Summary of Understanding and Mitigating Miscalibration in Prompt Tuning For Vision-language Models, by Shuoyuan Wang et al.
Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models
by Shuoyuan Wang, Yixuan Li, Hongxin Wei
First submitted to arxiv on: 3 Oct 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 A novel approach to confidence calibration is proposed to ensure the safe deployment of machine learning models, particularly in vision-language models like CLIP after fine-tuning. Existing prompt tuning methods are found to lead to a trade-off between calibration for base and new classes, with overconfidence in new classes and underconfidence in base classes. To address this issue, Dynamic Outlier Regularization (DOR) is introduced, which minimizes the feature deviation of novel textual labels sampled from a large vocabulary. This approach prevents increased textual divergence for new labels while easing restrictions on base classes. Experimental results show that DOR enhances the calibration performance of current fine-tuning methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Machine learning models need to be calibrated to work safely in real-world situations. A big problem is that some models become overconfident when dealing with new information, which can lead to bad decisions. Researchers have been trying to fix this issue by tuning the prompts used to train these models. However, they found that existing methods actually make things worse – they make the model too confident in new situations and not confident enough in familiar ones. To solve this problem, the authors propose a new method called Dynamic Outlier Regularization (DOR). This method helps the model avoid becoming overconfident by making sure it’s not getting too far away from what it already knows. The results show that DOR can improve how well models are calibrated. |
Keywords
» Artificial intelligence » Fine tuning » Machine learning » Prompt » Regularization
