Research map / Label-efficient and robust learning / Noisy labels
SELFIE: Refurbishing Unclean Samples for Robust Deep Learning
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 |
|---|---|---|---|
| restart training with refurbished labels (1st run to 3rd run) (ablation) | SELFIE · CIFAR-100 pair noise 40% | test error (%): 46.7 -> 41.3 | Helped |
| restart training: increases usable training samples (ablation) | SELFIE · CIFAR-100 pair noise 40% | ratio of samples used (%): 59.2% -> 90.2% | Helped |
| restart: training error across runs (ablation) | SELFIE · CIFAR-100 pair noise 40% | training error (%): 34.1% -> 19.8% | Helped |
| add Coteaching collaboration (CoSELFIE) to SELFIE, pair noise (ablation) | SELFIE · CIFAR-100 pair noise | best test error: reduction of 0.20pp-1.82pp | Helped |
| add Coteaching collaboration (CoSELFIE) to SELFIE, symmetry noise (ablation) | SELFIE · CIFAR-100 symmetry noise | best test error: reduction of 0.24pp-0.41pp | Helped |
| use SELFIE instead of Default, pair noise 40% | DenseNet · CIFAR-100 pair noise 40% | best test error (%): 53.2 -> 41.3 | Helped |
| use SELFIE instead of Coteaching, pair noise 40% | Coteaching · Tiny-ImageNet pair noise 40% | best test error (%): 71.5 -> 64.7 | Helped |
| Coteaching under pair noise vs Default (loss-based separation fails on pair noise) | Coteaching · Tiny-ImageNet pair noise 40% | best test error (%): 69.9 -> 71.5 | Hurt |
| use SELFIE instead of Coteaching, symmetry noise 40% | Coteaching · CIFAR-100 symmetry noise 40% | best test error (%): 37.5 -> 37.1 | No clear effect |
| use ActiveBias instead of Default, symmetry noise 40% | Default · CIFAR-10 symmetry noise 40% | best test error (%): 17.4 -> 15.7 | No clear effect |
| use SELFIE instead of Default on realistic noise | VGG-19 · ANIMAL-10N | best test error (%): 20.6 -> 18.2 | Helped |
| use SELFIE instead of Default on realistic noise | DenseNet · ANIMAL-10N | best test error (%): 17.9 -> 17.0 | Helped |
Closest papers by meaning
- Self-Training With Noisy Student Improves ImageNet Classification (CVPR 2020 · similarity 0.88)
- High-Quality Self-Supervised Deep Image Denoising (NeurIPS 2019 · similarity 0.85)
- Self-Supervised Image Denoising Using Implicit Deep Denoiser Prior (AAAI 2023 · similarity 0.85)
- Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach (CVPR 2017 · similarity 0.84)
- SELF: Learning to Filter Noisy Labels with Self-Ensembling (ICLR 2020 · similarity 0.84)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.83)
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.