Research map / Label-efficient and robust learning / Noisy labels
Unlocking the Power of Open Set: A New Perspective for Open-Set Noisy Label 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 CECL, a two-step contrastive learning method for open-set noisy label learning that absorbs some open-set examples into known classes (class expansion) and uses the rest as delimiters.
Problem
Learning from data containing both closed-set and open-set label noise, where identifying open-set examples is hard, especially in severely corrupted datasets.
Approach
Observes a "class expansion" phenomenon where open-set examples are absorbed into semantically similar known classes, which helps known-class separation; CECL uses prototype-based open-set decision to add indistinct open-set examples as positives and distinguishable open-set examples as negatives (delimiters) in a MoCo/SupCon-style contrastive loss, after a Promix-based clean/noisy split.
Key results
- On CIFAR80N Sym-20% CECL reaches 77.23% test accuracy versus 67.00% for PNP.
- On Food101N CECL achieves 90.24% versus 89.64% for NGC (best prior).
- Ablation on CIFAR80N sym-20%: removing both CONT and OSD drops accuracy from 77.23% to 70.54%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CECL instead of NGC | CIFAR80N Sym-20% · CIFAR80N Sym-20% | test accuracy (%): 74.26 ± 0.23 -> 77.23 ± 0.26 | Helped |
| Use CECL instead of NGC | CIFAR80N Sym-80% · CIFAR80N Sym-80% | test accuracy (%): 36.36 ± 0.48 -> 37.21 ± 0.11 | Helped |
| Use CECL instead of NGC | CIFAR80N Asym-40% · CIFAR80N Asym-40% | test accuracy (%): 65.73 ± 0.44 -> 68.48 ± 0.14 | Helped |
| Use CECL instead of NGC | CIFAR100N Sym-20% · CIFAR100N Sym-20% | test accuracy (%): 68.47 ± 0.28 -> 69.20 ± 0.09 | Helped |
| Use CECL instead of NGC | CIFAR100N Sym-80% · CIFAR100N Sym-80% | test accuracy (%): 37.17 ± 0.41 -> 36.37 ± 0.12 | Hurt |
| Use CECL instead of NGC | CIFAR100N Asym-40% · CIFAR100N Asym-40% | test accuracy (%): 64.79 ± 0.38 -> 65.49 ± 0.24 | Helped |
| Use CECL instead of strongest listed baseline | real-world noisy datasets · Web-Aircraft, Web-Bird, Web-Car, and Food101N | test accuracy (%): 86.79 -> 87.46; 83.61 -> 83.87; 91.83 -> 92.61; 89.64 -> 90.24 | Helped |
| Add contrastive learning (ablation) | CECL · CIFAR80N Sym-20% | test accuracy (%): 73.17 -> 77.23 | Helped |
| Add contrastive learning (ablation) | CECL · CIFAR80N Asym-40% | test accuracy (%): 64.57 -> 68.48 | Helped |
| Add open-set decision (ablation) | CECL · CIFAR80N Sym-20% | test accuracy (%): 72.87 -> 77.23 | Helped |
| Add open-set decision (ablation) | CECL · CIFAR80N Asym-40% | test accuracy (%): 64.26 -> 68.48 | Helped |
| Add contrastive learning and open-set decision (ablation) | Promix (Step 1) · CIFAR80N Sym-20% | test accuracy (%): 70.54 -> 77.23 | Helped |
Surprising findings
- Adding open-set examples with high closed-set prediction scores into training significantly improves closed-set classification, contrary to common assumptions that open-set examples are harmful.
- CECL also outperforms closed-set baselines on CIFAR100N, which the authors attribute to the model treating ambiguous examples as open-set.
Limitations
- Baseline results are copied directly from prior papers rather than re-run under the authors' own setup.
- Theoretical guarantee relies on assumptions such as (sigma, delta)-augmentation and Lipschitz embeddings, and the bound is only partially characterized in the main text.
- Open-set decision depends on threshold tau and on the quality of the Step 1 clean/noisy split.
Experimental setup
Datasets: CIFAR80N CIFAR100N CIFAR10 MNIST Tiny-ImageNet Web-Aircraft Web-Bird Web-Car Food101N
Benchmarks: CIFAR80N CIFAR100N Web-Aircraft Web-Bird Web-Car Food101N
Metrics: test accuracy (%)
Baselines: Standard Decoupling Co-teaching Co-teaching+ JoCoR MoPro NGC Jo-SRC PNP CleanNet SELFIE PENCIL
Builds on: ProMix MoCo SupCon PNP Co-teaching
Closest papers by meaning
- Open-set Label Noise Can Improve Robustness Against Inherent Label Noise (NeurIPS 2021 · similarity 0.90)
- Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space (AAAI 2025 · similarity 0.90)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.87)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.86)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.86)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 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.