Summary of Understanding Linear Probing Then Fine-tuning Language Models From Ntk Perspective, by Akiyoshi Tomihari and Issei Sato
Understanding Linear Probing then Fine-tuning Language Models from NTK Perspective
by Akiyoshi Tomihari, Issei Sato
First submitted to arxiv on: 27 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 two-stage fine-tuning (FT) method, linear probing (LP) followed by fine-tuning (LP-FT), outperforms LP and FT alone for both in-distribution (ID) and out-of-distribution (OOD) data. This paper analyzes the training dynamics of LP-FT for classification tasks using the neural tangent kernel (NTK) theory. The analysis decomposes the NTK matrix into two components, highlighting the importance of linear head norm alongside prediction accuracy at the start of the FT stage. The study also observes a significant increase in the linear head norm during LP, which stems from training with cross-entropy (CE) loss. This increased norm effectively reduces changes in learned features and can affect model calibration, which can be corrected using temperature scaling. Additionally, the paper extends its analysis to low-rank adaptation (LoRA) method and validates its effectiveness. Experimental results using a Transformer-based model on multiple natural language processing datasets confirm the theoretical analysis, demonstrating the effectiveness of LP-FT for fine-tuning language models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about a new way to make big language models better. The researchers looked at how these models learn from small amounts of data and found that adding an extra step helps them generalize well. They used something called the neural tangent kernel (NTK) theory to understand what’s going on, and it helped them figure out why their method works so well. They also tested it with different types of datasets and found that it really makes a difference. |
Keywords
» Artificial intelligence » Classification » Cross entropy » Fine tuning » Lora » Low rank adaptation » Natural language processing » Temperature » Transformer
