Summary of Forget Sharpness: Perturbed Forgetting Of Model Biases Within Sam Dynamics, by Ankit Vani et al.
Forget Sharpness: Perturbed Forgetting of Model Biases Within SAM Dynamics
by Ankit Vani, Frederick Tung, Gabriel L. Oliveira, Hossein Sharifi-Noghabi
First submitted to arxiv on: 10 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
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 challenges the conventional understanding of sharpness-aware minimization (SAM), a popular technique in machine learning. SAM aims to improve generalization by minimizing sharpness, but researchers found that this approach doesn’t always correlate with actual generalization error. Instead, the authors propose a new perspective on SAM’s training dynamics, suggesting that perturbations perform “perturbed forgetting,” discarding undesirable model biases and exhibiting better generalization signals. The paper relates this concept to the information bottleneck principle, explaining observations like smaller perturbation batches leading to better generalization. The authors also introduce a novel perturbation method targeting output biases, which outperforms standard SAM, GSAM, and ASAM on ImageNet, robustness benchmarks, and transfer learning tasks on CIFAR-10 and 100. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how a popular machine learning technique called sharpness-aware minimization (SAM) really works. Researchers found that even though SAM seems to improve generalization, it doesn’t always do so consistently. The authors propose a new way of understanding SAM’s training process, suggesting that it “forgets” old biases and adapts better to new situations. They also introduce a new method for perturbing the model, which outperforms other approaches on image recognition tasks. |
Keywords
» Artificial intelligence » Generalization » Machine learning » Sam » Transfer learning
