ARPruning: An automatic channel pruning based on attention map ranking.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Structured pruning is key for compressing Convolutional Neural Networks (CNNs).
• Existing methods often lack explainability and can lead to suboptimal compression by relying solely on weight magnitudes.
• Determining optimal pruning ratios across layers requires sophisticated search strategies.

Purpose of the Study:

• To develop an explainable and effective structured pruning method for CNNs.
• To introduce a new criterion for assessing intra-layer channel importance using attention maps.
• To design an efficient search algorithm for optimal inter-layer pruning ratios.

Main Methods:

• ARPruning utilizes attention maps to create an explainable intra-layer channel importance criterion.
• A local neighborhood search algorithm is developed for determining optimal inter-layer pruning ratios.
• The method integrates pruning criteria with an automatic search strategy for efficient optimization.

Main Results:

• ARPruning achieves high compression rates while maintaining excellent model accuracy.
• Experimental results demonstrate superior compression performance compared to state-of-the-art pruning methods.
• The proposed method offers a more interpretable approach to structured pruning.

Conclusions:

• ARPruning provides an effective and explainable solution for CNN model compression.
• The attention-map-based criterion and search strategy significantly improve pruning efficiency and performance.
• This work advances the field of model compression by offering a robust and accurate structured pruning technique.