Advanced Training Algorithms in Sigma-Delta Spiking YOLO for Energy-Efficient Object Detection on Neuromorphic Hardware

This paper explores methods for improving the training and inference performance of a Sigma-Delta spiking object detector by applying a series of training techniques and optimization strategies. Conventional CNN-based YOLO models, while effective in terms of accuracy, are generally unsuitable for deployment in low-power or real-time edge environments due to their computational demands. Spiking neural networks (SNNs), especially those employing Sigma-Delta neurons, offer a promising alternative with sparse, event-driven computation that reduces energy use. To this end, we enhance a Sigma-Delta spiking YOLOv3 framework by adapting several well-established training methods from both the SNN and CNN domains. These include Cosine-Annealing Learning-Rate Scheduling with Warm Restarts (SGDR), Multi-Gaussian surrogate gradients (MG-SGD), and a novel Epoch-Based Dynamic Surrogate Gradient Descent (ED-SGD) algorithm that we have devised. These additions are intended to improve both convergence behavior and overall model performance in spike-driven training. We evaluate the enhanced model using multiple real-world datasets, including BDD100K, Udacity Self-Driving, and SoccerNet. The experimental results show consistent improvements in detection accuracy and energy efficiency compared to both the conventional CNN-based YOLOv3 and the spiking baseline. Finally, we deploy the trained model on Intel’s Loihi 2 neuromorphic platform, confirming its feasibility for real-time, low-power operation.