Filter Pruning Based on Information Capacity and Independence.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are widely used but computationally intensive.
• Existing filter pruning methods face challenges with biased selection and high computational costs.
• Efficient CNN compression is crucial for practical deployment.

Purpose of the Study:

• To introduce a novel, interpretable, and lightweight filter pruning method for CNNs.
• To address the limitations of existing filter pruning techniques regarding bias and computational overhead.
• To improve the efficiency of CNNs without significant performance degradation.

Main Methods:

• Developed a multiperspective filter evaluation approach.
• Introduced 'information capacity' using interpretable entropy to measure individual filter importance.
• Designed 'information independence' to assess correlations among filters.
• Implemented a feature-guided approximation for efficient metric calculation.

Main Results:

• Achieved significant reduction in computational cost and model size on benchmark datasets.
• Demonstrated superior performance compared to state-of-the-art methods.
• For ResNet-50 on ILSVRC-2012, reduced FLOPs by 77.4% and parameters by 69.3% with only a 2.64% accuracy drop.

Conclusions:

• The proposed filter pruning method is effective, efficient, and practical.
• It offers a lightweight and interpretable approach to CNN compression.
• The method significantly enhances CNN performance while maintaining high accuracy.