Summary of Lora Vs Full Fine-tuning: An Illusion Of Equivalence, by Reece Shuttleworth et al.
LoRA vs Full Fine-tuning: An Illusion of Equivalence
by Reece Shuttleworth, Jacob Andreas, Antonio Torralba, Pratyusha Sharma
First submitted to arxiv on: 28 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 The paper investigates the differences between Low-Rank Adaptation (LoRA) and full fine-tuning methods for adapting pre-trained large language models to downstream tasks. It is found that while both methods achieve similar performance on target tasks, they alter the model’s weight matrices in distinct ways. LoRA introduces new, high-ranking singular vectors, known as “intruder dimensions,” which do not appear during full fine-tuning. These intruder dimensions lead to LoRA models becoming worse at modeling the pre-training distribution and adapting less robustly to multiple tasks sequentially. The paper concludes by discussing the implications of these findings and potential ways to minimize their effects. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how different methods for fine-tuning language models change what they know. It finds that two common methods, LoRA and full fine-tuning, make changes in very different ways. Even though both methods can get similar results on specific tasks, the way they change the model is distinct. This means that if you use one method or the other, it will have a different impact on what the model can do. |
Keywords
» Artificial intelligence » Fine tuning » Lora » Low rank adaptation
