Research map / Label-efficient and robust learning / Noisy labels
Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space
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 a dual-space joint learning method (projection network with learnable prototypes plus a One-Vs-All network) with class-independent margin criteria to handle both closed-set and open-set noisy labels in learning with open-world noisy data.
Problem
Learning with noisy labels when noisy labels may come from unknown (open-set) classes during both training and inference, requiring both a classifier for known classes and an OOD detector for open-set noise.
Approach
Jointly trains a prototype-based projection network and a One-Vs-All network in a dual representation space, adds bi-level contrastive learning and consistency regularization, and uses class-independent margins (neighbor margin and negative margin) to select clean samples, weight closed-set noise, and filter open-set noise.
Key results
- On CIFAR80N, the method reaches 49.30% accuracy at Sym-80% and 67.27% at Asym-40%, outperforming SED and other baselines.
- On CIFAR80N OOD detection, average AUROC of 70.10% versus 64.93% for the NGC baseline (per the reported table).
- On Web datasets, average accuracy of 86.77% across Web-Aircraft, Web-Bird, and Web-Car, versus 85.83% for SED.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dual-space learning with clean-sample OVA and prototype losses rather than warm-up only (ablation) | LOND-DRS · CIFAR80N Sym-80% | Avg.: 45.03 -> 54.82 | Helped |
| add pseudo-label loss L_PU (ablation) | LOND-DRS baseline · CIFAR80N Asym-40% | accuracy: 66.15 -> 67.00 | Helped |
| add pseudo-label loss L_PU (ablation) | LOND-DRS baseline · CIFAR80N Sym-80% | Avg.: 54.82 -> 54.73 | Hurt |
| add consistency regularization L_Con (ablation) | LOND-DRS with L_PU · CIFAR80N Sym-80% | AUROC: 62.51 -> 63.03 | Helped |
| add bi-level contrastive learning L_BCL (ablation) | LOND-DRS with L_PU and L_Con · CIFAR80N Sym-80% | accuracy: 47.65 -> 49.30 | Helped |
| use Ours rather than SED | CIFAR80N open-world classification · CIFAR80N Sym-80% | accuracy: 42.57 -> 49.30 | Helped |
| use Ours rather than NGC | CIFAR80N OOD detection · CIFAR80N Avg. | AUROC: 64.93 -> 70.10 | Helped |
| use Ours rather than SED | real-world noisy-data classification · Web-Bird | accuracy: 82.00 -> 80.95 | Hurt |
| increase alpha_ID | LOND-DRS · CIFAR80N | average metric | Mixed |
| increase alpha_OOD | LOND-DRS · CIFAR80N | average metric | Hurt |
| replace neighbor-margin sample selection with UNICON | LOND-DRS · CIFAR80N Asym-40% | label accuracy and model performance | Hurt |
Surprising findings
- On Web-Bird, the method (80.95%) is below SED (82.00%), so it does not win on every dataset.
Limitations
- Relies on the warm-up phase and hyper-parameter choices (alpha_ID, alpha_OOD, lambda terms) tuned per setting.
- Open-set noise evaluation is mostly on synthetic CIFAR80N; web datasets have unknown noise rates.
Experimental setup
Datasets: CIFAR100 CIFAR80N CIFAR100N Web-Aircraft Web-Bird Web-Car TinyImageNet Places-365
Benchmarks: CIFAR100N CIFAR80N Web-Aircraft Web-Bird Web-Car
Metrics: top-1 accuracy AUROC FPR95
Baselines: Standard Co-teaching Co-teaching+ DivideMix NGC NCE CIR NPN-hard SED Jo-SRC UNICON DISC
Builds on: DivideMix NGC UNICON SED mixup
Closest papers by meaning
- Unlocking the Power of Open Set: A New Perspective for Open-Set Noisy Label Learning (AAAI 2024 · similarity 0.90)
- Open-set Label Noise Can Improve Robustness Against Inherent Label Noise (NeurIPS 2021 · similarity 0.89)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.87)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.87)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.87)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · 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.