Summary of Stat: Shrinking Transformers After Training, by Megan Flynn et al.
STAT: Shrinking Transformers After Training
by Megan Flynn, Alexander Wang, Dean Edward Alvarez, Christopher De Sa, Anil Damle
First submitted to arxiv on: 29 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 In this paper, researchers introduce STAT, a novel algorithm for pruning transformer models without requiring fine-tuning. The approach eliminates attention heads and neurons while maintaining accuracy by applying a correction to the weights of subsequent layers. By compressing each layer block using matrix factorizations that preserve network structure, STAT achieves significant compression with minimal computational cost. The entire algorithm takes minutes to compress BERT on a single GPU and less than three hours for models with 7B parameters. STAT outperforms existing gradient-free pruning methods using only several hundred data examples, matching the performance of fine-tuned methods in some cases. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Pruning transformer models helps reduce their size and computational needs, making them more efficient to use. This paper presents a new way to do this called STAT. It removes parts of the model that aren’t as important while keeping the important parts the same. This means the model still works well but is smaller and faster. The algorithm takes just minutes or hours to shrink down big models like BERT, which has many parameters. What’s more, it does better than other methods at removing parts without retraining the entire model. |
Keywords
» Artificial intelligence » Attention » Bert » Fine tuning » Pruning » Transformer
