Summary of Allora: Adaptive Learning Rate Mitigates Lora Fatal Flaws, by Hai Huang et al.
ALLoRA: Adaptive Learning Rate Mitigates LoRA Fatal Flaws
by Hai Huang, Randall Balestriero
First submitted to arxiv on: 13 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
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 proposes an alternative to Low-Rank Adaptation (LoRA), a widely used technique for finetuning Large Language Models (LLMs). LoRA learns an additive low-rank perturbation of a pretrained matrix parameter to align the model with a new task or dataset. However, the authors identify three limitations of LoRA: its reliance on Dropout, which is only suitable for long training episodes; the initialization of B at 0, which creates slow training dynamics; and the scaling factor multiplying each LoRA additive perturbation, which creates “short-sighted” interactions between layers. The proposed solution, Adaptive Learning rate LoRA (ALLoRA), addresses these limitations by removing Dropout and scaling, and introducing an adaptive learning rate that scales gradients with the inverse of parameters’ L2 norm. ALLoRA reduces two hyper-parameters from LoRA and achieves better accuracy on various settings, including against recent LoRA variants. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper solves a problem in fine-tuning large language models. They want to make these models work better for new tasks or datasets. To do this, they use an idea called Low-Rank Adaptation (LoRA). But, they found that LoRA has three big problems. The first is that it uses something called Dropout, which only works well when training takes a long time. The second problem is that LoRA starts with all its weights being zero, which makes training slow. The third problem is that LoRA’s way of adding new information to the model creates problems between different layers of the model. To fix these problems, they created a new method called ALLoRA. ALLoRA gets rid of Dropout and the complicated scaling factor, and uses a special learning rate to make everything work better together. This makes ALLoRA more accurate than LoRA, even when compared to other ways of fine-tuning language models. |
Keywords
» Artificial intelligence » Dropout » Fine tuning » Lora » Low rank adaptation
