ABSTRACT Photonic spiking neural networks (SNNs) hold great promise for high‐speed and energy‐efficient computing by integrating the advantages of photonics and neuromorphic computation. However, conventional photonic SNNs are limited by device properties and can only implement algorithms with non‐negative weights. In this work, we propose a photonic SNN computing architecture based on the intrinsic plasticity of the distributed feedback semiconductor laser with saturable absorber, enabling the implementation of spiking convolutional networks with both positive and negative weights. We experimentally demonstrate the feasibility of the multiply–accumulate operations with this architecture and apply it to classify neuromorphic datasets, including DVS128 Gesture, N‐MNIST, and CIFAR10‐DVS, in simulations. To facilitate hardware deployment, network weights are quantized during training. Simulation results show that the hardware model achieves accuracies of 89.58% (DVS128 Gesture), 99.06% (N‐MNIST), and 68.70% (CIFAR10‐DVS) under 8‐bit quantization—comparable to the software baseline. This work contributes to the development of integrated photonic neuromorphic systems that bridge sensing and computing.
Yu et al. (2026) studied this question.