Research map / Label-efficient and robust learning / Noisy labels
Dirichlet-Based Prediction Calibration for Learning with 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 Dirichlet-based Prediction Calibration (DPC), which breaks softmax translation invariance with a constant in the exponent, trains with an evidential deep learning loss, and selects clean examples with a large-margin criterion, combined with MixMatch-style semi-supervised learning for noisy-label training.
Problem
Deep networks trained on noisy labels produce over-confident softmax predictions, which makes clean/noisy example selection and label correction unreliable.
Approach
Break the translation invariance of softmax by adding a constant gamma to the exponentiated logits, train with an evidential deep learning (Dirichlet) loss that pushes given-label logits up and complementary logits down, and select clean examples by a logit margin rather than loss, all within a two-head MixMatch framework.
Key results
- On CIFAR-100 with 50% symmetric noise, DPC reaches 76.5% test accuracy versus 74.2% for DivideMix (averaged over last 10 epochs).
- On CIFAR-10N worst (40.21% noise), DPC achieves 93.82% versus 93.24% for SOP.
- On WebVision top-1, DPC* reaches 81.1% versus 77.6% for UniCon.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPC instead of the strongest listed baseline | CIFAR-10 and CIFAR-100 with synthetic label noise · Synthetic-noise CIFAR benchmarks | test accuracy | Mixed |
| Add strong data augmentation to DPC | DPC · CIFAR-10 and CIFAR-100 synthetic-noise benchmarks | test accuracy: 96.1 -> 96.5; 95.2 -> 95.9; 93.5 -> 94.8; 95.5 -> 95.4; 94.5 -> 95.3; 93.6 -> 94.9; 79.4 -> 81.0; 76.5 -> 78.5; 63.0 -> | Helped |
| Add the EDL loss (ablation) | DPC · CIFAR-10 symmetric 50% | last-epoch test accuracy: 94.7 -> 95.2 | Helped |
| Add the large-margin selection criterion instead of small-loss selection (ablation) | DPC · CIFAR-10 symmetric 50% | last-epoch test accuracy: 94.8 -> 95.2 | Helped |
| Use two classification heads instead of one (ablation) | DPC · CIFAR-10 symmetric 50% | last-epoch test accuracy: 94.9 -> 95.2 | Helped |
| Add EDL loss to UniCon (ablation) | UniCon · CIFAR-10 symmetric noise | test accuracy: 94.2 -> 94.6; 95.6 -> 95.5; 94.2 -> 94.2 | Mixed |
| Add EDL loss to UniCon (ablation) | UniCon · CIFAR-100 symmetric noise | test accuracy: 79.0 -> 79.5; 77.1 -> 77.5; 65.0 -> 65.4 | Helped |
| Vary the EDL tradeoff parameter beta (ablation) | DPC · CIFAR-10 symmetric 50% | test accuracy | No clear effect |
| Vary the EDL tradeoff parameter beta (ablation) | DPC · CIFAR-100 symmetric 50% | test accuracy | Mixed |
| Vary the calibration constant gamma (ablation) | DPC · CIFAR-10 and CIFAR-100 symmetric 50% | test accuracy | Mixed |
Limitations
- Gains over the best prior methods on CIFAR-10 at low noise are small (e.g., 96.1 vs 95.8 at 20% symmetric).
- Large-margin criterion and two-head design add components whose benefits vary by setting; ablation shows performance drop without LM at 80% noise on CIFAR-100.
- Hyperparameters beta and gamma require tuning; sensitivity is significant on CIFAR-100.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ILSVRC12
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ILSVRC12
Metrics: test accuracy top-1 accuracy top-5 accuracy expected calibration error
Baselines: Cross-Entropy Mixup PENCIL JPL DivideMix PES ELR+ MOIT+ UniCon GCE Co-teaching CORES
Builds on: DivideMix MixMatch evidential deep learning UniCon
Closest papers by meaning
- The Devil Is in the Margin: Margin-Based Label Smoothing for Network Calibration (CVPR 2022 · similarity 0.89)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.87)
- From Noisy Prediction to True Label: Noisy Prediction Calibration via Generative Model (ICML 2022 · similarity 0.87)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.86)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.86)
- Towards Calibrated Multi-label Deep Neural Networks (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.