lowsupervision

Lowsupervision refers to learning under restricted supervision, where labeled data is scarce, expensive to obtain, or noisy. The term is not widely standardized; in practice it encompasses several related settings such as weak supervision, semi-supervised learning, self-supervised learning, and few-shot learning, all oriented toward reducing the need for extensive manually labeled data.

In many domains, labeling costs dominate, and models must generalize from limited supervision. Lowsupervision emphasizes data efficiency and robustness to imperfect labels rather than maximizing performance with fully labeled corpora.

Approaches commonly associated with lowsupervision include semi-supervised learning, which leverages large pools of unlabeled data alongside a small labeled set, often via consistency regularization or pseudo-labeling. Weak supervision uses noisy or indirect signals, such as labeling functions, heuristics, or external knowledge bases. Self-supervised learning constructs pretext tasks from unlabeled data to learn useful representations. Few-shot or low-shot learning aims to generalize from a small number of labeled examples per class. Active learning and data augmentation are additional tools used to reduce labeling requirements.

Applications of lowsupervision span natural language processing, computer vision, medical imaging, speech recognition, and remote sensing, especially in domains where labeled data is scarce or costly to obtain. Challenges include handling label noise, distribution shift, evaluating performance with limited labels, and mitigating bias, as well as computational efficiency. Ongoing research seeks to improve reliability, interpretability, and transferability of low-supervision methods.