(☆ Award Candidate)

Authors: Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, Guoqi Li

Research Introduction:

A long-standing and difficult challenge in the field of spiking neural networks (SNNs) is achieving competitive performance on large-scale, complex detection tasks. The proposed SpikeYOLO significantly bridges the performance gap between SNNs and ANNs in object detection through two main contributions. First, in terms of network architecture, the team found that overly complex architectures in SNNs lead to spike degradation, so they simplified the design (Figure 1). Second, for spiking neuron design, they proposed a new spiking neuron that uses integer-valued training and spike-driven inference, effectively reducing quantization error while preserving spike-driven computation characteristics (Figure 2). Experimental results show that this approach greatly improves task performance while maintaining low power consumption. On the static COCO dataset, SpikeYOLO achieved an mAP of 48.9%, 18.7% higher than the previous SOTA in the SNN field. On the neuromorphic Gen1 dataset, SpikeYOLO outperformed an ANN of the same architecture by 2.7% mAP with 5.7x better energy efficiency.

This research demonstrates the broad application potential of SNNs in ultra-low-power edge vision scenarios. Currently, the team is not only studying SNN applications in more typical edge vision use cases but is also conducting hardware simulation and design work for general-purpose SNN architectures. This work explores the performance/efficiency potential of SNNs at the algorithmic level, proving their future viability as low-power alternatives to conventional ANNs and providing important guidance for next-generation neuromorphic algorithms and chips.

Figure 1. SpikeYOLO Architecture Design

Figure 2. I-LIF Neuron