What is it about?
This paper introduces FATE (Firstly Adapt, Then catEgorize), a new semi-supervised learning framework designed for situations where labeled data are extremely limited, possibly only one labeled example per class. FATE uses a two-stage prompt-tuning approach: First, it adapts a pre-trained model to the target data distribution using large amounts of unlabeled data. Then, it categorizes samples through a semi-supervised learning stage tailored for pre-trained models. FATE works for both vision and vision-language models and shows substantial performance improvements over existing SSL methods.
Featured Image
Why is it important?
In many real-world tasks, collecting labeled data is expensive and time-consuming, while unlabeled data are abundant. Traditional SSL methods fail when labeled data are extremely scarce. FATE provides a practical and effective solution by efficiently leveraging pre-trained models and unlabeled data, enabling high performance even under extreme label scarcity. This makes it valuable for applications like medical imaging, rare event detection, and other low-data environments.
Perspectives
I am quite satisfied with the theoretical progress achieved in this work. FATE demonstrates a meaningful step toward addressing the challenge of semi-supervised learning under extreme label scarcity. However, there remain several limitations. The framework has not yet been fully deployed or validated in real-world applications, and its integration with other learning paradigms or modalities has not been explored. In future work, I hope to extend FATE toward more practical and cross-modal scenarios, bridging the gap between theoretical development and real-world implementation.
Hezhao Liu
Xiamen University
Read the Original
This page is a summary of: FATE: A Prompt-Tuning-Based Semi-Supervised Learning Framework for Extremely Limited Labeled Data, October 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3746027.3754998.
You can read the full text:
Contributors
The following have contributed to this page
