributed feedback lasers with a saturable absorber (DFB-SA laser) as the photonic spiking neurons. We also introduce hardware implementations for key computational steps in photonic SCNNs, including the nonlinear activation function, the convolutional layer, the fully connected layer, and the max-pooling layer. Additionally, a hardware-aware training method is proposed. Furthermore, we apply the designed network to classify the MNIST, Fashion-MNIST, and CIFAR-10 datasets, achieving accuracies of 96.52%, 90.48%, and 88.45%, respectively, in simulations on the test sets. And we experimentally validate the nonlinear activation function in the MNIST dataset classification task using the DFB-SA laser, achieving a classification accuracy of 96.06%. This study introduces a novel approach to deploying neural networks on hardware, presenting a portable, modular hardware simulation model and training method. This model is anticipated to be seamlessly integrated into the future hardware–software co-design of large-scale photonic SCNNs. Part of the hardware-aware training code is available at https://github.com/Oo-Fish-oO/Hardware-aware-training