Research map / Label-efficient and robust learning / Noisy labels
Robust Logit Adjustment for Learning with Long-Tailed Noisy Data
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 Robust Logit Adjustment, which combines label refurbishment (GMM-based clean probability mixing noisy and pseudo-labels) with logit adjustment using an online EMA-updated estimate of refurbished-label class frequencies, to train under long-tailed noisy labels.
Problem
Learning with noisy labels methods break down under long-tailed class distributions, and long-tailed methods rely on label frequencies that are corrupted by label noise, so both fail on long-tailed noisy data.
Approach
Use label refurbishment to correct noisy labels with per-sample clean probabilities, then estimate class frequencies online from the refurbished labels (EMA) and feed them into logit adjustment, so the class-rebalancing offsets track the current training targets rather than the noisy label frequencies.
Key results
- On long-tailed CIFAR-10 with noise rate 0.5 and imbalance ratio 0.01, Robust Logit Adjustment reaches 86.71% best accuracy versus 72.96% for SSBL2.
- On long-tailed Food-101N, it achieves 75.58%, 68.76%, and 63.88% at imbalance ratios 0.05, 0.02, and 0.01.
- Ablation on CIFAR-10 shows removing logit adjustment drops accuracy from 86.71% to 72.96% at noise 0.5 and imbalance 0.01.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Robust Logit Adjustment instead of Logit Adjustment | CIFAR-10 · CIFAR-10, noise 0.5, imbalance 0.01 | accuracy (best): 78.74 -> 86.71 | Helped |
| Use Robust Logit Adjustment instead of Logit Adjustment | CIFAR-100 · CIFAR-100, noise 0.5, imbalance 0.01 | accuracy (best): 47.16 -> 50.84 | Helped |
| Use Robust LA instead of SSBL2 | Animal-10N · Animal-10N, imbalance 0.01 | accuracy: 69.36 -> 71.00 | Helped |
| Use Robust LA instead of SSBL2 | Food-101N · Food-101N, imbalance 0.01 | accuracy: 63.67 -> 63.88 | Helped |
| Add logit adjustment (ablation) | Robust Logit Adjustment · CIFAR-10, noise 0.5, imbalance 0.01 | accuracy (best): 72.96 -> 86.71 | Helped |
| Add logit adjustment (ablation) | Robust Logit Adjustment · CIFAR-10, noise 0.2, imbalance 0.01 | accuracy (best): 78.47 -> 86.39 | Helped |
| Add online update of class distribution (ablation) | Robust Logit Adjustment · CIFAR-10, noise 0.5, imbalance 0.01 | accuracy (best): 44.30 -> 86.71 | Helped |
| Add online update of class distribution (ablation) | Robust Logit Adjustment · CIFAR-10, noise 0.2, imbalance 0.01 | accuracy (best): 64.52 -> 86.39 | Helped |
| Use logit adjustment rather than inverse class frequency, inverse square-root class frequency, or effective-number reweighting (ablation) | Robust Logit Adjustment · Long-tailed noisy CIFAR-10 | accuracy | Helped |
| Vary the EMA smoothing coefficient α (ablation) | Robust Logit Adjustment · Long-tailed noisy CIFAR-10 | accuracy | No clear effect |
Surprising findings
- Increasing synthetic noise rate from 0.2 to 0.5 did not reduce accuracy at imbalance 0.01 (86.71% vs 86.39%), which the authors attribute to noise partially mitigating class imbalance.
- On real-world Animal-10N, the method does not achieve the best result at every imbalance ratio; SSBL2 is higher at 0.05 and 0.02.
Limitations
- Assumes class-independent noise; authors note class-dependent or instance-dependent noise is future work.
- Relies on model predictions after warm-up being reasonably balanced for clean-probability estimation.
- Not best on all real-world imbalance settings (e.g., Animal-10N vs SSBL2).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Animal-10N Food-101N
Benchmarks: CIFAR-10 CIFAR-100 Animal-10N Food-101N
Metrics: test accuracy (best) test accuracy (last 10 epoch average)
Baselines: ERM Co-teaching ELR+ DivideMix Logit Adjustment SSBL2 DivideMix-LA DivideMix-DRW
Builds on: logit adjustment DivideMix label refurbishment (Reed et al. 2014) FixMatch SSBL
Closest papers by meaning
- Long-tail learning via logit adjustment (ICLR 2021 · similarity 0.93)
- CoLA: Co-Calibrated Logit Adjustment for Long-Tailed Semi-Supervised Learning (ICLR 2026 · similarity 0.90)
- Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition (ICML 2024 · similarity 0.86)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.86)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.86)
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.