Active Learning

Active Learning is a label-efficient Machine Learning setting in which an Algorithm proactively selects potentially informative data points and queries an Oracle—typically a human expert—for their labels to improve a model’s training. It is distinct from Semi-Supervised Learning, which uses unlabeled data alongside labeled data during training without necessarily requesting additional labels, though the two approaches can be combined. Active learning is particularly valuable when Unlabeled Data is abundant but manual Data Labeling is costly. Using a Query Strategy such as Uncertainty Sampling, Query-by-Committee, or Expected Model Change, the system aims to achieve comparable or higher predictive Accuracy with fewer labeled training examples than Passive Learning, where the model does not choose which examples receive labels. Common settings include Pool-Based Sampling, which selects instances from an available collection of unlabeled data, and Stream-Based Selective Sampling, which evaluates incoming instances to decide whether to request their labels. These approaches are used in Natural Language Processing and Computer Vision to make Training Data acquisition more efficient.