Active Learning

machine-learning

Definition

Active Learning

Active learning is a paradigm where the learning algorithm interactively queries an information source (typically a human oracle) to obtain labels for specific unlabelled data points. Formally, given a pool of unlabelled data $U$, the learner selects instances $x^{\ast }=\arg {\max }_{x\in U}Q(x)$ using a query strategy $Q$ to maximise model performance while minimising labelling costs.

Strategic Data Acquisition

By selecting the most informative samples (e.g., those near the decision boundary or with highest uncertainty), active learning achieves target accuracy with significantly fewer labels than passive learning. This enhanced sample efficiency is however accompanied by the risk of sampling bias, where the selected instances may not fully represent the underlying distribution $P(X,Y)$.