Tripartite Taxonomy of Active Learning Phase Transitions

Active learning dynamics shift between data-driven, transition, and model-driven phases based on budget-induced changes in generalization mechanisms. The framework identifies these phases using segmented regression and measurable proxies, explaining why uncertainty, coverage, and representativeness strategies succeed at different labeling stages.