Scholarlink

<p>Deploying deep convolutional neural networks (CNNs) on embedded or mobile devices is challenging due to their significant storage demands and high computational cost. To address this issue, we propose an efficient channel pruning method based on a sparse learning strategy that reduces computational complexity while preserving high performance. To effectively identify and remove less important channels, we formulate the problem as a bi-objective optimization task to explore the trade-off between model complexity and accuracy. Our approach searches for the optimal combination of scaling and shifting factors in batch normalization layers, which determine the importance of each channel, enabling structured pruning. Subsequently, redundant and insignificant channels are pruned without significantly impacting accuracy. A smoothed <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10586_2025_5504_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(L_0\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mn>0</mn> </msub> </math></EquationSource> </InlineEquation> regularization is introduced into the model training to achieve channel-level sparsity. The results of applying this method to modern models across various benchmark datasets demonstrate its effectiveness compared to state-of-the-art.</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

Multiobjective pruning for sparse deep neural networks: leveraging \(L_0\) regularization on batch normalization scaling factors

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Khalid Elghazi,
  • Mostafa Bakhouya,
  • Hassan Ramchoun,
  • Tawfik Masrour

摘要

Deploying deep convolutional neural networks (CNNs) on embedded or mobile devices is challenging due to their significant storage demands and high computational cost. To address this issue, we propose an efficient channel pruning method based on a sparse learning strategy that reduces computational complexity while preserving high performance. To effectively identify and remove less important channels, we formulate the problem as a bi-objective optimization task to explore the trade-off between model complexity and accuracy. Our approach searches for the optimal combination of scaling and shifting factors in batch normalization layers, which determine the importance of each channel, enabling structured pruning. Subsequently, redundant and insignificant channels are pruned without significantly impacting accuracy. A smoothed \(L_0\) L 0 regularization is introduced into the model training to achieve channel-level sparsity. The results of applying this method to modern models across various benchmark datasets demonstrate its effectiveness compared to state-of-the-art.