Burst Spiking Neural Networks

This paper proposes Burst Spiking Neural Networks (BuSNNs) to improve robustness without giving up the energy advantages of spiking models. Its Burst-enhanced Spiking Neurons (BSNs) provide graded burst-based activations intended to reduce abrupt state changes caused by small input perturbations. A Dynamic Weight Constraint (DWC) mechanism penalizes connection weights according to activation states, reducing sensitivity while preserving accuracy. According to the abstract, BuSNNs outperform SNN and ANN counterparts on CIFAR-10. On ImageNet with an MS ResNet-34 backbone, they improve top-1 accuracy by 3.18 percentage points and corruption robustness by 2.66 points over the corresponding SNN baseline, surpassing 4-bit activation-quantized ANNs and approaching 8-bit ANN baselines.