Summary of Comq: a Backpropagation-free Algorithm For Post-training Quantization, by Aozhong Zhang et al.
COMQ: A Backpropagation-Free Algorithm for Post-Training Quantization
by Aozhong Zhang, Zi Yang, Naigang Wang, Yingyong Qi, Jack Xin, Xin Li, Penghang Yin
First submitted to arxiv on: 11 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 A novel post-training quantization (PTQ) algorithm called COMQ is proposed to compress large neural networks efficiently without compromising their original accuracy. This algorithm sequentially minimizes layer-wise reconstruction errors by treating scaling factors and integer bit-codes as variables. COMQ requires no hyper-parameter tuning, only dot products and rounding operations. It achieves remarkable results in quantizing 4-bit Vision Transformers with a negligible loss of less than 1% in Top-1 accuracy. In addition, COMQ maintains near-lossless accuracy in 4-bit INT quantization of convolutional neural networks with a minimal drop of merely 0.3% in Top-1 accuracy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary COMQ is an innovative algorithm that helps compress large neural networks without losing their original accuracy. It works by minimizing errors in each layer, using a clever combination of scaling factors and bit-codes. This makes it easy to use and doesn’t require any tricky adjustments. COMQ is great at reducing the size of 4-bit Vision Transformers, keeping their accuracy almost perfect. It also does well with convolutional neural networks, losing only a tiny amount of accuracy when reduced to 4-bit. |
Keywords
* Artificial intelligence * Quantization
