Summary of Mimiq: Low-bit Data-free Quantization Of Vision Transformers with Encouraging Inter-head Attention Similarity, by Kanghyun Choi et al.
MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity
by Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, Jinho Lee
First submitted to arxiv on: 29 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The abstract proposes a novel technique called MimiQ, which improves data-free quantization (DFQ) of vision transformer (ViT) architectures in low-bit settings. The existing DFQ methods fail to achieve efficacy due to misaligned attention maps in synthetic datasets. The authors identify the importance of aligning attention maps and devise MimiQ by generating synthetic data that aligns head-wise attention responses with spatial query patches, followed by structural attention distillation to align the quantized network’s attention maps with those of the full-precision teacher. Experimental results show that MimiQ outperforms baselines and sets a new state-of-the-art performance for DFQ ViT quantization. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary MimiQ is a new way to make vision transformer (ViT) networks smaller without losing accuracy. This technique, called data-free quantization (DFQ), creates a lightweight network from its full-precision counterpart using synthetic data. The problem with current DFQ methods for ViTs is that they don’t work well in low-bit settings because their synthetic data have misaligned attention maps. MimiQ fixes this by generating synthetic data that aligns head-wise attention responses with spatial query patches, and then adjusts the quantized network’s attention maps to match those of its full-precision teacher. The results show that MimiQ performs better than previous methods. |
Keywords
» Artificial intelligence » Attention » Distillation » Precision » Quantization » Synthetic data » Vision transformer » Vit
