Summary of Hyper-compression: Model Compression Via Hyperfunction, by Fenglei Fan et al.
Hyper-Compression: Model Compression via Hyperfunction
by Fenglei Fan, Juntong Fan, Dayang Wang, Jingbo Zhang, Zelin Dong, Shijun Zhang, Ge Wang, Tieyong Zeng
First submitted to arxiv on: 1 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Emerging Technologies (cs.ET)
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 proposed hyper-compression method addresses the issue of model compression by leveraging a hyperfunction to represent the parameters of the target network. This approach, inspired by the relationship between genotype and phenotype, enables preferential compression ratios, no post-hoc retraining, affordable inference time, and short compression time. The hyper-compression achieves close-to-int4-quantization performance without retraining and with a performance drop of less than 1%. Additionally, it compresses LLaMA2-7B in an hour, demonstrating its efficiency. This work facilitates the harmony between the scaling law and stagnation of hardware upgradation by saving both computation and data. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper proposes a new way to shrink big models on computers so they don’t take up too much memory. They use a special function called “hyper-compression” that helps reduce the size of the model while keeping its performance good enough. This is helpful because as models get bigger, it’s harder for computers to handle them. The method works well and can even shrink big models like LLaMA2-7B in just an hour. It also doesn’t require extra training or make the model perform much worse. Overall, this work helps solve a problem that has been holding back advancements in machine learning. |
Keywords
» Artificial intelligence » Inference » Machine learning » Model compression » Quantization
