Summary of Improving Calibration by Relating Focal Loss, Temperature Scaling, and Properness, By Viacheslav Komisarenko and Meelis Kull
Improving Calibration by Relating Focal Loss, Temperature Scaling, and Properness
by Viacheslav Komisarenko, Meelis Kull
First submitted to arxiv on: 21 Aug 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 This paper explores the relationship between loss functions and classifier calibration in machine learning. Proper losses like cross-entropy encourage models to produce well-calibrated class probabilities on training data, but they often become overconfident on test data due to the generalization gap. The focal loss is not proper, yet training with it has been shown to lead to better-calibrated classifiers. The authors demonstrate that focal loss can be decomposed into a confidence-raising transformation and a proper loss, explaining why it pushes models to provide under-confident predictions on training data. This leads to improved calibration on test data due to the generalization gap. Additionally, the paper reveals a strong connection between temperature scaling and focal loss through its confidence-raising transformation, proposing a new post-hoc calibration method called focal temperature scaling. Experiments on three image classification datasets show that focal temperature scaling outperforms standard temperature scaling. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how machine learning models make predictions. It shows that some models are more accurate than others because they can adjust their confidence levels. The authors explain why this happens and propose a new way to improve the accuracy of these models. They tested this method on several image classification tasks and found that it works better than previous methods. |
Keywords
» Artificial intelligence » Cross entropy » Generalization » Image classification » Machine learning » Temperature
