Research map / Label-efficient and robust learning / Recognition and long-tail
Inverse Weight-Balancing for Deep Long-Tailed Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes inverse weight-balancing (IWB), a two-stage long-tailed learning method that fine-tunes the classifier of a fixed CE-trained encoder toward an inverted class- and sample-wise weight-norm distribution, using KL-divergence to compensate for encoder imbalance.
Problem
Deep long-tailed recognition, where models trained on imbalanced data perform poorly on tail classes; prior work balances only the classifier weight norms and ignores imbalance in the encoder.
Approach
Overbalance the classifier by learning a target distribution P = λ·NC·HC, where NC is an inverted log class-frequency term and HC is a sample-wise difficulty term from cosine similarities to all class weights, and align classifier weight norms to P with a KL-divergence loss while the encoder stays fixed.
Key results
- On CIFAR100-LT with imbalance factor 100, IWB+CB reaches 54.8% top-1, 1.25% above WD+MaxNorm (48.60% without CB; 53.55% with CB comparison).
- On ImageNet-LT, IWB achieves 55.2% top-1 vs 53.9% for WD+MaxNorm.
- On iNaturalist2018, IWB achieves 71.5% top-1 vs 70.2% for WD+MaxNorm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IWB+CB instead of WD+MaxNorm+CB | ResNet32 · CIFAR100-LT (IF 100) | top-1 accuracy: 53.55 -> 54.8 | Helped |
| Use IWB+CB instead of BCL | ResNet32 · CIFAR100-LT (IF 50) | top-1 accuracy: 56.59 -> 57.82 | Helped |
| Use IWB instead of WD+MaxNorm | ResNeXt50 · ImageNet-LT test (All) | top-1 accuracy: 53.9 -> 55.2 | Helped |
| Use IWB instead of MARC | ResNet50 · iNaturalist2018 test (All) | top-1 accuracy: 70.4 -> 71.5 | Helped |
| Add HC to NC using vector cosine similarity (ablation) | IWB on ResNet32 · CIFAR100-LT (IF 100) | top-1 accuracy: 54.6 -> 54.8 | Helped |
| Use vector rather than scalar cosine similarity for HC (ablation) | IWB on ResNet32 · CIFAR100-LT (IF 100), Few classes | top-1 accuracy: 33.3 -> 35.7 | Helped |
| Use vector rather than scalar cosine similarity for HC (ablation) | IWB on ResNet32 · CIFAR100-LT (IF 100), All | top-1 accuracy: 54.8 -> 54.8 | No clear effect |
| Replace KL-divergence loss with MSE loss (ablation) | IWB on ResNet32 · CIFAR100-LT (IF 100) | top-1 accuracy: 54.8 -> 54.4 | Hurt |
| Set λ to 0.15 instead of 0.05 (ablation) | IWB on CIFAR100-LT · CIFAR100-LT (IF 100) | top-1 accuracy: 54.4% -> 54.8% | Helped |
| Increase λ from 0.05 to 0.2 (ablation) | IWB on CIFAR100-LT, Few classes · CIFAR100-LT (IF 100) | top-1 accuracy: 33. 5% -> 36. 1% | Helped |
Surprising findings
- Lambda must be kept moderate: very small lambda reduces overall accuracy (54.4% at 0.05 vs 54.8% at 0.15), and the optimal lambda decreases as dataset size grows.
Limitations
- Hyperparameter lambda is sensitive and dataset-dependent, requiring tuning per dataset.
- Gains over the strongest prior logit adjustment baseline (WD+MaxNorm) are modest (about 1.3%).
Experimental setup
Datasets: CIFAR100-LT ImageNet-LT iNaturalist2018
Benchmarks: CIFAR100-LT ImageNet-LT iNaturalist2018
Metrics: top-1 accuracy many-shot accuracy medium-shot accuracy few-shot accuracy
Baselines: CE LogitAdjust tau-norm IIF BALMS MARC WD+MaxNorm RIDE KD SSD Paco BCL
Builds on: cross-entropy class-balanced loss tau-normalization WD+MaxNorm inverse image frequency
Closest papers by meaning
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.89)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.88)
- Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation (NeurIPS 2024 · similarity 0.87)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.87)
- Distribution Alignment Optimization through Neural Collapse for Long-tailed Classification (ICML 2024 · similarity 0.87)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.87)
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.