Concept mask-aware pruning and augmentation for few sample model compression

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-sample model compression is crucial for efficient deep learning with limited data.
• Existing methods often overlook pruning configuration impacts and sample diversity for performance recovery.
• Deep learning interpretability reveals the significance of feature map regions linked to filters.

Purpose of the Study:

• To propose a novel few-sample model compression method addressing limitations of prior approaches.
• To introduce concept masks for guiding pruning configuration and enhancing sample diversity.
• To unify model compression and performance recovery within a flexible framework.

Main Methods:

• Concept mask-aware (CMa) pruning: Determines block-wise pruning by measuring concept mask divergence.
• CMaMix augmentation: Mixes salient image regions using concept masks from uncompressed models to boost sample diversity.
• Flexible pruning granularity: Supports both filter-level and block-level compression.

Main Results:

• Achieved new state-of-the-art results on image classification and semantic segmentation tasks.
• Demonstrated significant accuracy improvements (0.2% to 8.1%) over prior works with only 50 samples.
• Successfully unified model compression and performance recovery, showcasing effectiveness across diverse architectures.

Conclusions:

• The proposed CMa pruning and CMaMix augmentation offer a powerful approach for few-sample model compression.
• Leveraging concept masks and sample diversity effectively recovers performance in compressed models.
• This method provides a flexible and state-of-the-art solution for efficient deep learning with minimal data.