Summary of Investigating the Benefits Of Projection Head For Representation Learning, by Yihao Xue et al.
Investigating the Benefits of Projection Head for Representation Learning
by Yihao Xue, Eric Gan, Jiayi Ni, Siddharth Joshi, Baharan Mirzasoleiman
First submitted to arxiv on: 18 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Vision and Pattern Recognition (cs.CV)
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 abstract proposes an innovative technique for obtaining high-quality representations in machine learning models by adding a projection head during training and then discarding it. The authors investigate the reasons behind this technique’s effectiveness, revealing that the implicit bias of training algorithms leads to layer-wise progressive feature weighting, where lower layers tend to have more normalized and less specialized representations. This phenomenon is theoretically characterized as beneficial in certain scenarios, improving robustness in supervised contrastive learning and supervised learning. Experimental results on various datasets, including CIFAR-10/100, UrbanCars, and shifted ImageNet versions, validate the authors’ findings. Furthermore, the abstract presents an alternative to projection head that offers a more interpretable and controllable design. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper explores why adding a projection head during training and then discarding it makes representations better in machine learning models. The authors look into why this technique works so well, and they found that something called “layer-wise progressive feature weighting” happens. This means that lower layers in the model learn to focus on certain features more than others, which can be helpful in some situations. They also discovered that adding non-linearity to the network helps lower layers learn new features that aren’t present in higher layers. All of this leads to better results when using this technique for supervised learning and contrastive learning. The authors tested their findings on several datasets and found that they worked well. |
Keywords
* Artificial intelligence * Machine learning * Supervised
