Selective Coupling of Decoupled Informative Regions: Masked Attention Alignment for Data-Free Quantization of Vision Transformers
Researchers have developed a new method called Masked Attention Alignment (MaskAQ) for data-free quantization of Vision Transformers. This technique identifies and focuses on the most informative regions within image patches, which are crucial for the self-attention mechanism. MaskAQ aligns these key regions between full-precision and quantized models, improving the quality of synthesized data and enhancing quantization performance across various tasks. AI
IMPACT Enhances efficiency of Vision Transformers by improving data-free quantization techniques.