Scholarlink

<p>Spiking Neural Networks (SNNs) represent a promising paradigm for energy-efficient machine learning. However, achieving high performance under significant sparsity constraints remains challenging. In this work, we present Staged Scale-Free Topology Evolution (S2TE), a novel sparse training framework for SNNs that systematically combines layer-wise scale-free topology initialization with a conceptually inspired, three-stage weight pruning and regeneration process comprising exploratory growth, topology refinement, and stability maintenance. S2TE uses the Barabási-Albert model to establish sparse layer-wise scale-free topologies from the outset, and the staged dynamic topology optimization approach enables efficient balance between structural exploration and exploitation. Experiments on CIFAR10, CIFAR100, and DVS-CIFAR10 benchmarks show that S2TE-trained SNNs achieve competitive accuracy while maintaining merely 30% connectivity. The resulting networks exhibit layer-wise scale-free-like connectivity, which is associated with efficient and robust information flow in the evaluated benchmarks. These findings demonstrate the promise of staged topology evolution for improving sparse SNN training and motivate further study in broader scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

S2TE: Staged Scale-Free Topology Evolution for Sparse Spiking Neural Networks

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Zaipeng Xie,
  • Wei Zhu,
  • Peixin Li,
  • Haotian Ding,
  • WenZhan Song

摘要

Spiking Neural Networks (SNNs) represent a promising paradigm for energy-efficient machine learning. However, achieving high performance under significant sparsity constraints remains challenging. In this work, we present Staged Scale-Free Topology Evolution (S2TE), a novel sparse training framework for SNNs that systematically combines layer-wise scale-free topology initialization with a conceptually inspired, three-stage weight pruning and regeneration process comprising exploratory growth, topology refinement, and stability maintenance. S2TE uses the Barabási-Albert model to establish sparse layer-wise scale-free topologies from the outset, and the staged dynamic topology optimization approach enables efficient balance between structural exploration and exploitation. Experiments on CIFAR10, CIFAR100, and DVS-CIFAR10 benchmarks show that S2TE-trained SNNs achieve competitive accuracy while maintaining merely 30% connectivity. The resulting networks exhibit layer-wise scale-free-like connectivity, which is associated with efficient and robust information flow in the evaluated benchmarks. These findings demonstrate the promise of staged topology evolution for improving sparse SNN training and motivate further study in broader scenarios.