Scholarlink

Spiking Neural Networks (SNN) offer significant potential for energy-efficient edge computing due to their bio-inspired, event-driven mechanisms. However, hardware implementations employing approximate computing techniques, such as approximate adders, introduce nonlinear error accumulation across spatial and temporal dimensions, degrading membrane potential dynamics and spike timing reliability. Existing approaches often overlook real-world data heterogeneity and lack systematic error modeling tailored to SNN-specific mechanisms. To address these challenges, this work proposes a hardware-algorithm co-designed optimization framework that dynamically models error propagation and implements compensation strategies. By quantifying layer-wise error scaling coefficients and introducing a global error metric ( \(E_{\text {global}}\) ), the framework establishes a strong correlation ( \(|r| >0.74\) ) between \(E_{\text {global}}\) and task accuracy, enabling robust deployment in high-noise scenarios. Experimental results on MNIST demonstrate that the framework achieves up to 29% energy efficiency improvement while maintaining classification accuracy ( \(<1\%\) degradation). This study provides a systematic methodology for co-optimizing energy efficiency and error resilience in SNN, advancing their practical applicability in edge computing.

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

An Energy Efficient Optimization Framework for Spiking Neural Networks Using Error-Resilient Approximation Techniques

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Weikang Xu,
  • Yan Sun,
  • Jianmin Zhang,
  • Zhiqiang Wen,
  • Yu Ma

摘要

Spiking Neural Networks (SNN) offer significant potential for energy-efficient edge computing due to their bio-inspired, event-driven mechanisms. However, hardware implementations employing approximate computing techniques, such as approximate adders, introduce nonlinear error accumulation across spatial and temporal dimensions, degrading membrane potential dynamics and spike timing reliability. Existing approaches often overlook real-world data heterogeneity and lack systematic error modeling tailored to SNN-specific mechanisms. To address these challenges, this work proposes a hardware-algorithm co-designed optimization framework that dynamically models error propagation and implements compensation strategies. By quantifying layer-wise error scaling coefficients and introducing a global error metric ( \(E_{\text {global}}\) ), the framework establishes a strong correlation ( \(|r| >0.74\) ) between \(E_{\text {global}}\) and task accuracy, enabling robust deployment in high-noise scenarios. Experimental results on MNIST demonstrate that the framework achieves up to 29% energy efficiency improvement while maintaining classification accuracy ( \(<1\%\) degradation). This study provides a systematic methodology for co-optimizing energy efficiency and error resilience in SNN, advancing their practical applicability in edge computing.