Scholarlink

<p>Network pruning provides a promising approach for deploying costly Deep Neural Network&#xa0;(DNN) models on resource-constrained devices. However, most existing pruning works focus on compressing traditional convolutional DNN models for image-related classification and object-detection tasks in 2D scenarios. Due to the complicated structure of 3D CNN models, pruning such models has not been well studied. In this paper, we analyze the different properties between 2D and 3D tasks, then propose a Filter/Depth-wise Independence Score&#xa0;(FDIS) to evaluate the importance of each filter. In addition, we adopt several granularity schemes to improve the performance of the proposed method. To achieve fine-grained pruning, we prune the networks gradually using an iterative pruning procedure. In addition, we experimentally show that weights with low independence scores contain less important information, enabling the removal of filters without serious accuracy degradation. Our proposed FDIS-based approach maintains high accuracy with certain FLOP reduction and practical acceleration.</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

Pruning 3D Convolutional Neural Networks via Channel Independence

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月6日

查看全文
  • Yang Sui,
  • Khizar Anjum,
  • Dario Pompili,
  • Bo Yuan

摘要

Network pruning provides a promising approach for deploying costly Deep Neural Network (DNN) models on resource-constrained devices. However, most existing pruning works focus on compressing traditional convolutional DNN models for image-related classification and object-detection tasks in 2D scenarios. Due to the complicated structure of 3D CNN models, pruning such models has not been well studied. In this paper, we analyze the different properties between 2D and 3D tasks, then propose a Filter/Depth-wise Independence Score (FDIS) to evaluate the importance of each filter. In addition, we adopt several granularity schemes to improve the performance of the proposed method. To achieve fine-grained pruning, we prune the networks gradually using an iterative pruning procedure. In addition, we experimentally show that weights with low independence scores contain less important information, enabling the removal of filters without serious accuracy degradation. Our proposed FDIS-based approach maintains high accuracy with certain FLOP reduction and practical acceleration.