← AI Terminology

Active Learning

Active learning lets a model select the most informative unlabeled examples for human annotation to maximise accuracy per labeling budget.

A classical strategy for expensive labels.
Why It Matters in AI
Labeling is often the bottleneck. Uncertainty sampling and diversity criteria focus human effort where it reduces error most — still highly relevant for enterprise classifiers and eval set construction.
Key Points
Aspect Description
Loop Train → query informative x → label → retrain
Risk Bias if query strategy skewed
Benefit Fewer labels for same performance
LLM era Also for preference data and eval mining
Related Weak supervision, human-in-the-loop
Criteria Uncertainty, margin, committee disagreement, coresets
Simple Analogy
A teacher who asks students to bring the homework problems that confused them most — mark those first for maximum learning per minute.
Common Usage Examples
  • Uncertainty sample low-confidence tickets
  • Label only cluster exemplars
  • Build golden sets via active mining
  • Stop when gains plateau
Summary
In short: Active learning chooses the best examples to label next — maximising model quality per unit of human annotation.