AI glossary · Advanced · Data & evaluation · Classic ML, recsys & search
Active Learning
Active learning is a training approach in which the model picks out data it judges to be "confusing" and asks a person to label it with the correct answer. The goal is to improve performance efficiently while reducing labeling costs.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Foundations & applications
- Added
- 2026-10-09
How it works
First, the model is trained on a small amount of labeled data. Next, the model is run on a large amount of unlabeled data, and the items it is least confident about, or that differ most from the existing data, are selected. A person labels only the selected data, which is then added and the model is retrained. This process is repeated until performance is good enough.
The active learning loop
Until performance is good enough
- 1Train model
- 2Make predictions
- 3Pick confusing items
- 4Human labeling
- 1. Train model On labeled data
- 2. Make predictions On unlabeled data
- 3. Pick confusing items Low-confidence or unfamiliar data
- 4. Human labeling Only the selected items
Why it matters
In fields where only experts can provide the correct answers, such as medical image reading or legal document classification, each label is expensive. When you cannot label all the data, labeling the data that helps the model most first is how you build a better model on the same budget. It also saves the effort of labeling easy data the model already gets right.
Things to keep in mind
If you only pick data the model finds confusing, many outliers or noisy items may be selected. Also, if the selected data is concentrated in one area, it will not represent the whole dataset evenly, which introduces bias. For this reason, methods that consider both "confusion" and "diversity" when selecting are commonly used. Because a person must take part in every round, it is also important to have a well-organized labeling workflow.
Example
Many data labeling tools have a feature where the model makes predictions first and only low-confidence items are shown to annotators. For example, when building a defect inspection model for a manufacturing plant, if inspectors check only the product photos the model found ambiguous between normal and defective, out of a large number of photos, the model can be improved quickly with little labeling work. The same idea is used in data flywheels, which collect answers corrected by users in a live service and retrain on them.