Active Learning
Letting the model choose which unlabeled examples to have labeled next — labeling the most informative ones first.
Active learning is a strategy to make labeling efficient: instead of labeling data at random, the model identifies the unlabeled examples it is most uncertain about (or that would most improve it), and those are sent to human annotators first. You reach good accuracy with far fewer labels.
Worked example: a classifier flags the 100 examples nearest its decision boundary (where it is least confident); labeling those 100 improves it far more than 100 random labels, because the confident ones teach it little. Gotcha: active learning saves labeling budget but introduces sampling bias — the labeled set skews toward hard/boundary cases, which can distort class balance and metrics — and it needs a well-calibrated uncertainty signal; a badly-calibrated model queries the wrong examples and the strategy backfires.