Label-Efficient Learning: Building AI Models with Limited Data

Abstract:

Recent advances in artificial intelligence have significantly improved many aspects of daily life. However, most modern AI systems rely on deep learning frameworks that demand large-scale datasets with high-quality labels, which are expensive and time-consuming to acquire. As a result, in many real-world scenarios only a limited number of labeled samples are available, which is insufficient for training well-generalized models. This thesis explores and advances label-efficient learning across several major strategies: self-supervised learning, multi-domain learning, domain adaptation, and domain generalization.

We first propose a self-supervised learning method that addresses the limitations of contrastive learning in contrast-agnostic applications, where visually similar images are difficult to distinguish. We then explore multi-domain learning, transferring knowledge from related auxiliary datasets to improve performance on a target task, with applications to cervical image analysis and transportation mode classification.

Finally, we investigate domain adaptation and domain generalization for transportation mode classification, developing frameworks that remain robust when models are deployed across different cities or regions. Together, these studies offer practical and broadly applicable strategies for building accurate models when labeled data is scarce.