Research map / Label-efficient and robust learning / Recognition and long-tail
Feature Directions Matter: Long-Tailed Learning via Rotated Balanced Representation
Gao Peifeng, Qianqian Xu, Peisong Wen, Zhiyong Yang, Huiyang Shao, Qingming Huang
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 |
|---|---|---|---|
| Replace fixed ETF classifier with learnable orthogonal rotation of the balanced feature (RBL) (ablation) | ETF-based balanced representation learning · CIFAR100-LT (IF 50/100/200) | test accuracy: +1.0, +1.4, +1.7 over Fixed | Improved |
| Learnable rotation (vs fixed ETF direction) without logit adjustment (ablation) | ETF balanced representation · CIFAR100-LT (IF 200) | test accuracy: 43.4 vs 41.7 | Improved |
| Fixing ETF classifier direction without logit adjustment (ablation) | CE baseline with ETF classifier · CIFAR100-LT (IF 50/100/200) | test accuracy: -1.3, -0.2, -1.0 vs CE | Hurt |
| Add post-hoc logit adjustment (LD) at test time (ablation) | CE baseline · CIFAR100-LT (IF 200) | test accuracy: +3.9 (42.7 to 46.6) | Improved |
| Add post-hoc logit adjustment (LD) to learnable rotation model (ablation) | RBL · CIFAR100-LT (IF 200) | test accuracy: 43.4 to 48.9 | Improved |
| Full RBL vs Fixed ETF on CIFAR10-LT | RBL · CIFAR10-LT (IF 50/100/200) | test accuracy: +0.5, +0.7, +1.0 | Improved |
| Full RBL vs Fixed ETF on ImageNet-LT | RBL · ImageNet-LT | All accuracy: +2.1 (51.2 to 53.3) | Improved |
| Full RBL vs Fixed ETF on Places-LT | RBL · Places-LT | All accuracy: +0.8 (38.0 to 38.8) | Improved |
| Full RBL vs LADE on Places-LT | RBL · Places-LT | All accuracy: 0.0 (38.8 vs 38.8) | No clear effect |
| Full RBL vs best baseline LADE on ImageNet-LT | RBL · ImageNet-LT | All accuracy: +0.3 (53.0 to 53.3) | Improved |
Closest papers by meaning
- MLC-NC: Long-Tailed Multi-Label Image Classification Through the Lens of Neural Collapse (AAAI 2025 · similarity 0.86)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.86)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.86)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.85)
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.
