Abstract

Modern machine and deep learning require large amounts of training data. Yet, even if the data itself is abundantly available, the fraction of annotated data may still be proportionally small or missing. Hence, learning with limited labeled data is an important research field. Two streams of research attack this problem from opposite directions [64]. On the one hand, semi-supervised learning aims to leverage all information by directly incorporating unlabeled data. On the other hand, active learning finds unlabeled data for that annotations would be most beneficial for learning, and queries humans-in-the-loop of model training. This chapter discusses both concepts and their essential principles, methodological overlaps, and strengths and weaknesses. Furthermore, we elaborate on possible combinations and their advantages ands disadvantages. Finally, the conclusion refers to recent state-of-the-art and provides an outlook into the future of learning with few labeled data.