Active Learning

Query strategies that minimise labelling cost — uncertainty sampling, BALD, core-set, and pool-based loops for chemical discovery.

Courses

2 resources

Active Learning Literature Survey — Burr Settles

The definitive free monograph covering uncertainty sampling, query-by-committee, BALD, core-set, and stream-based vs. pool-based settings.

University of Wisconsin · Free PDF

Active Learning: A Survey — Aggarwal et al.

Book chapter covering the mathematical framework, query strategies, and stopping criteria for active learning.

Tutorials

2 resources

modAL Tutorials

scikit-learn-compatible active learning framework with notebooks demonstrating uncertainty sampling, query-by-committee, and Bayesian AL.

modal-python.readthedocs.io · Free notebooks

Small Data Drug Discovery — Pat Walters

Notebook demonstrating active learning for hit discovery using molecular fingerprints and a Gaussian Process surrogate.

GitHub · Colab notebook

Papers

4 resources

Houlsby et al. — "Bayesian Active Learning for Classification and Preference Learning" (2011)

Introduces BALD: selecting points that maximise mutual information between model parameters and predictions.

arXiv:1112.5745

Sener & Savarese — "Active Learning for Convolutional Neural Networks: A Core-Set Approach" (2018)

Core-set selection as a geometry-based AL strategy: select points that minimise the largest uncovered distance.

ICLR 2018 · arXiv:1708.00489

Reker & Schneider — "Active-learning strategies in computer-assisted drug discovery" (2015)

Chemistry-specific review of active learning applications in QSAR and virtual screening.

Drug Discovery Today 2015

Lookman et al. — "Active learning in materials science with emphasis on adaptive sampling in experimental design" (2019)

Review of AL applied to accelerate materials discovery via adaptive experimental design.

npj Computational Materials 2019

Tools

3 resources

Modular active learning framework for scikit-learn. Implements uncertainty sampling, query-by-committee, BALD, and expected error reduction.

Bayesian active learning library supporting MC-Dropout and deep ensembles for uncertainty estimation in deep learning models.

Comprehensive AL toolbox implementing 30+ query strategies with a unified interface. Good for benchmarking AL strategies.

Datasets

3 resources

134k molecules with 19 DFT properties. Its large size and known ground truth make it well-suited for simulating pool-based active learning loops.

20M DFT energies for small organic molecules across conformational space. Built via active learning — a real example of AL-driven dataset construction.

Isayev group · Free

Curated collection of molecular property datasets (ESOL, BBBP, Tox21, BACE) with standardised splits. Practical for AL experiments on real drug-like tasks.

DeepChem · Free