active data acquisition

Labels cost money, so do not buy them at random — buy the ones that teach the model the most. Active data acquisition is the loop where the current model points at the unlabeled examples whose labels would most improve it, an oracle (often a human) labels just those, the model retrains, and the cycle repeats. With a good acquisition rule you can reach target accuracy with a fraction of the labels a random budget would need.

Acquisition functions encode 'most informative' in different ways. Uncertainty sampling picks examples where the model is least confident (smallest margin, highest predictive entropy). Query-by-committee picks where an ensemble disagrees most. Bayesian methods (BALD) pick examples that maximize expected information gain about the parameters, distinguishing model uncertainty from data noise. Diversity- and coverage-based methods (core-set, BADGE, which combines gradient magnitude with k-means++ diversity) counter the failure mode where pure uncertainty selects redundant, near-identical hard examples in one batch. The output each round is a batch to send for labeling.

The honest difficulties: acquisition is biased sampling, so the labeled set is non-i.i.d. and naive evaluation on it is optimistic; cold-start with a weak initial model can mislead the scores; and gains shrink under distribution shift between the pool and deployment.