Summary of Elucidating and Overcoming the Challenges Of Label Noise in Supervised Contrastive Learning, by Zijun Long et al.
Elucidating and Overcoming the Challenges of Label Noise in Supervised Contrastive Learning
by Zijun Long, George Killick, Lipeng Zhuang, Richard McCreadie, Gerardo Aragon Camarasa, Paul Henderson
First submitted to arxiv on: 25 Nov 2023
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- 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 In this paper, researchers investigate how mislabeled data in image classification datasets affects supervised contrastive learning (SCL) methods. They show that even with advanced SCL models, labeling errors can significantly impact the algorithm’s performance and cause it to incorrectly cluster data points of different classes together. To address this issue, they propose a novel Debiased Supervised Contrastive Learning (D-SCL) objective that mitigates the bias introduced by labeling errors. The authors demonstrate that D-SCL outperforms state-of-the-art techniques for representation learning on various vision benchmarks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Labeling errors in image classification datasets can be a big problem! When we train models to group similar images together, mistakes in the labels can cause our model to get confused and misclassify new pictures. The researchers found that these labeling errors usually make it easy for the model to incorrectly group certain images as “correct” when they’re actually not. They came up with a new way to teach the model, called Debiased Supervised Contrastive Learning (D-SCL), which helps fix this problem. With D-SCL, the model does better on various image classification tasks and is less affected by mistakes in the labels. |
Keywords
* Artificial intelligence * Image classification * Representation learning * Supervised
