Active learning operates as a loop. (1) A base model is trained on a small initial labeled seed set. (2) A query strategy scores the informativeness of unlabeled instances via a utility function — e.g. prediction uncertainty (least confidence, margin, entropy), committee disagreement (query-by-committee), expected gradient/model change, expected error or variance reduction, optionally weighted by density/diversity to avoid selecting near-duplicate points within a batch. (3) The most informative instances (one at a time or in a batch of size B) are sent to the oracle for labeling. (4) The new labels are added to the training set and the model is retrained. The loop repeats until the labeling budget is exhausted or a stopping criterion is met. Scenarios differ in the candidate source: pool-based scores an entire pool and picks the best, stream-based decides on each incoming instance individually, and membership query synthesis generates synthetic instances to query.
Labeling data is often the most expensive and slowest part of building ML models, while raw unlabeled data is cheap and plentiful. Passive learning (labeling a random sample) wastes the annotation budget on uninformative examples. Active learning minimizes the number of labels required by directing annotator effort toward the examples that will most improve the model.
An authoritative source of labels — usually a human expert, sometimes another system. It answers the model's queries by providing ground-truth labels for selected instances.
Official
The ML model trained on the growing labeled set; its state (uncertainty, gradients, predictions) drives the selection of the next queries.
Official
A set or stream of unlabeled instances from which candidates for labeling are selected.
The rule selecting instances: uncertainty sampling, query-by-committee, expected model change, expected error/variance reduction, density/diversity-based methods.
Official
A utility (acquisition) function assigning each instance an expected information-gain score used to rank candidates.
Official
The condition that ends the loop: exhausted labeling budget, validation-accuracy plateau, or a confidence threshold.
Official
An actively collected set is not iid with respect to the target distribution, which can hurt other models trained on that data and complicate fair evaluation.
With a very small seed set the model is weak, so its uncertainty signal is unreliable and early queries are poorly chosen.
Pure uncertainty sampling in batch mode selects highly similar, mutually redundant instances, wasting the labeling budget.
The assumption of an infallible oracle is often unrealistic; label noise and variable annotation cost erode active-learning gains.
Softmax in deep networks tends to be overconfident, so raw uncertainty is a poor measure of informativeness.
Dana Angluin formalizes membership-query learning — the theoretical basis for the membership query synthesis scenario.
Seung, Opper and Sompolinsky introduce a committee of models voting on instances; points of maximal disagreement are selected.
Cohn, Atlas and Ladner introduce the term and show improved generalization from selective instance selection.
Lewis and Gale propose a sequential algorithm for training text classifiers based on uncertainty sampling.
Burr Settles publishes the Active Learning Literature Survey, organizing scenarios and query strategies into a single, widely cited taxonomy.
Gal, Islam and Ghahramani combine active learning with Bayesian deep-network uncertainty (MC-dropout, BALD) for image data.
Sener and Savarese cast batch selection as a core-set problem, emphasizing representativeness over uncertainty alone in active learning with CNNs.
Time complexity: O(R · (C_train + |U|·c_score)).
Choice of acquisition function (uncertainty, QBC, expected model change, diversity/core-set) — governs sampling effectiveness.
Maximum number of oracle queries; determines cost and usually the accuracy ceiling.
Number of instances labeled per round; large B requires a diversity strategy to avoid redundancy.
Type of model (e.g. logistic regression, SVM, deep net); affects the quality and calibration of the uncertainty signal.
Number of labels at the start; too small causes the cold-start problem and poor early queries.
Condition for ending the loop (budget, accuracy plateau, confidence threshold).
The canonical loop is sequential (each round depends on the previous one). Batch-mode active learning parallelizes the labeling of many instances within a single round.
Active learning is a hardware-independent data-selection paradigm; it works with any base model and infrastructure.
