Research map / Label-efficient and robust learning / Noisy labels
AutoDO: Robust AutoAugment for Biased Data With Label Noise via Scalable Probabilistic Implicit Differentiation
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 per-point augmentation, loss reweighting, and soft-labeling | AutoDO on SVHN · IR-NR 100-0.1 | top-1 test error rate (%): 15.3 ±1.07 -> 6.3 ±0.57 | Helped |
| add per-point augmentation rather than shared augmentation hyperparameters (ablation) | AutoDO on SVHN · IR-NR 100-0.1 | top-1 test error rate (%): 13.7 ±0.77 -> 12.3 ±0.80 | Helped |
| add loss reweighting to per-point augmentation (ablation) | AutoDO on SVHN · IR-NR 100-0.1 | top-1 test error rate (%): 12.3 ±0.80 -> 8.1 ±0.14 | Helped |
| add soft-labeling to per-point augmentation and loss reweighting (ablation) | AutoDO on SVHN · IR-NR 100-0.1 | top-1 test error rate (%): 8.1 ±0.14 -> 6.3 ±0.57 | Helped |
| use per-point augmentation, loss reweighting, and soft-labeling | AutoDO on CIFAR-10 · IR-NR 10-0.1 | top-1 test error rate (%): 17.9 ±0.32 -> 13.1 ±0.16 | Helped |
| add loss reweighting to per-point augmentation (ablation) | AutoDO on CIFAR-10 · IR-NR 10-0.1 | top-1 test error rate (%): 16.6 ±0.20 -> 15.1 ±0.19 | Helped |
| add soft-labeling to per-point augmentation and loss reweighting (ablation) | AutoDO on CIFAR-10 · IR-NR 10-0.1 | top-1 test error rate (%): 15.1 ±0.19 -> 13.1 ±0.16 | Helped |
| use per-point augmentation, loss reweighting, and soft-labeling | AutoDO on CIFAR-100 · IR-NR 10-0.1 | top-1 test error rate (%): 48.9 ±0.51 -> 39.6 ±0.06 | Helped |
| add soft-labeling to per-point augmentation and loss reweighting (ablation) | AutoDO on CIFAR-100 · IR-NR 10-0.1 | top-1 test error rate (%): 48.0 ±0.38 -> 39.6 ±0.06 | Helped |
| use AutoDO with all three components rather than FAA | AutoDO on ImageNet · IR-NR 10-0.1 | top-1 test error rate (%): 44.7 -> 44.1 | Helped |
| add soft-labeling (ablation) | AutoDO on CIFAR-100 · undistorted train data, IR-NR 1-0.0 | top-1 test error rate (%): 30.6% -> 29.4% | Hurt |
| estimate hyperparameters using test data instead of validation data (ablation) | AutoDO · SVHN | test error rate (%): 2.48 ±0.04 -> 2.48 ±0.05 | No clear effect |
Closest papers by meaning
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.88)
- Adversarial AutoAugment (ICLR 2020 · similarity 0.88)
- On Random Deep Weight-Tied Autoencoders: Exact Asymptotic Analysis, Phase Transitions, and Implications to Training (ICLR 2019 · similarity 0.86)
- Adversarial Auto-Augment with Label Preservation: A Representation Learning Principle Guided Approach (NeurIPS 2022 · similarity 0.85)
- AutoGO: Automated Computation Graph Optimization for Neural Network Evolution (NeurIPS 2023 · similarity 0.85)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.84)
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.