semiüberwachte

Semiüberwachte, in German-language literature often written as semi-supervised learning, refers to learning paradigms that use a small amount of labeled data together with a large amount of unlabeled data. The aim is to improve generalization when labeling is costly or impractical by exploiting the structure present in the unlabeled data. It sits between supervised learning (labeled data only) and unsupervised learning (unlabeled data only).

Common approaches include self-training or pseudo-labeling, where a model trained on labeled data assigns labels to unlabeled examples and iterates, and co-training, which uses multiple views of the data. Graph-based methods perform label propagation over a data graph, while semi-supervised variants of support vector machines impose decision boundaries that respect unlabeled data structure. In modern deep learning, techniques such as consistency regularization and teacher-student frameworks rely on unlabeled data to stabilize and guide training.

Key assumptions behind semiüberwachte methods include the cluster assumption (points in the same cluster share a label), low-density separation, and the manifold hypothesis. When these assumptions hold, unlabeled data can substantially improve performance with limited labels. Challenges include confirmation bias from incorrect pseudo-labels, sensitivity to distribution shifts, and the potential for degraded performance if labeled data are unrepresentative.

Applications span image and video recognition, natural language processing, and speech analysis, especially where labeling is expensive. The field gained prominence in the early 2000s, with foundational work on co-training and a comprehensive treatment in Chapelle, Schölkopf, and Zien’s Semi-Supervised Learning (2006).