Research map / Label-efficient and robust learning / Recognition and long-tail
Procrustean Training for Imbalanced Deep Learning
Han-Jia Ye, De-Chuan Zhan, Wei-Lun Chao
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 Major Feature Weakening | ERM · CIFAR-10 step (rho=100) | accuracy: 65.8 -> 80.1 | Improved |
| add Major Feature Weakening | ERM · CIFAR-10 long-tailed (rho=100) | accuracy: 71.1 -> 78.5 | Improved |
| add Deferred Re-Weight to MFW (ablation) | MFW · CIFAR-10 step (rho=100) | accuracy: 80.1 -> 81.6 | Improved |
| add DRW to MFW (ablation) | MFW · CIFAR-100 step (rho=100) | accuracy: 46.9 -> 48.4 | Improved |
| add DRW to MFW (ablation) | MFW · CIFAR-100 (rho=10) long-tailed | accuracy: 60.1 -> 59.1 | Hurt |
| apply MFW after 2nd conv group rather than input layer (ablation) | MFW · CIFAR-10 step (rho=100) | accuracy: 68.1 -> 80.1 | Improved |
| apply MFW after 2nd group rather than 4th group (ablation) | MFW · CIFAR-100 long-tailed (rho=100) | accuracy: 40.3 -> 44.7 | Improved |
| keep labels unmixed (MFW) instead of mixup (ablation) | mixup · CIFAR-10 long-tailed (rho=100) | accuracy: 73.1 -> 78.5 | Improved |
| increase beta-distribution alpha from 0.1 to 5 (ablation) | MFW · CIFAR-10 step | accuracy: 73.3 -> 80.1 | Improved |
| increase alpha from 0.1 to 1 (ablation) | MFW · CIFAR-100 long-tailed | accuracy: 43.1 -> 42.1 | Mixed |
| add MFW under extreme imbalance rho=1000 | ERM · CIFAR-10 long-tailed (rho=1000) | accuracy: 56.5 -> 62.0 | Improved |
| add MFW to a DenseNet-121 backbone | DenseNet-121 · CIFAR-10 long-tailed (rho=100) | accuracy: 77.6 -> 81.2 | Improved |
Closest papers by meaning
- Neural Collapse for Cross-entropy Class-Imbalanced Learning with Unconstrained ReLU Features Model (ICML 2024 · similarity 0.87)
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.85)
- Restoring balance: principled under/oversampling of data for optimal classification (ICML 2024 · similarity 0.85)
- Neural Collapse To Multiple Centers For Imbalanced Data (NeurIPS 2024 · similarity 0.84)
- Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models (NeurIPS 2024 · similarity 0.84)
- Identifying and Understanding Cross-Class Features in Adversarial Training (ICML 2025 · similarity 0.84)
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.
