Research map / Label-efficient and robust learning / Noisy labels
Generalized Jensen-Shannon Divergence Loss for Learning with Noisy Labels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GJS instead of GCE | ResNet-34 · CIFAR-100, 60% symmetric noise | test accuracy: 65.21 ± 0.16 -> 70.15 ± 0.30 | Helped |
| Use GJS instead of the strongest table baseline BS | ResNet-34 · CIFAR-100, 40% asymmetric noise | test accuracy: 64.67 ± 0.69 -> 63.70 ± 0.22 | Hurt |
| Use GJS instead of GCE | ResNet-34 · CIFAR-100, 80% symmetric noise | test accuracy: 49.68 ± 0.84 -> 44.49 ± 0.53 | Hurt |
| Train two independent GJS networks and average predictions (ablation) | GJS on WebVision and ILSVRC12 · WebVision validation | Top 1 accuracy: 77.99 ± 0.35 -> 79.28 ± 0.24 | Helped |
| Train two independent GJS networks and average predictions (ablation) | GJS on WebVision and ILSVRC12 · ILSVRC12 validation | Top 1 accuracy: 74.33 ± 0.46 -> 75.50 ± 0.17 | Helped |
| Add the consistency term to the mean-prediction JS variant (ablation) | GJS · CIFAR-100, 40% symmetric noise | validation accuracy: 68.7 -> 74.3 | Helped |
| Apply GJS to noisy examples rather than only clean examples (ablation) | mixed JS/GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 72.9 -> 74.1 | Helped |
| Apply GJS to both clean and noisy examples instead of noisy examples only (ablation) | mixed JS/GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 74.1 -> 74.7 | Helped |
| Remove CutOut (ablation) | GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 74.8 -> 71.3 | Hurt |
| Remove RandAug (ablation) | GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 74.8 -> 70.6 | Hurt |
| Use weak rather than full augmentation (ablation) | GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 74.8 -> 66.5 | Hurt |
| Reduce training from 400 to 200 epochs (ablation) | GJS training · CIFAR-100, 40% symmetric noise | validation accuracy: 74.8 -> 72.9 | Hurt |
Closest papers by meaning
- Decoupled Kullback-Leibler Divergence Loss (NeurIPS 2024 · similarity 0.86)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.85)
- Loss Functions and Operators Generated by f-Divergences (ICML 2025 · similarity 0.85)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.85)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.85)
- Robust Loss Functions for Training Decision Trees with Noisy Labels (AAAI 2024 · similarity 0.85)
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.