Summary of Lora Learns Less and Forgets Less, by Dan Biderman et al.
LoRA Learns Less and Forgets Less
by Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, Cody Blakeney, John P. Cunningham
First submitted to arxiv on: 15 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 The paper compares the performance of Low-Rank Adaptation (LoRA) and full finetuning on large language models in two target domains: programming and mathematics. In both instruction finetuning and continued pretraining settings, LoRA substantially underperforms full finetuning. However, it better maintains the base model’s performance outside the target domain. The study also shows that LoRA mitigates forgetting more effectively than common regularization techniques like weight decay and dropout. Furthermore, LoRA helps maintain more diverse generations compared to full finetuning. The results suggest that full finetuning learns perturbations with a much higher rank than typical LoRA configurations, which could explain some of the performance gaps. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary LoRA is a method used to make language models smaller and faster. This paper compares how well it works to a different way of making changes called “full finetuning”. The researchers looked at two areas: programming and math. They tried LoRA in both situations where they’re teaching the model new things (instruction finetuning) and when they’re giving it more text to learn from (continued pretraining). The results show that LoRA doesn’t do as well as full finetuning, but it does a better job of keeping the model’s original abilities. This is important because sometimes models forget what they already knew when they’re learning new things. |
Keywords
» Artificial intelligence » Dropout » Lora » Low rank adaptation » Pretraining » Regularization
