What Is Active Learning in Machine Learning?

What is active learning?

In machine learning, active learning is a strategy for labeling smarter, not more. Rather than labeling data at random, you train a model on what you have, let it score the remaining unlabeled pool, and send the most informative samples, typically the ones it is least certain about, to be labeled next. Those new labels are added, the model retrains, and the loop repeats. Because the model keeps asking for the examples that most improve it, you often reach the same accuracy with a fraction of the labels random sampling would need. This is the machine learning sense of the term, distinct from the education and teaching concept of the same name.

Key takeaways

The model selects which unlabeled samples to label next, focusing on the most informative ones.
It is a curation strategy, it changes what you label, not how you label it.
The payoff is reaching target accuracy with fewer labels, so lower cost and time.

Common selection strategies

Uncertainty sampling: pick the samples the model is least confident on.
Margin or entropy: pick samples where the top classes are closest or the prediction is most spread.
Diversity and representativeness: pick samples that cover under-represented regions, not just uncertain ones.
Hybrid: balance uncertainty with diversity to avoid labeling near-duplicates of the same hard case.

How it works, and how FiftyOne fits

The loop is: train, score the unlabeled pool, select the most valuable samples, label them, repeat. FiftyOne supports the selection step directly, use model uncertainty, embeddings, and similarity to find the samples worth labeling next, send them to annotation, and bring them back. That is active learning made operational.

Active learning vs related approaches

Decisions and label origins for active learning and related approaches

Decisions and label origins for active learning and related approaches
Approach	What it decides	Where the labels come from
Active learning	Which unlabeled samples to label next	Humans label the selected samples
Random sampling	Nothing, samples are chosen at random	Humans label a random subset
Auto-labeling	Not what to label, but how	A model generates labels, humans review

Why it matters

Active learning attacks the most expensive part of ML, labeling, by spending the budget only where it moves the model. Pure uncertainty sampling has a well-known trap, it tends to select outliers and near-duplicate hard cases, so you label the same confusing thing over and over and waste budget. That is why diversity-aware selection usually beats raw uncertainty, a point formalized by the core-set view of batch active learning (Sener and Savarese, 2018) and grounded in the broader query-strategy literature (Settles, 2009), and why embeddings, which spread samples by similarity, are the practical backbone of active learning rather than confidence scores alone.

Frequently asked questions