Active Learning
Active learning is a machine learning training approach in which the model (or an associated selection algorithm) actively selects which unlabeled examples should be labeled next — typically the most informative or uncertain ones — rather…
Definition
Active learning is a machine learning training approach in which the model (or an associated selection algorithm) actively selects which unlabeled examples should be labeled next — typically the most informative or uncertain ones — rather than labeling data randomly or exhaustively, in order to reach strong performance with fewer labeled examples.
Overview
Labeling training data is often the most expensive and time-consuming part of building a supervised machine learning system, particularly when labels require expert annotators (e.g., medical diagnoses, legal classifications) or significant human effort. Active learning addresses this by flipping the usual passive data collection process: instead of randomly sampling a large pool of data and labeling all of it, the system is given a large pool of unlabeled data and a small labeled seed set, trains an initial model, then uses a query strategy to select the specific unlabeled examples predicted to be most valuable for improving the model if labeled next, sends only those to a human (or oracle) annotator, retrains, and repeats this loop iteratively. Common query strategies include uncertainty sampling (selecting examples where the model's prediction confidence is lowest), query-by-committee (using disagreement among an ensemble of models to identify informative examples), and expected model change or expected error reduction methods (estimating which unlabeled examples, if labeled, would most change or improve the model). The core premise, supported by substantial empirical research, is that not all data points are equally informative — near-boundary or ambiguous examples typically teach a model more per label than examples it already classifies confidently and correctly. Active learning is widely used in domains where labeling is expensive relative to the value of each label: medical imaging annotation, specialized legal or scientific document classification, and industrial defect detection are common examples. In the context of large language models, related ideas appear in techniques for selecting which examples to include in instruction-tuning or RLHF preference datasets, and in data curation pipelines that prioritize labeling diverse or high-uncertainty examples over exhaustively labeling an entire corpus. The main trade-off is that active learning adds algorithmic and infrastructure complexity (managing the iterative train-select-label-retrain loop) in exchange for potentially large reductions in total labeling cost and time to reach a target accuracy.
Key Concepts
- Model or selection algorithm chooses which unlabeled examples to have labeled next
- Aims to reach target accuracy with substantially fewer labeled examples than random sampling
- Common query strategies: uncertainty sampling, query-by-committee, expected error reduction
- Operates as an iterative loop: train, select, label, retrain
- Most valuable when labeling is expensive relative to data volume (e.g., medical, legal domains)
- Prioritizes ambiguous or near-decision-boundary examples over confidently classified ones
- Related to data curation strategies used in instruction-tuning and RLHF dataset construction
- Adds pipeline complexity in exchange for reduced total annotation cost