Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Adaptive Training: Bridging Supervised and Self-Supervised Learning
Lang Huang, Chao Zhang, Change Institutions to: University of Waterloo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add EMA of model predictions as targets (ablation) | supervised self-adaptive training · CIFAR-10 symmetric label noise 40% | test accuracy: 72.00 -> 92.64 | Improved |
| add EMA of model predictions as targets (ablation) | supervised self-adaptive training · CIFAR-100 symmetric label noise 80% | test accuracy: 11.57 -> 38.72 | Improved |
| add sample re-weighting (ablation) | supervised self-adaptive training · CIFAR-100 symmetric label noise 40% | test accuracy: 69.52 -> 71.38 | Improved |
| add sample re-weighting (ablation) | supervised self-adaptive training · CIFAR-10 symmetric label noise 40% | test accuracy: 92.49 -> 92.64 | No clear effect |
| raise EMA momentum alpha from 0.9 to 0.99 (ablation) | supervised self-adaptive training · CIFAR-10 symmetric label noise 40% | test accuracy: 92.64 -> 84.38 | Hurt |
| start target updates at epoch 60 instead of 20 (ablation) | supervised self-adaptive training · CIFAR-10 symmetric label noise 40% | test accuracy: 89.58 -> 92.64 | Improved |
| replace ERM with self-adaptive training | ResNet-101 · ImageNet val, label noise 40% | top-1 accuracy: 70.2 -> 73.5 | Improved |
| replace TRADES cross entropy with self-adaptive training | TRADES · CIFAR-10 adversarial training | robust accuracy: 1% to 3% robust accuracy gain | Improved |
| use two views instead of one view (ablation) | self-supervised self-adaptive training · ImageNet linear evaluation (200 epochs) | linear top-1 accuracy: 67.4 -> 72.8 | Improved |
| use two views instead of one view (ablation) | self-supervised self-adaptive training · CIFAR-10 linear evaluation (single view) | linear top-1 accuracy: 92.27 -> 91.56 | No clear effect |
| use one view instead of two views (ablation) | BYOL · CIFAR-10 linear evaluation (single view) | linear top-1 accuracy: 91.78 -> 44.36 | Hurt |
| add predictor (momentum 0.99) (ablation) | BYOL · CIFAR-10 linear evaluation | linear top-1 accuracy: 26.87 -> 91.78 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- SELF — SELF: Learning to Filter Noisy Labels with Self-Ensembling (ICLR 2020) · better on 0 of 1 shared result
- SelectiveNet — SelectiveNet: A Deep Neural Network with an Integrated Reject Option (ICML 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Adaptive Training: beyond Empirical Risk Minimization (NeurIPS 2020 · similarity 0.94)
- Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data (ICLR 2021 · similarity 0.90)
- Neighborhood-Regularized Self-Training for Learning with Few Labels (AAAI 2023 · similarity 0.87)
- Debiased Self-Training for Semi-Supervised Learning (NeurIPS 2022 · similarity 0.87)
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.87)
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 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.
