Photonic Neuromorphic Accelerator Based on Multiplication-Free Convolutional Spiking Neural Network

Spiking neural networks (SNNs), widely recognized as the third generation of neural networks, transmit information via binary spikes. Neuromorphic hardware typically prioritizes multiplication-free operations for efficient computing. Here, we propose a photonic neuromorphic accelerator based on a multiplication-free binary convolutional spiking neural network. It maps input data and convolutional kernels into spatiotemporal spikes through spatiotemporal encoding, leveraging the refractory period and threshold dynamics of photonic spiking neurons (DFB-SA lasers) for reshaping spikes and XNOR-like operations. Additionally, it utilizes optical power meters as integrators to process convolution results. The results show that convolution examples can be performed successfully. Moreover, a spiking neural network with binary weights achieved recognition accuracies of 96.6% on the MNIST dataset and 91.8% on the Fashion-MNIST dataset. In addition to decoupling the system from constraints imposed by hardware scale, the photonic convolutional accelerator also lessens the need for high-precision digital-to-analog converters and high-speed analog-to-digital converters. Boasting low-power consumption and high operational efficiency, this engineered accelerator is well-suited for edge computing scenarios, such as enabling real-time image identification in intelligent monitoring systems and facilitating low-latency feature extraction in wearable smart gadgets.