Summary of Entropy Is Not Enough For Test-time Adaptation: From the Perspective Of Disentangled Factors, by Jonghyun Lee et al.
Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors
by Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, Sungroh Yoon
First submitted to arxiv on: 12 Mar 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
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 Test-time adaptation (TTA) fine-tunes pre-trained deep neural networks for unseen test data. The primary challenge of TTA is limited access to the entire test dataset during online updates, causing error accumulation. To mitigate this issue, novel TTA methods have utilized entropy as a confidence metric, aiming to determine which samples are less likely to cause errors. However, experimental studies revealed the unreliability of entropy under biased scenarios, attributing it to neglecting latent disentangled factors’ influence on predictions. Building upon these findings, we introduce DeYO, a novel TTA method leveraging Pseudo-Label Probability Difference (PLPD), which quantifies shape’s influence on prediction by measuring pre- and post-object-destructive transformation differences. DeYO consists of sample selection and weighting using entropy and PLPD concurrently. For robust adaptation, DeYO prioritizes samples incorporating dominant shape information when making predictions. Our experiments demonstrate the superiority of DeYO over baseline methods across various scenarios. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about improving how deep neural networks work on new data they’ve never seen before. When we try to fine-tune these networks for new data, it’s hard because we can’t see all the test data at once. Some people tried using a special tool called entropy to help figure out which samples might be wrong. But this paper shows that entropy doesn’t work well when there are biases or problems in the data. So, they came up with a new way called DeYO (Destroy Your Object) that uses a different tool called PLPD (Pseudo-Label Probability Difference). This helps us prioritize the most important samples and make better predictions. The paper shows that DeYO works really well across many different scenarios. |
Keywords
* Artificial intelligence * Probability
