Research map / Architectures and efficiency / Attention and transformers
Sharpness-Aware Minimization Leads to Low-Rank Features
Maksym Andriushchenko, Dara Bahri, Hossein Mobahi, Nicolas Flammarion
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SAM (ASAM variant, rho 0 -> 4.0) | ResNet on CIFAR-10 · CIFAR-10 | feature rank: 5048 -> 4383 | Improved |
| add SAM (GAM variant, rho 0 -> 0.8) | ResNet on CIFAR-10 · CIFAR-10 | feature rank: 7633 -> 6897 | Improved |
| increase ASAM rho from 0.0 to 1.0 | ResNet on CIFAR-10 · CIFAR-10 | test error: 7.29 -> 6.38 | Improved |
| increase ASAM rho from 1.0 to 4.0 | ResNet on CIFAR-10 · CIFAR-10 | test error: 6.38 -> 10.64 | Hurt |
| increase GAM rho from 0.0 to 0.4 | ResNet on CIFAR-10 · CIFAR-10 | test error: 4.04 -> 3.64 | Improved |
| increase GAM rho from 0.4 to 1.6 | ResNet on CIFAR-10 · CIFAR-10 | test error: 3.64 -> 4.81 | Hurt |
| train ViT-L/16 with SAM instead of standard, after 3/4 blocks | ViT-L/16 · ImageNet-1k | feature rank: 595 -> 314 | Improved |
| train Mixer-B/16 with SAM, after 3/4 blocks | MLP-Mixer-B/16 · ImageNet-1k | feature rank: 301 -> 477 | Hurt |
| train ViT-B/16 with SAM, last block | ViT-B/16 · ImageNet-1k | feature rank: 680 -> 617 | Improved |
| train Mixer-B/16 with SAM, last block | MLP-Mixer-B/16 · ImageNet-1k | feature rank: 695 -> 620 | Improved |
| add linear bottleneck layer to image encoder (imagewise retrieval) (ablation) | contrastive image-text model · MS-COCO imagewise retrieval | retrieval error: 23.3 -> 22.7 | No clear effect |
| add linear bottleneck layer (textwise retrieval) (ablation) | contrastive image-text model · MS-COCO textwise retrieval | retrieval error: 22.2 -> 22.9 | Hurt |
Closest papers by meaning
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.91)
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.91)
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.90)
- Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.90)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.90)
- Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
