Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition
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 label shift correction (LSC), which estimates the agnostic test label distribution with a neural estimator trained on simulated class-prior subsets (generalized BBSE) plus adaptive logit clipping, then adjusts pre-trained long-tail model logits.
Problem
Long-tail recognition when the test label distribution is unknown and may be forward-long-tailed, uniform, or backward-long-tailed, rather than uniform as assumed by standard long-tail methods.
Approach
Generalizes black box shift estimation into a learned neural estimator trained on subsets of training data resampled under varied class priors, feeding it class-averaged logits with an adaptive logit-clipping step that removes tail-class overconfidence; the estimated test prior is then used for logit post-adjustment.
Key results
- On CIFAR100-LT, LSC reaches 48.3% accuracy under backward-50 and 68.1% under forward-50, improving over SADE by 4.2 and 2.1 points.
- On ImageNet-LT, LSC improves over SADE by 2.9 (forward-50) and 4.2 (backward-50) points.
- Adding LSC to NCL raises CIFAR100-LT backward-50 accuracy from 41.4% to 47.9%, and it improves RIDE, PaCo and SHIKE across settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply LSC label-shift correction | NCL · CIFAR100-LT-100 Forward 50 | accuracy: 65.9 -> 68.1 | Helped |
| apply LSC label-shift correction | NCL · ImageNet-LT Backward 50 | accuracy: 53.1 -> 57.3 | Helped |
| apply LSC label-shift correction | NCL · Places-LT Backward 50 | accuracy: 41.6 -> 44.4 | Helped |
| combine LSC with NCL | NCL · CIFAR100-LT-100 Forward 50 | accuracy: +5.1 | Helped |
| combine LSC with NCL | NCL · CIFAR100-LT-100 Uniform | accuracy: +0.0 | No clear effect |
| remove logit clipping during estimator training (ablation) | LSC on ImageNet-LT · Backward 50 | accuracy: 57.3 -> 48.2 | Hurt |
| remove logit clipping during estimator training (ablation) | LSC on ImageNet-LT · Forward 50 | accuracy: 72.3 -> 72.2 | No clear effect |
| remove logit clipping at test time (ablation) | LSC on ImageNet-LT · Forward 50 | accuracy: 72.3 -> 64.5 | Hurt |
| replace adaptive k with a fixed value (ablation) | LSC on ImageNet-LT · Backward 50 | accuracy: 57.3 -> 53.4 | Hurt |
| train the estimator with imbalance range F200-B200 instead of F100-B100 (ablation) | LSC on CIFAR100-LT-100 · Backward 50 | accuracy: 48.3 -> 48.4 | Helped |
| train the estimator with imbalance range F10-B10 instead of F100-B100 (ablation) | LSC on CIFAR100-LT-100 · Backward 50 | accuracy: 48.3 -> 45.2 | Hurt |
Surprising findings
- Directly estimating the test prior from base-model predictions (BBSE, RLLS, MLLS) often underperforms, and fitting normalized probabilities instead of average logits fails to generalize.
Limitations
- Estimator is trained only on test label distributions with imbalance ratios within [1/100, 100]; generalization beyond this range is shown empirically rather than guaranteed.
- Theoretical guarantees depend on label shift assumption P_S(X|Y)=P_T(X|Y) and invertible confusion matrix.
- Estimation accuracy depends on the pre-trained model's calibration; logit clipping threshold selection uses heuristic lambda=1.5.
Experimental setup
Datasets: CIFAR100-LT ImageNet-LT Places-LT
Benchmarks: CIFAR100-LT ImageNet-LT Places-LT
Metrics: micro-accuracy
Baselines: Balanced Softmax PaCo RIDE NCL SHIKE LADE SADE BalPoE BBSE RLLS MLLS MiSLAS
Builds on: black box shift estimation (BBSE) NCL SADE LADE Balanced Softmax
Closest papers by meaning
- Harnessing Hierarchical Label Distribution Variations in Test Agnostic Long-tail Recognition (ICML 2024 · similarity 0.91)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.87)
- Robust Logit Adjustment for Learning with Long-Tailed Noisy Data (AAAI 2025 · similarity 0.86)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
- LaSCal: Label-Shift Calibration without target labels (NeurIPS 2024 · similarity 0.85)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 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.