Filter Pruning via Learned Representation Median in the Frequency Domain.

Area of Science:

• Deep Learning
• Computer Vision
• Network Optimization

Background:

• Existing filter pruning methods often focus on spatial domain importance, which may not generalize well.
• Redundancy and low-frequency representations, crucial for visual tasks, are better analyzed in the frequency domain.

Purpose of the Study:

• To propose a novel filter pruning method that operates in the frequency domain.
• To identify and remove absolutely unimportant filters based on learned representation median (LRMF).
• To avoid the computationally expensive fine-tuning process typically required after pruning.

Main Methods:

• Calculating the learned representation median (RM) in the frequency domain using discrete cosine transform (DCT).
• Pruning filters identified as absolutely unimportant in the frequency domain.
• Applying the method to various deep learning architectures like ResNet and VGG.

Main Results:

• LRMF achieves significant reduction in Floating Point Operations (FLOPs) across multiple datasets and networks.
• Demonstrates improved or comparable Top-1 accuracy compared to state-of-the-art pruning techniques.
• Achieved 52.3% FLOPs reduction on ResNet110 (CIFAR-10) with a 0.04% accuracy increase and 35.9% FLOPs reduction on VGG16 (CIFAR-100) with a 0.5% accuracy increase.

Conclusions:

• The proposed LRMF method offers an effective and efficient approach to filter pruning in deep learning.
• Frequency domain analysis provides a more robust criterion for filter importance than spatial domain methods.
• LRMF enables significant model compression and acceleration without compromising, and sometimes improving, accuracy.