Research map / Label-efficient and robust learning / Noisy labels
Directional Label Diffusion Model for Learning from Noisy Labels
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 Directional Label Diffusion (DLD), a label diffusion model that splits the forward process into directional and random diffusion paths, paired with a neighbor-consistency label pre-correction, for learning from noisy labels in image classification.
Problem
Learning from noisy labels (LNL) in image classification, where label noise (especially instance-dependent noise) degrades DNN generalization and discriminative LNL methods are limited.
Approach
Reformulates LNL as generative label diffusion: the forward process is disentangled into a directional diffusion toward an estimated noise distribution and a random Gaussian diffusion, with the endpoint yT = yn + ε, and labels are generated from features conditioned on a one-step, feature-based label pre-correction.
Key results
- On CIFAR-100 with 60% IDN, DLD+ViT reaches 69.57% accuracy versus 57.21% for LRA-diffusion+ViT.
- On Animal-10N, DLD achieves 89.4% accuracy, above the prior best LRA-diffusion at 88.6%.
- On WebVision, DLD reaches 84.51% versus 84.16% for LRA-diffusion; on ILSVRC12 83.74% versus 82.56%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-10 IDN 20% · CIFAR-10 IDN 20% | accuracy: 96.65±0.18 -> 96.98±0.12 | Helped |
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-10 IDN 40% · CIFAR-10 IDN 40% | accuracy: 93.68±0.34 -> 96.49±0.07 | Helped |
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-10 IDN 60% · CIFAR-10 IDN 60% | accuracy: 78.12±1.17 -> 83.14±0.81 | Helped |
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-100 IDN 20% · CIFAR-100 IDN 20% | accuracy: 73.59±0.13 -> 78.62±0.11 | Helped |
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-100 IDN 40% · CIFAR-100 IDN 40% | accuracy: 67.67±0.21 -> 76.03±0.14 | Helped |
| Use DLD+ViT instead of LRA-diffusion+ViT | CIFAR-100 IDN 60% · CIFAR-100 IDN 60% | accuracy: 57.21±0.12 -> 69.57±0.17 | Helped |
| Use DLD instead of J-ViT | Animal10N · Animal10N | accuracy: 89.0 -> 89.4 | Helped |
| Use DLD instead of SANM | Clothing1M · Clothing1M | accuracy: 75.63 -> 75.69 | Helped |
| Use DLD instead of LRA-diffusion | WebVision · WebVision | accuracy: 84.16 -> 84.51 | Helped |
| Use DLD instead of LRA-diffusion | ILSVRC2012 · ILSVRC2012 | accuracy: 82.56 -> 83.74 | Helped |
| Add label pre-correction (ablation) | DLD · CIFAR-10 and CIFAR-100 noisy-label experiments | label accuracy | Helped |
| Add the DLD diffusion model after label pre-correction (ablation) | Pre-correction · CIFAR-10 and CIFAR-100 noisy-label experiments | label accuracy | Helped |
Surprising findings
- The LRA neighbor-label re-estimation method disrupts originally correct labels, whereas the proposed pre-correction improves label accuracy.
- DLD+ResNet underperforms in low-noise settings, which the authors attribute to introducing too much robust information.
Limitations
- Performance gains partly rely on strong pre-trained feature extractors such as ViT-L/14.
- Low-noise performance with ResNet features is weaker than baselines.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Animal-10N Clothing1M WebVision ILSVRC2012
Benchmarks: CIFAR-10 (IDN) CIFAR-100 (IDN) Animal-10N Clothing1M WebVision ILSVRC2012
Metrics: classification accuracy label accuracy
Baselines: Co-teaching GCE DAC DMI SEAL NCR CC+DivideMix DISC LRA-diffusion EPL J-ViT DivideMix
Builds on: DDPM RDDM CARD LRA-diffusion DDIM
Closest papers by meaning
- Label-Retrieval-Augmented Diffusion Models for Learning from Noisy Labels (NeurIPS 2023 · similarity 0.91)
- Label-Noise Robust Diffusion Models (ICLR 2024 · similarity 0.89)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.89)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.88)
- Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness (AAAI 2025 · similarity 0.88)
- L2B: Learning to Bootstrap Robust Models for Combating Label Noise (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.