Active Learning
The short answer · 22 words
In short
Active learning is a machine learning approach in which the model selects for which data points it needs additional labels or feedback.
Flow-alerta.cab (2026): Active Learning. https://flow-alerta.cab/en/glossar/active-learning. Editorially reviewed on: 09 October 2026.
How much detail would you like?
Practical
Relevant when expert annotation is expensive or slow, such as quality inspection, document classification or image data. Its benefit depends strongly on the selection strategy, data distribution and label quality.
Possible example: A manufacturer owns thousands of product images. The model initially classifies only the most uncertain cases of skilled workers and is iteratively improved with these feedbacks.
Possible applications: efficient creation of domain-specific training data, document and image classification, quality control and human-in-the-loop learning processes.
At a glance
- Core
- Rather than requiring a large randomly labeled dataset, an active-learning method selects particularly informative examples and has an oracle, often a person, annotate them. Its aim is to achieve good model performance with less labeling effort.
- Use
- Relevant when expert annotation is expensive or slow, such as quality inspection, document classification or image data. Its benefit depends strongly on the selection strategy, data distribution and label quality.
- Example
- Possible example: A manufacturer owns thousands of product images. The model initially classifies only the most uncertain cases of skilled workers and is iteratively improved with these feedbacks.
- Limits
- Equate uncertainty with actual information gain; create distorted choice; adopt weak or inconsistent labels; include evaluation data in the selection; underestimate the effort required for human feedback.
In context
Questions about this term
Does Active Learning Always Need People?
No, the “Oracle” can also be another reliable source depending on the application.
Does it guarantee labels?
No, the advantage depends on the problem, strategy and data.
All details and background
Rather than requiring a large randomly labeled dataset, an active-learning method selects particularly informative examples and has an oracle, often a person, annotate them. Its aim is to achieve good model performance with less labeling effort.
ELI5 – Explained simply
The AI specifically asks for selected examples instead of indiscriminately labeling everything – for example in cases where it is particularly uncertain.
Why it matters in practice
Relevant when expert annotation is expensive or slow, such as quality inspection, document classification or image data. Its benefit depends strongly on the selection strategy, data distribution and label quality.
Example
Possible example: A manufacturer owns thousands of product images. The model initially classifies only the most uncertain cases of skilled workers and is iteratively improved with these feedbacks.
Pitfalls and limits
- Equate uncertainty with actual information gain; create distorted choice; adopt weak or inconsistent labels; include evaluation data in the selection; underestimate the effort required for human feedback.
Origin & history
In 2009, Burr Settles summarized the state of the literature on active learning in a survey; the author's version reviewed here was updated in 2010.
Use in mid-sized companies
- Possible applications: efficient creation of domain-specific training data, document and image classification, quality control and human-in-the-loop learning processes.
Related terms
Matching Flow-alerta service
AI opportunity analysis & roadmapAI potential analysis & roadmap: clarifying data quality and prerequisites for a possible first AI use case.Matching tool / use case
View use casesUse Case: Have particularly informative training examples labeled in a targeted manner; selecting and documenting labels in a comprehensible manner.Sources
- burrsettles.comSettles, Active Learning Literature Survey (2009, Autorenfassung 2010)
Quellenprüfung: 09.10.2026. Anwendungsbeispiele sind mögliche Szenarien, keine behaupteten Kundenfälle.
Editorially reviewed on: