Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Masked Unsupervised Self-training for Label-free Image Classification
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 |
|---|---|---|---|
| Adapt CLIP using unlabeled images with MUST | ViT-B/16 · ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: +9.4, +8.7, +4.0, +22.1, +9.4, +12.3 | Helped |
| Adapt CLIP using unlabeled images with MUST | ViT-L/14 · ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: +6.6, +8.0, +2.4, +18.1, +7.2, +8.7 | Helped |
| Use unlabeled-image MUST adaptation instead of 16-shot CLIP adaptation | ViT-B/16 · ImageNet | accuracy: 71.5 -> 77.7 | Helped |
| Add global-local feature alignment (ablation) | MUST with self-training and MIM · Average across ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: 72.6 -> 74.1 | Helped |
| Add masked image modeling (ablation) | MUST with self-training · Average across ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: 71.9 -> 72.6 | Helped |
| Add fairness regularization (ablation) | Self-training · Average across ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: 67.8 -> 71.9 | Helped |
| Align [CLS] with only the nearest 10 [MSK] tokens instead of all tokens (ablation) | MUST · Average across ImageNet, SUN397, Food101, GTSRB, DTD, UCF101 | accuracy: 74.1 -> 74.1 | No clear effect |
| Warm up MUST on ImageNet before adaptation | MUST · Pets | accuracy: 93.1 -> 94.3 | Helped |
| Warm up MUST on ImageNet before adaptation | MUST · Caltech101 | accuracy: 93.1 -> 93.7 | Helped |
| Adapt on ImageNet, then evaluate without transductive target adaptation | Inductive MUST · ImageNet-R and ImageNet-A | accuracy: 77.7 -> 68.7; 49.9 -> 41.6 | Hurt |
| Use transductive MUST on the target distribution instead of inductive ImageNet MUST | MUST · ImageNet-R and ImageNet-A | accuracy: 68.7 -> 87.3; 41.6 -> 56.9 | Helped |
| Use MUST instead of CLIP | ViT-B/16 · ImageNet-C | mCE: 70.7 -> 53.4 | Helped |
Closest papers by meaning
- Big Self-Supervised Models are Strong Semi-Supervised Learners (NeurIPS 2020 · similarity 0.88)
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.87)
- Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning (AAAI 2021 · similarity 0.87)
- Unlabeled Data Improves Fine-Grained Image Zero-shot Classification with Multimodal LLMs (NeurIPS 2025 · similarity 0.87)
- Let Go of Your Labels with Unsupervised Transfer (ICML 2024 · similarity 0.87)
- S3A: Towards Realistic Zero-Shot Classification via Self Structural Semantic Alignment (AAAI 2024 · 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.