Summary of Uncertainty-calibrated Test-time Model Adaptation Without Forgetting, by Mingkui Tan et al.
Uncertainty-Calibrated Test-Time Model Adaptation without Forgetting
by Mingkui Tan, Guohao Chen, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Peilin Zhao, Shuaicheng Niu
First submitted to arxiv on: 18 Mar 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 proposed Efficient Anti-Forgetting Test-Time Adaptation (EATA) method tackles two key challenges in test-time adaptation: optimization costs and performance degradation on in-distribution data. EATA develops an active sample selection criterion for test-time entropy minimization, introduces a Fisher regularizer to constrain important model parameters from drastic changes, and proposes EATA with Calibration (EATA-C) to separately exploit reducible model uncertainty and inherent data uncertainty for calibrated TTA. The method also utilizes the disagreement among predicted labels as an indicator of data uncertainty, and devises a min-max entropy regularizer to selectively increase and decrease prediction confidence. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper proposes a new test-time adaptation (TTA) method that solves two major issues: it’s too expensive to adapt to each test sample, and adapting to out-of-distribution data often makes the model worse at recognizing familiar data. The method uses an “active sample selection” idea to choose only the most important test samples to adapt to, and a special kind of regularizer to keep the model from changing too much. It also has a way to adjust the confidence it has in its predictions based on how uncertain the data is. |
Keywords
* Artificial intelligence * Optimization
