Scholarlink

<p>The deployment of neural networks on resource-constrained devices is fundamentally limited by their inherent computational and memory complexity. While unstructured pruning has proven effective in compressing models, the automatic optimization of per-layer pruning ratios remains a critical challenge. This paper proposes an optimization framework based on an Improved Sparrow Search Algorithm (ISSA). Specifically, ISSA restructures the population into a dual-role “discoverer–joiner” architecture and integrates an Enhanced Lens Imaging Opposition-based Learning (ELLO) mechanism to dynamically modulate exploration intensity. Additionally, a small-step stochastic perturbation strategy is introduced in the late search phase to balance global exploration with local exploitation. Extensive experiments on eight benchmark functions demonstrate that ISSA reduces the average solution error by 3–12 orders of magnitude compared to the original SSA. Preliminary application to MNIST-LeNet-5 achieves an 84.81% parameter reduction with only a 0.59% accuracy drop. Crucially, strictly validating on modern architectures (CIFAR-10) reveals ISSA’s superior robustness: for ResNet-20, it achieves 86.70% accuracy at ~ 11.5% retention, outperforming the Uniform baseline; for VGG-16 under extreme compression (&gt; 88%), ISSA maintains 86.21% accuracy while the standard Global Magnitude Pruning (GlobalMP) baseline collapses to random guessing (10.00%). Further generalization experiments on three engineering datasets—concrete strength, airfoil aerodynamics, and building energy efficiency—show that ISSA adaptively allocates layer-wise pruning ratios, ensuring high accuracy while enabling efficient model compression. This study provides a novel and automated approach to model compression for deployment on resource-limited devices.</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

Layer-by-layer unstructured pruning rate optimization based on improved sparrow search algorithm

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Youyu Liu,
  • Xiangxiang Zhou,
  • Jianwen Zhang,
  • Zhouyi Gu,
  • Qiang Wang

摘要

The deployment of neural networks on resource-constrained devices is fundamentally limited by their inherent computational and memory complexity. While unstructured pruning has proven effective in compressing models, the automatic optimization of per-layer pruning ratios remains a critical challenge. This paper proposes an optimization framework based on an Improved Sparrow Search Algorithm (ISSA). Specifically, ISSA restructures the population into a dual-role “discoverer–joiner” architecture and integrates an Enhanced Lens Imaging Opposition-based Learning (ELLO) mechanism to dynamically modulate exploration intensity. Additionally, a small-step stochastic perturbation strategy is introduced in the late search phase to balance global exploration with local exploitation. Extensive experiments on eight benchmark functions demonstrate that ISSA reduces the average solution error by 3–12 orders of magnitude compared to the original SSA. Preliminary application to MNIST-LeNet-5 achieves an 84.81% parameter reduction with only a 0.59% accuracy drop. Crucially, strictly validating on modern architectures (CIFAR-10) reveals ISSA’s superior robustness: for ResNet-20, it achieves 86.70% accuracy at ~ 11.5% retention, outperforming the Uniform baseline; for VGG-16 under extreme compression (> 88%), ISSA maintains 86.21% accuracy while the standard Global Magnitude Pruning (GlobalMP) baseline collapses to random guessing (10.00%). Further generalization experiments on three engineering datasets—concrete strength, airfoil aerodynamics, and building energy efficiency—show that ISSA adaptively allocates layer-wise pruning ratios, ensuring high accuracy while enabling efficient model compression. This study provides a novel and automated approach to model compression for deployment on resource-limited devices.