.png){kind=icon}
Return
Network slimming using Lp(p < 1) regularization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.111711.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Convolutional neural networks (CNNs) require significant computational resources, making them challenging to deploy on battery-powered devices. Channel pruning provides a practical solution by reducing model size to facilitate efficient deployment. Among various pruning methods, L-0 regularization is an effective approach for identifying sparse sub-networks. However, solving it directly is NP-hard. To overcome this challenge, we approximate the L-0 norm using L-p regularization with p < 1, and employ the Alternating Direction Method of Multipliers (ADMM) to solve the corresponding optimization problem. Since the L-p regularization term introduces a non-convex subproblem, we linearize it to transform the subproblem into a convex optimization problem. Nevertheless, when the model parameter approaches zero, the L-p norm grows excessively, leading to no convergence in ADMM. To address this, we introduce a synaptic inhibition mechanism to suppress the excessive growth of the L-p regularization term, ensuring ADMM algorithm convergence. Experiments on CIFAR, ImageNet, and GLUE benchmark dataset demonstrate that our method achieves state-of-the-art performance in channel pruning.
Keywords:
Channel pruning
Alternating direction method of multipliers
Efficient inference
L-p(p < 1) regularization
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveShareSave
ShareSaveShareSave
View more