Active Learning (Machine Learning)
A training strategy where the model identifies which unlabeled examples would be most informative to label next, reducing the amount of labeled data needed to reach good performance.
Active learning is a machine learning training strategy in which the model plays an active role in selecting its own training data. Rather than training on a fixed labeled dataset, the model is iteratively trained and then queries a human (or another process) to label the examples it is most uncertain about or that would most improve its performance.
The key insight is that not all training examples are equally informative. A well-chosen set of labeled examples can produce a model nearly as good as one trained on a much larger randomly labeled set — which matters in science because generating high-quality labels is often expensive (expert annotation, experimental measurement, DFT calculation).
How it’s used in science:
- GNoME (materials discovery): DeepMind’s materials discovery model uses active learning to iteratively predict new crystal structures, synthesize the most promising predictions into its training data, and improve over successive rounds. This allowed it to explore a much larger chemical space than sequential screening would permit.
- Drug discovery: Active learning guides which compounds to synthesize and measure next, prioritizing experiments that will most improve a property prediction model.
- Systematic review screening: Active learning tools (Abstrackr, Rayyan AI) learn from the reviewer’s decisions to prioritize the remaining abstracts most likely to be relevant.
Not to be confused with: “Active learning” in education (student-centered pedagogy). The machine learning usage is unrelated.
Related guides: Materials Science, Health & Medicine
Related tools: Elicit