Research map / Label-efficient and robust learning / Noisy labels
On Learning Contrastive Representations for Learning With Noisy Labels
Check 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 CTRR instead of the strongest listed baseline | PreAct ResNet18 · CIFAR-10, symmetric 90% label noise | test accuracy: 81.65±2.46 vs 51.13±0.27 | Helped |
| Use CTRR instead of the strongest listed baseline | PreAct ResNet18 · CIFAR-100, symmetric 80% label noise | test accuracy: 43.69±0.28 vs 34.22±1.06 | Helped |
| Use CTRR instead of the strongest listed baseline | ResNet18 · ANIMAL-10N | test accuracy: 86.71±0.15 vs 85.73±0.27 | Helped |
| Use CTRR instead of the strongest listed baseline | ImageNet-pretrained ResNet18 · Clothing1M | test accuracy: 72.71±0.19 vs 72.18±0.21 | Helped |
| Replace L′ctr with the proposed log-based regularizer L̃ctr (ablation) | CTRR, PreAct ResNet18 · CIFAR-10, symmetric label noise | test accuracy: 86.05±0.33 -> 93.05±0.32; 82.34±0.25 -> 92.16±0.31; 74.35±0.54 -> 87.34±0.84 | Helped |
| Use SimCLR rather than SimSiam as the contrastive framework (ablation) | CTRR · CIFAR-10, symmetric 20% noise | test accuracy: 93.05±0.32 -> 92.50±0.35 | Hurt |
| Use BYOL rather than SimSiam as the contrastive framework (ablation) | CTRR · CIFAR-10, symmetric 80% noise | test accuracy: 83.66±0.52 -> 86.99±0.59 | Helped |
| Enable label correction (ablation) | CTRR, CIFAR-10 · symmetric 60% label noise | test accuracy: 87.34±0.84 -> 89.23±0.18 | Helped |
| Combine CTRR with GCE (ablation) | CIFAR-10 · symmetric 80% label noise | test accuracy: 83.66±0.52 -> 90.25±0.40 | Helped |
| Set confidence threshold τ to 0, accepting any pair as a confident pair (ablation) | CTRR · CIFAR-10, symmetric 60% label noise | test accuracy: ~20% | Hurt |
| Use a regularization strength λ that is too weak or too strong (ablation) | CTRR · CIFAR-10, symmetric 60% label noise | test accuracy | Mixed |
Closest papers by meaning
- Mitigating Memorization of Noisy Labels via Regularization between Representations (ICLR 2023 · similarity 0.91)
- Learning from Noisy Labels with Complementary Loss Functions (AAAI 2021 · similarity 0.91)
- Investigating Why Contrastive Learning Benefits Robustness against Label Noise (ICML 2022 · similarity 0.90)
- Selective-Supervised Contrastive Learning With Noisy Labels (CVPR 2022 · similarity 0.90)
- Multi-Objective Interpolation Training for Robustness To Label Noise (CVPR 2021 · similarity 0.89)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 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.