Scholarlink

<p>Spiking neural networks (SNNs) offer event-driven computation and promising energy efficiency for neuromorphic systems. However, the structural evolution from conventional artificial neural networks (ANNs) to fully event-driven SNNs remains conceptually fragmented, especially in the rapidly growing literature on Leaky Integrate-and-Fire (LIF)-based, directly trained SNNs. This review proposes a unified four-stage structural evolution framework that interprets directly trained SNN architectures as progressively incorporating: (1) Binary ANN; (2) Temporal Dimension Introduction; (3) Temporal Accumulation Mechanism; and (4) Reset and Sparsity Control. By mapping diverse “neuron modifications” onto these fundamental computational components, the framework provides consistent operational terminology, enables stage-wise comparison of methods, and clarifies their performance trade-offs. We further present representative solutions and open challenges at each stage, offering an ANN oriented pathway for understanding and designing modern LIF-SNNs.</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

A Four-Stage Structural Evolution Framework for Spiking Neural Networks: A Review and Perspective from Binary ANN to Event-Driven Models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月7日

查看全文
  • Huaxu He,
  • Shuchao Gao

摘要

Spiking neural networks (SNNs) offer event-driven computation and promising energy efficiency for neuromorphic systems. However, the structural evolution from conventional artificial neural networks (ANNs) to fully event-driven SNNs remains conceptually fragmented, especially in the rapidly growing literature on Leaky Integrate-and-Fire (LIF)-based, directly trained SNNs. This review proposes a unified four-stage structural evolution framework that interprets directly trained SNN architectures as progressively incorporating: (1) Binary ANN; (2) Temporal Dimension Introduction; (3) Temporal Accumulation Mechanism; and (4) Reset and Sparsity Control. By mapping diverse “neuron modifications” onto these fundamental computational components, the framework provides consistent operational terminology, enables stage-wise comparison of methods, and clarifies their performance trade-offs. We further present representative solutions and open challenges at each stage, offering an ANN oriented pathway for understanding and designing modern LIF-SNNs.