BuSNN boosts SNN accuracy by 3.18% and robustness on ImageNet

Spiking Neural Networks (SNNs) promise ultra-low power consumption but historically lag behind Artificial Neural Networks (ANNs) in accuracy—and, as this new paper highlights, in robustness. The authors identify two root causes: binary spiking activations cause large state changes under small input perturbations, and the lack of effective weight constraints makes outputs overly sensitive. To solve this, they introduce BuSNNs built on two innovations. First, Burst-enhanced Spiking Neurons (BSNs) replace single spikes with graded burst firing, smoothing activation transitions. Second, a Dynamic Weight Constraint (DWC) penalizes connection weights based on activation states, shrinking magnitudes without sacrificing accuracy. Theoretical analysis backs the robustness gains.

Experimental results are compelling. On CIFAR-10, BuSNNs outperform both SNN and ANN counterparts in accuracy and robustness. Scaling to ImageNet with the MS ResNet-34 backbone, BuSNN achieves a 3.18% top-1 accuracy improvement and a 2.66% boost in corruption robustness over the baseline SNN. Remarkably, despite using binary spikes, BuSNN surpasses 4-bit activation-quantized ANN baselines and comes close to 8-bit ANN performance—all while maintaining the inherent energy efficiency of SNNs. This work positions BuSNNs as a practical path toward robust, low-power neural networks for real-world deployment.