Skip to content
Back to search results

Active Learning

Introduced by
Burr Settles
Year
2009
Type
Interactive supervised learning framework
Family
Interactive ML
Font size

What does it do?

Active learning is an iterative machine-learning paradigm in which a learning algorithm selectively queries an oracle — typically a human annotator — for labels on the most informative unlabeled examples.

Formalized by Burr Settles in his seminal 2009 literature survey, active learning addresses the practical bottleneck of annotation cost by achieving high model accuracy with far fewer labeled examples than passive supervised learning requires.

The intuition behind the method

Imagine a medical student who, instead of memorizing every textbook page, asks the attending physician only about the most puzzling cases. The student learns faster by focusing on uncertainty rather than reviewing examples they already understand. Active learning works the same way: the model identifies examples it is most confused about, requests human labels for those specific instances, and updates itself — repeating this loop until performance is satisfactory.

History

The concept of selectively querying informative training examples dates to Fedorov's optimal experimental design (1972) and Cohn et al.'s query learning (1994). The field was substantially unified and popularized by Burr Settles' 2009 technical report, which catalogued query strategies, theoretical bounds, and applications across NLP, bioinformatics, and computer vision. Deep active learning emerged after 2016 as neural networks became the dominant model class, prompting new uncertainty estimation techniques.

When to use it

Active learning is most beneficial when unlabeled data is abundant but expert annotation is expensive, slow, or scarce — common in medical imaging, legal document classification, scientific data curation, and natural language processing. It assumes oracle availability and that the query strategy can meaningfully identify informative examples. It is less suitable when annotation cost is negligible, the unlabeled pool is small, or the task distribution shifts substantially between queries. Alternatives include semi-supervised learning (when labels cannot be obtained interactively) and self-training (when oracle queries are infeasible).

Limitations and considerations

  • Query strategies can introduce sampling bias, causing the labeled set to be non-representative of the true data distribution.
  • Performance gains depend heavily on the choice of query strategy, which is task- and model-specific.
  • Retraining the model after each query is computationally expensive for large neural networks.
  • Oracle noise (annotator disagreement or errors) can degrade performance if not explicitly modeled.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Drastically reduces labeling cost by focusing annotation effort on maximally informative examples.
  • Applicable to any supervised learner — classification, regression, structured prediction — without modifying the underlying model architecture.
  • Theoretically grounded: uncertainty sampling and query-by-committee have PAC-learning and Bayesian justifications.
  • Compatible with crowdsourcing workflows and modern annotation platforms.

Applications

  • Medical image annotation (radiology, pathology) where expert labeling is costly and scarce.
  • Named-entity recognition and relation extraction in NLP, reducing hand-labeled corpus requirements.
  • Drug discovery: selecting which molecular candidates to synthesize and assay.
  • Remote sensing and satellite imagery classification with limited ground-truth labels.
  • Autonomous driving: prioritizing annotation of rare or safety-critical driving scenarios.

How it works

How it works

1

The process begins with a large pool of unlabeled examples U and an initially empty labeled set L. A base model may be pretrained on any available labeled seed data, or initialized randomly. This separation of labeled and unlabeled data is the defining structural assumption of pool-based active learning.

Formula
\mathcal{U} = \{\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_n\}, \quad \mathcal{L} = \emptyset
2

At each iteration the learner scores every unlabeled instance according to an informativeness criterion — here, uncertainty sampling selects the example whose maximum class posterior probability is lowest, indicating maximum model uncertainty. Alternative criteria include query-by-committee disagreement and expected model change.

Formula
\mathbf{x}^* = \arg\max_{\mathbf{x} \in \mathcal{U}} \phi(\mathbf{x}; \theta), \quad \phi(\mathbf{x}) = 1 - \max_{y} P_\theta(y \mid \mathbf{x})
3

The selected instance x* is presented to an oracle (human expert, crowdsourcing platform, or automated annotator) who provides a label y*. The newly labeled pair is moved from U into L, and the query budget counter is decremented. Oracle cost is typically treated as uniform per query.

Formula
y^* = \mathcal{O}(\mathbf{x}^*), \quad \mathcal{L} \leftarrow \mathcal{L} \cup \{(\mathbf{x}^*, y^*)\}, \quad \mathcal{U} \leftarrow \mathcal{U} \setminus \{\mathbf{x}^*\}
4

The model parameters are retrained (or fine-tuned) on the current labeled set L by minimizing an empirical loss. The updated model then re-scores the remaining pool, and the query-select-label-retrain cycle repeats until the budget is exhausted or a target accuracy threshold is reached.

Formula
\theta^* = \arg\min_\theta \frac{1}{|\mathcal{L}|} \sum_{(\mathbf{x},y) \in \mathcal{L}} \mathcal{L}_{\text{loss}}(f_\theta(\mathbf{x}), y)

Common pitfalls

  • Cold-start problem: with very few initial labels, uncertainty estimates are unreliable, leading to poor early queries.
  • Batch-mode active learning may select redundant instances if inter-query correlations are ignored.
  • Applying active learning to deep neural networks with dropout-based uncertainty (e.g., MC Dropout) can underestimate uncertainty in out-of-distribution regions.
  • Comparing active learning runs across seeds or query budgets without confidence intervals leads to misleading conclusions about strategy superiority.

Frequently asked questions

How is active learning different from semi-supervised learning?

Semi-supervised learning exploits the structure of unlabeled data (e.g., clustering, manifold assumptions) without querying an oracle, updating the model using pseudo-labels or consistency regularization. Active learning explicitly requests true labels from an oracle for selected examples. The two approaches are complementary and can be combined.

Which query strategy should I use in practice?

Uncertainty sampling is the most widely used baseline and works well for binary and multi-class classifiers. Query-by-committee is preferable when an ensemble is already available. Expected model change or BALD (Bayesian Active Learning by Disagreement) tend to perform better for deep models but are computationally heavier. Start with uncertainty sampling and benchmark against alternatives.

Can active learning be applied to regression tasks?

Yes. For regression, informativeness is typically measured by predictive variance (e.g., from a Gaussian process or ensemble) rather than class probability entropy. The learner queries points where variance is highest. This is standard practice in Bayesian optimization and surrogate-model-based design of experiments.

Related methods

Sources

  1. Settles, B. (2009). Active learning literature survey. University of Wisconsin-Madison Computer Sciences Technical Report 1648.

    Open publication reference
    Copied.