Summary of On the Generalization and Causal Explanation in Self-supervised Learning, by Wenwen Qiang et al.
On the Generalization and Causal Explanation in Self-Supervised Learning
by Wenwen Qiang, Zeen Song, Ziyin Gu, Jiangmeng Li, Changwen Zheng, Fuchun Sun, Hui Xiong
First submitted to arxiv on: 1 Oct 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 The paper investigates self-supervised learning (SSL) methods, which achieve high generalization performance on downstream tasks but may suffer from overfitting to their training data and lose adaptability to new tasks. The authors conduct experiments on various SSL methods and datasets, making two key observations: overfitting occurs abruptly in later layers and epochs, while early layers learn generalizing features for all epochs; and coding rate reduction can be used as an indicator of overfitting degree in SSL models. To mitigate this issue, the authors propose Undoing Memorization Mechanism (UMM), a plug-and-play method that aligns feature distributions to maximize coding rate reduction. The learning process is a bi-level optimization process, which improves generalization performance on various downstream tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at self-supervised learning methods that learn from data without labels and do well on new tasks. But they can get stuck in their training data and not adapt well to new things. To see what happens, the authors tried different methods and datasets and found two important things: overfitting happens suddenly in later layers and epochs, but early layers learn general features for all epochs; and a measure called coding rate reduction can show how much they’re overfitting. To fix this problem, the authors came up with a new method that helps the model get unstuck by aligning its early and late features to work better. |
Keywords
» Artificial intelligence » Generalization » Optimization » Overfitting » Self supervised
