Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Vision Transformers
Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, Yu-Gang Jiang
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 unlabeled data via FixMatch-style pseudo labels to ViT-S | ViT-S · ImageNet 10% labels | top-1 accuracy: 48.6 -> 59.0 (+10.4) | Improved |
| use CNN pseudo labels instead of ViT pseudo labels (conv-labeled) | ViT-S · ImageNet 10% labels | top-1 accuracy: 59.0 -> 67.2 | Improved |
| add cross-stream feature fusion to conv-labeled (ablation) | ViT-S + Conv · ImageNet 10% labels | top-1 accuracy: 67.2 -> 72.4 (+5.2) | Improved |
| add cross-stream feature fusion to conv-labeled (PiT-S) (ablation) | PiT-S + Conv · ImageNet 10% labels | top-1 accuracy: 67.8 -> 70.8 | Improved |
| add cross-stream feature fusion to conv-labeled (ConvMixer) (ablation) | ViT-S + ConvMixer · ImageNet 10% labels | top-1 accuracy: 66.7 -> 72.9 | Improved |
| average predictions of both streams at inference instead of transformer stream only (ablation) | Semiformer ViT-S + Conv · ImageNet 10% labels | top-1 accuracy: 72.4 -> 73.5 | Improved |
| generate pseudo labels with CNN stream instead of transformer stream (ablation) | Semiformer · ImageNet 10% labels | top-1 accuracy: 67.4 -> 73.5 | Improved |
| average CNN and transformer predictions for pseudo labels instead of CNN only (ablation) | Semiformer · ImageNet 10% labels | top-1 accuracy: no better than 73.5 | Hurt |
| add Semiformer on 5% labels over conv-labeled (ablation) | ViT-S + Conv · ImageNet 5% labels | top-1 accuracy: 62.0 -> 66.3 (+4.3) | Improved |
| add Semiformer on 20% labels over conv-labeled (ablation) | ViT-S + Conv · ImageNet 20% labels | top-1 accuracy: 74.8 -> 78.1 (+3.3) | Improved |
| add Semiformer on Places205 5% labels over conv-labeled (ablation) | ViT-S + Conv · Places205 5% labels | top-1 accuracy: 52.5 -> 53.8 (+1.3) | Improved |
| add Semiformer on Places205 over vanilla ViT-S | ViT-S · Places205 5% labels | top-1 accuracy: 46.9 -> 53.8 (+6.9) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FixMatch — FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) · 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
- Semi-supervised Vision Transformers at Scale (NeurIPS 2022 · similarity 0.91)
- SemiCVT: Semi-Supervised Convolutional Vision Transformer for Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Big Self-Supervised Models are Strong Semi-Supervised Learners (NeurIPS 2020 · similarity 0.88)
- Going Deeper With Image Transformers (ICCV 2021 · similarity 0.87)
- An Empirical Study of Training Self-Supervised Vision Transformers (ICCV 2021 · similarity 0.87)
- Revisiting Semi-Supervised Learning in the Era of Foundation Models (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.
