Active Learning
The model picks which unlabeled examples a human should label next, choosing the ones it is most unsure about, so you need far fewer labels.
Also query strategy · uncertainty sampling · human in the loop
A student who asks the teacher only about the homework problems they find confusing, instead of asking about every single one. They learn just as much with far fewer questions.
Where it fits
+All relationships as a list
How it works
Labels are often the most expensive part of ML: a radiologist’s time, a lawyer’s review. Active learning spends that budget where it helps most.
The loop:
- Train a model on the few labeled examples you have.
- Score all unlabeled examples. Pick the most informative ones, typically those the model is least sure about.
- Ask a human to label just those.
- Retrain and repeat.
Examples the model already gets right with high confidence teach it little. Examples near its decision boundary teach it a lot.
It is a close cousin of semi-supervised learning (which uses unlabeled data directly) and a core part of practical data labeling pipelines. One caveat: the actively chosen examples are not a random sample, so always evaluate on a separately drawn test set.
Everyday use cases
- Case 01
Medical image annotation
Radiologists label only the scans the model finds ambiguous instead of thousands of obvious ones.
- Case 02
Content moderation
Borderline posts go to human reviewers, and their decisions improve the model where it is weakest.
Strengths, weaknesses and when to use it
Strengths
- Can cut labeling costs dramatically when labels need experts.
Weaknesses
- Needs an annotation workflow and repeated retraining.
- Gains vary; random sampling is a surprisingly strong baseline.
When to use itWhen unlabeled data is plentiful and each label is expensive.
Compare with
Check yourself
- Q01
Which example would an uncertainty-sampling strategy send to a labeler first?
Show answer
B: One the model scores 0.51 cat