Summary of Sparse Autoencoders Reveal Selective Remapping Of Visual Concepts During Adaptation, by Hyesu Lim et al.
Sparse autoencoders reveal selective remapping of visual concepts during adaptation
by Hyesu Lim, Jinho Choi, Jaegul Choo, Steffen Schneider
First submitted to arxiv on: 6 Dec 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 This paper develops a new type of Sparse Autoencoder (SAE) called PatchSAE, which is designed to extract interpretable concepts from foundation models like the CLIP vision transformer. The goal is to identify specific features such as shape, color, or semantics and their spatial attributes in downstream image classification tasks. By analyzing how these concepts influence model outputs, the authors investigate recent prompt-based adaptation techniques that modify the association between input data and these concepts. Surprisingly, they find that most gains from adaptation can be attributed to existing concepts already present in the non-adapted foundation model. This work provides a framework for training and using SAEs with vision transformers and sheds light on the mechanisms driving adaptation. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at how we can make machine learning models better suited for specific tasks by modifying them slightly. The authors create a new way to analyze what these modified models are “looking” at when they make predictions. They find that most of the improvement comes from using features that were already present in the original model, rather than discovering entirely new ones. |
Keywords
» Artificial intelligence » Autoencoder » Image classification » Machine learning » Prompt » Semantics » Vision transformer
