Research map / Label-efficient and robust learning / Recognition and long-tail
Long-Tailed Recognition via Weight Balancing
Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, Shu Kong
Check 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 |
|---|---|---|---|
| Tune weight decay during cross-entropy feature learning (ablation) | CIFAR100-LT (IF=100), ResNet32 · CIFAR100-LT (IF=100) | top-1 accuracy: 38.38 -> 46.08 | Improved |
| Add tau-normalization to the tuned-WD model (ablation) | CIFAR100-LT (IF=100), ResNet32 · CIFAR100-LT (IF=100) | top-1 accuracy: 46.08 -> 51.31 | Improved |
| Add L2 normalization to the tuned-WD model (ablation) | CIFAR100-LT (IF=100), ResNet32 · CIFAR100-LT (IF=100) | top-1 accuracy: 46.08 -> 49.60 | Improved |
| Retrain the classifier with CB loss only (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 46.08 -> 47.09 | Improved |
| Add classifier weight decay with CB loss (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 47.09 -> 52.42 | Improved |
| Add MaxNorm to classifier weight decay and CB loss (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 52.42 -> 53.35 | Improved |
| Use L2 normalization rather than classifier WD and MaxNorm with CB loss (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 53.35 -> 50.10 | Hurt |
| Fine-tune the last two layers instead of only the classifier (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 53.35 -> 53.55 | Improved |
| Fine-tune the last five layers rather than only the classifier (ablation) | Tuned-WD model on CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 53.35 -> 51.45 | Hurt |
| Use WD + MaxNorm weight balancing instead of the naive model | CIFAR100-LT (IF=100) · CIFAR100-LT (IF=100) | top-1 accuracy: 38.38 -> 53.35 | Improved |
| Use weight balancing instead of the naive model | ImageNet-LT · ImageNet-LT test | top-1 accuracy: 38.0 -> 53.9 | Improved |
| Use weight balancing instead of the naive model | iNaturalist2018 · iNaturalist2018 | top-1 accuracy: 46.1 -> 70.2 | Improved |
Closest papers by meaning
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.94)
- Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · similarity 0.91)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.90)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.90)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.89)
- Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 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.
