ABSTRACT Spiking neural networks (SNNs)‐based radar gesture recognition has gained significant attention for its ability to enhance computational efficiency. To better exploit their potential, SNNs are often integrated with effective compression techniques to enable practical deployment in resource‐constrained edge environments. However, conventional pruning approaches either necessitate computationally intensive pre‐training or rely on gradient‐dependent mechanisms, which are incompatible with SNN's event‐driven nature. In this paper, we propose a spike timing–driven pruning‐before‐training method (ST‐PBT) for SNNs, which introduces a gradient‐independent pruning criterion derived from spike timing correlations within the STDP rule. By leveraging the inherent spatio‐temporal dynamics of SNNs, it enables a more accurate assessment of connection importance than conventional magnitude‐based approaches. Furthermore, our method performs one‐shot pruning during the model initialisation, thereby avoiding the computational cost incurred in the model pre‐training. We evaluate the ST‐PBT on a radar‐ based human gesture dataset and compare it against several pruning baselines. Experimental results demonstrate that the ST‐PBT achieves a 10× reduction in pruning time and computational cost while maintaining superior performance, thereby validating the biologically inspired ST‐PBT's efficacy in resource‐constrained radar gesture recognition.
Zhang et al. (2026) studied this question.