Scholarlink

<p>With the growing scale and complexity of convolutional neural networks (CNNs), model pruning has experienced rapid development. However, most existing methods use fixed importance criteria across the entire network, lacking adaptability and flexibility in selecting channels across different layers. To address the above issues, this paper proposes a dynamic pruning method based on reinforcement learning (RL), which utilizes a designed dual-action agent and an iterative weight reconstruction mechanism to allocate sparsity and select channel importance for different layers of the network. First, a weight reconstruction mechanism is introduced to recover the accuracy of the pruned sub-network and stabilize the RL training process. Second, since weight updates in the compressed network induce a non-stationary environment, this work models the state of compressed networks and reconstructs the agent’s policy learning. Extensive experiments using ResNets and MobileNets on CIFAR-10/100 and ImageNet demonstrate that the proposed method achieves excellent performance in both accuracy and inference efficiency.</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

Sparsity allocation and channel importance selection through dual-action agent driven and weight reconstruction

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Rui Cai,
  • Yixin Zhao

摘要

With the growing scale and complexity of convolutional neural networks (CNNs), model pruning has experienced rapid development. However, most existing methods use fixed importance criteria across the entire network, lacking adaptability and flexibility in selecting channels across different layers. To address the above issues, this paper proposes a dynamic pruning method based on reinforcement learning (RL), which utilizes a designed dual-action agent and an iterative weight reconstruction mechanism to allocate sparsity and select channel importance for different layers of the network. First, a weight reconstruction mechanism is introduced to recover the accuracy of the pruned sub-network and stabilize the RL training process. Second, since weight updates in the compressed network induce a non-stationary environment, this work models the state of compressed networks and reconstructs the agent’s policy learning. Extensive experiments using ResNets and MobileNets on CIFAR-10/100 and ImageNet demonstrate that the proposed method achieves excellent performance in both accuracy and inference efficiency.