Learn · Beginner
Active learning: letting the model choose what to label next
Active learning is a training strategy where the model chooses which examples get labeled next. Instead of labeling a random sample of data and hoping it covers what matters, you train on a small seed set, ask the model which unlabeled examples it finds most uncertain, send just those to a human or an instrument for labeling, and repeat. The point is efficiency: when labels are expensive and raw data is cheap, the order you label things in changes how much you learn per dollar.
The economics that motivate it are everywhere. A hospital has millions of scans and a handful of radiologist-hours. A biology lab has a design space of billions of candidate proteins and a bench that can test a few hundred a week. A content platform has endless posts and a small review team. In every case the bottleneck is not data, it is annotation, and random sampling spends that budget on examples the model already handles correctly.
How the loop works
The cycle has four steps. Train a model on whatever labeled data you have. Run it over the unlabeled pool and score every example by how much labeling it would help. Send the top-scoring examples for labeling. Add them to the training set and go around again.
Everything interesting is in the scoring rule, called the acquisition function. The oldest and still most common is uncertainty sampling: pick the examples where the model's prediction is closest to a coin flip, on the reasoning that a confident correct prediction teaches nothing and a genuinely ambiguous case sits near the decision boundary the model is trying to find. Variants measure uncertainty as low margin between the top two classes, high entropy across all classes, or disagreement among an ensemble of models -- an approach known as query-by-committee, where you label the examples your models argue about.
The teaching analogy is a good student with limited study time. Re-reading the chapters you already understand feels productive and teaches you nothing. The efficient move is to find the problems you get wrong half the time and work those. Active learning is that instinct made into an algorithm.
The complication: batches and diversity
Pure uncertainty sampling breaks in practice, and the reason is worth understanding. Labeling one example at a time and retraining is far too slow, so real systems select a batch of hundreds at once. But the most uncertain examples tend to be uncertain for the same reason -- they cluster. You end up paying for five hundred near-identical hard cases and learning roughly what one of them would have taught you.
The fix is to combine uncertainty with diversity, selecting a batch that is both informative and spread across the data. Core-set approaches choose points that cover the feature space; gradient-based methods like BADGE pick examples whose expected updates to the model point in different directions. This is the main practical difference between the textbook version of active learning and one that works.
Where it shows up now
Active learning has quietly become central to two modern areas. The first is autonomous experimentation. When an AI agent runs a physical instrument -- iterating on a liquid-handling parameter, screening protein designs, tuning a laser -- it is doing active learning with the world as the labeling oracle: propose the experiment whose result is least predictable, run it, update, repeat. Anthropic's hardware standard for AI-operated instruments is infrastructure for exactly this loop, and the closed-loop optimisation it demonstrates is active learning wearing a lab coat. It is closely related to Bayesian optimization, which formalises the same idea when each experiment is very expensive.
The second is preference data for language models. Human preference labels are the costly ingredient in reinforcement learning from human feedback, and asking annotators to compare responses the reward model already scores confidently is waste. Selecting the comparisons where the model is genuinely torn is active learning applied to alignment.
The honest failure modes
Active learning is not free. It depends on the model's uncertainty being meaningful, and neural networks are notoriously overconfident, so the technique inherits every problem covered in calibration and confidence. It has a nasty affinity for garbage: the examples a model is least sure about are often the ones that are mislabeled, corrupted, or genuinely ambiguous to humans too, so an uncertainty-driven loop can spend its whole budget on noise. And the selected training set is deliberately not a random sample, which means it is biased by construction -- fine for training, misleading if you then try to estimate real-world accuracy from it. Always keep a separate, randomly sampled evaluation set, or you will have optimised your way into a number that does not transfer, which is a close cousin of the problem described in benchmark contamination.
Active Learning Literature Survey (Settles, 2009)
A Sequential Algorithm for Training Text Classifiers (Lewis and Gale, 1994)
Deep Bayesian Active Learning with Image Data
Active Learning for Convolutional Neural Networks: A Core-Set Approach
Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds (BADGE)
Key questions
What problem does active learning solve?
How does the model decide what to ask for?
When does active learning fail?
Cite this
APA
Ground Truth. (2026, August 27). Active learning: letting the model choose what to label next. Ground Truth. https://groundtruth.day/learn/active-learning.html
BibTeX
@misc{groundtruth:active-learning,
title = {Active learning: letting the model choose what to label next},
author = {{Ground Truth}},
year = {2026},
month = {aug},
url = {https://groundtruth.day/learn/active-learning.html}
}