Research map / Label-efficient and robust learning / Recognition and long-tail
No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problems
Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, Christopher Ré
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 |
|---|---|---|---|
| train with GEORGE (cluster-based GDRO) instead of ERM | ResNet · Waterbirds | robust accuracy: 63.3 -> 76.2 | Improved |
| train with GEORGE instead of ERM | ResNet · CelebA | robust accuracy: 41.1 -> 52.4 | Improved |
| train with GEORGE instead of ERM | ResNet · U-MNIST | robust accuracy: 93.9 -> 95.7 | Improved |
| train with GEORGE instead of ERM | ResNet · ISIC histopathology subclass | robust AUROC: .872 -> .881 | No clear effect |
| train with GEORGE instead of ERM | ResNet · ISIC non-patch subclass | robust AUROC: .920 -> .918 | No clear effect |
| train with GEORGE instead of ERM | ResNet · Waterbirds | overall accuracy: 97.2 -> 95.5 | Hurt |
| train with GEORGE instead of ERM | ResNet · CelebA | overall accuracy: 95.7 -> 94.8 | Hurt |
| cluster BiT pretrained embeddings rather than ERM features in step 1 (ablation) | GEORGE · CelebA | robust accuracy: 52.4 -> 86.0 | Improved |
| cluster BiT pretrained embeddings rather than ERM features in step 1 (ablation) | GEORGE · Waterbirds | robust accuracy | Hurt |
| use true subclass labels for GDRO instead of estimated clusters (ablation) | GEORGE · CelebA | robust accuracy: 52.4 -> 85.9 | Improved |
| use true subclass labels for GDRO instead of estimated clusters (ablation) | GEORGE · Waterbirds | robust accuracy: 76.2 -> 90.7 | Improved |
| apply UMAP dimensionality reduction before clustering (ablation) | GEORGE · hidden stratification datasets | robust accuracy | Improved |
Closest papers by meaning
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.87)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.87)
- Out-Of-Domain Unlabeled Data Improves Generalization (ICLR 2024 · similarity 0.87)
- Unsupervised Learning for Class Distribution Mismatch (ICML 2025 · similarity 0.87)
- Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification (NeurIPS 2026 · similarity 0.86)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.86)
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.
