Research map / Label-efficient and robust learning / Semi-supervised learning
Time-Consistent Self-Supervision for Semi-Supervised 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 |
|---|---|---|---|
| remove consistency loss (ablation) | TC-SSL · CIFAR10 WRN-28-135 500/44500 to 4000/41000 | test error rate: removing it raises error at 500 labels (6.04 to 7.51) and 1000 labels (3.81 to 5.31); small gap at 2000 and 4000 | Helped |
| remove contrastive loss (ablation) | TC-SSL · CIFAR10 WRN-28-135 500/44500 to 4000/41000 | test error rate: error rises from 6.04 to 7.56 at 500 labels, 3.81 to 5.35 at 1000, 3.79 to 3.96 at 2000, 3.54 to 3.66 at 4000 | Helped |
| remove pseudo-label cross entropy on unlabeled data (ablation) | TC-SSL · CIFAR10 WRN-28-135 500/44500 to 4000/41000 | test error rate: error rises sharply, e.g. 6.04 to 41.05 at 500 labels and 3.54 to 9.87 at 4000 labels | Helped |
| replace TC-based selection with uniform sampling of unlabeled data (ablation) | TC-SSL · CIFAR10 WRN-28-135 500/44500 to 4000/41000 | test error rate: error rises from 6.04 to 12.25 at 500 labels, 3.81 to 6.39 at 1000, 3.79 to 4.68 at 2000, 3.54 to 4.05 at 4000 | Helped |
| use TC-SSL (time-consistency selection + self-supervised losses) instead of MixMatch | WRN-28-2 on CIFAR10 · CIFAR10 500 to 4000 labels | test error rate: better at 1000/2000/4000 labels (6.15 vs 7.75 at 1000); worse than ReMixMatch at 1000 labels (5.73) | Mixed |
| use TC-SSL instead of MixMatch on CIFAR100 with WRN-28-135 | WRN-28-135 on CIFAR100 · CIFAR100 2500 to 10000 labels | test error rate: error reduced by about 12 to 13 points at each split (e.g. 44.20 to 31.95 at 2500 labels; 25.88 to 22.10 at 10000) | Helped |
| use TC-SSL with a 5.9M-param model instead of MixMatch (23.8M) on STL10 | STL10 5k/100k · STL10 5k/100k | test error rate: 4.82 vs 5.59 for MixMatch | Helped |
| use TC-based curriculum selection so training uses fewer batches than MixMatch | TC-SSL vs MixMatch training efficiency · CIFAR10 four splits, small and large WRN | test accuracy vs training batches: qualitative: reaches higher test accuracy much earlier; paper says consistently uses <20% of the batches the best baseli | Helped |
Closest papers by meaning
- Realistic Evaluation of Deep Semi-Supervised Learning Algorithms (NeurIPS 2018 · similarity 0.89)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.88)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.88)
- On Data-Augmentation and Consistency-Based Semi-Supervised Learning (ICLR 2021 · similarity 0.88)
- Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning (ICML 2024 · similarity 0.87)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
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.