Summary of Dual Risk Minimization: Towards Next-level Robustness in Fine-tuning Zero-shot Models, by Kaican Li et al.
Dual Risk Minimization: Towards Next-Level Robustness in Fine-tuning Zero-Shot Models
by Kaican Li, Weiyan Xie, Yongxiang Huang, Didan Deng, Lanqing Hong, Zhenguo Li, Ricardo Silva, Nevin L. Zhang
First submitted to arxiv on: 29 Nov 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 The proposed dual risk minimization (DRM) method improves the robustness of fine-tuned models by preserving their core features. Building on foundation models, DRM combines empirical and worst-case risk minimization to balance expected performance and worst-case performance. The approach uses LLM-generated descriptions to induce core-based zero-shot predictions, serving as proxies to estimate the worst-case risk. This state-of-the-art method achieves improved out-of-distribution performance on various real-world benchmarks, including ImageNet, WILDS-iWildCam, and WILDS-FMoW. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Foundation models can be fine-tuned for specific tasks, but this often reduces their ability to handle distribution shifts. To solve this, researchers have developed methods that preserve the pre-trained features. However, not all features are robust, and these methods don’t prioritize which ones to keep. A new approach called dual risk minimization (DRM) tries to fix this by combining two types of risks: expected performance and worst-case performance. This helps models perform well on average while also being prepared for unexpected situations. The results show that DRM can improve the out-of-distribution performance of a model like CLIP ViT-L/14@336. |
Keywords
» Artificial intelligence » Vit » Zero shot
