Active learning is a machine learning approach in which the model itself chooses which unlabeled examples should be labeled next by a human expert. Labels are often the most expensive part of building a supervised learning system, because a radiologist, lawyer or linguist must review each example. Active learning reduces that cost. Instead of labeling a random sample, the algorithm asks for labels on the examples from which it expects to learn the most, so it can reach a target accuracy with far fewer labeled items. It is closely related to semi-supervised learning, since both make use of a large pool of unlabeled data.
How active learning works
The process is a loop that combines a model with a human annotator, often called the oracle.
- Start with a small labeled set and a large pool of unlabeled examples.
- Train a model on the labeled data.
- Use the model to score the unlabeled examples with a query strategy that estimates how informative each one would be.
- Send the top-ranked examples to the oracle for labeling.
- Add the new labels to the training set, retrain the model and repeat until the accuracy is good enough or the labeling budget is spent.
The key idea is that not all examples are equally useful. A photo that the model already classifies with near certainty teaches little, while one close to the decision boundary teaches a lot.
Query strategies
The choice of which examples to ask about is called the query strategy.
- Uncertainty sampling: pick the examples the model is least sure about. Measures include the lowest top probability, the smallest gap between the top two classes and the highest entropy.
- Query by committee: train several models and ask about the examples on which they disagree most.
- Expected model change: pick examples that would change the model’s parameters the most if labeled.
- Expected error reduction: pick examples that are expected to lower future error the most, a powerful but costly method.
- Diversity and density-based sampling: choose examples that represent the whole data distribution, so that the labeled set is not full of near duplicates or outliers.
In practice, strategies that mix uncertainty with diversity often work best, especially when labels are requested in batches.
Scenarios of active learning
Three settings are commonly described. In pool-based sampling, the learner sees a large pool of unlabeled data and picks from it, which is the most common case. In stream-based selective sampling, examples arrive one by one and the learner decides on the spot whether to request a label. In membership query synthesis, the learner creates new examples and asks for their labels, which can produce odd inputs that are hard for humans to judge, such as nonsense images.
Applications of active learning
Active learning is valuable wherever labels are costly. In medical imaging, it directs expert attention to the scans that matter most. In natural language processing, it speeds up labeling for named entity recognition, sentiment analysis and translation quality. In speech recognition, it picks which recordings to transcribe. It is used in drug discovery to choose which compounds to test in the laboratory, in materials science to guide experiments, and in content moderation to find borderline posts for reviewers. Data labeling platforms often include active learning so that annotators spend time on examples that improve the model, and language model teams use related ideas to select training data for human feedback.
Advantages and limitations
The main advantage is efficiency. Studies often show that active learning reaches a given accuracy with a fraction of the labels random sampling needs, saving money and expert time. It also directs attention to hard cases, which improves robustness. The limitations are real. The method can be biased, because the model’s own mistakes shape what it asks about, and a poor initial model can steer the process badly, a problem called the cold start. Selecting only uncertain examples can overlook large regions of data, and noisy labels from tired annotators can mislead the model, especially since uncertain examples are also hard for people. Retraining after every batch adds computing cost, and the gains over random sampling are not guaranteed for every dataset.
A worked example
A hospital wants a model to detect pneumonia in chest X-ray images. It has fifty thousand unlabeled images and funds for radiologists to review two thousand. It begins with two hundred random labeled images and trains a first model. The model scores all unlabeled images, and the system sends the three hundred images with the most uncertain predictions to radiologists. After they label them, the model is retrained and the cycle repeats. Midway, the team adds a diversity rule so that near duplicate images are not all selected. After several rounds, the model matches the accuracy that would have needed several times as many randomly chosen labels. The team still tests the final model on a separate random sample to check that the targeted sampling has not hidden weaknesses.
Active learning vs. passive learning
In ordinary, or passive, learning, the training examples are chosen at random or by whoever collected the data. In active learning, the learner takes part in selecting them. Active learning is not the same as active learning in education, which is a teaching approach in which students participate in lessons, even though the two terms share a name. It also differs from reinforcement learning, in which an agent acts in an environment and receives rewards instead of asking for labels.
Practical tips for active learning
Success depends on a few habits. Start with a small but balanced labeled set so that the first model is not hopelessly wrong. Request labels in batches instead of one at a time, because retraining after every label is slow and annotators work more efficiently on groups of items. Combine uncertainty with diversity so that a batch does not contain fifty near copies of the same confusing image. Keep a random sample labeled alongside the chosen ones, which gives an unbiased test set and a check that the targeted approach is working. Finally, make the interface easy for annotators, show them the model’s guess and give clear guidelines, since the examples chosen by active learning are the hardest ones and disagreement between reviewers will be higher than usual.
FAQs
Is active learning supervised or unsupervised? It is usually a form of supervised learning, because labels are eventually obtained, but it uses unlabeled data to decide what to label.
What is uncertainty sampling? It is the most common strategy, in which the model asks for labels on the examples where its prediction is least confident.
When should I not use active learning? When labels are cheap, when the dataset is small, or when you need an unbiased sample for statistics rather than a model.
More in Machine Learning
Key facts
- In active learning, the algorithm queries an oracle, usually a human annotator, for labels on selected unlabeled examples.1
- Common query strategies include uncertainty sampling, query by committee and expected model change.1
- The goal is to reach a target accuracy with fewer labeled examples than random sampling would need.2
Related concepts
Quick checkWhat is the main goal of active learning?Show answer
To reach good accuracy with as few human-labeled examples as possible by letting the model choose which examples to label.
Sources
- Settles, Active Learning Literature Survey, https://burrsettles.com/pub/settles.activelearning.pdf
- Wikipedia, Active learning (machine learning), https://en.wikipedia.org/wiki/Active_learning_(machine_learning)
- modAL, Active learning in Python, https://modal-python.readthedocs.io/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
