Active Learning Theory

Recording

Learning Objectives

• Formalise the pool-based active learning setting and contrast it with stream-based and membership-query paradigms
• Implement uncertainty sampling, margin sampling, and entropy-based query strategies
• Explain the BALD acquisition function and its connection to mutual information
• Design a realistic active learning loop for a DFT screening campaign

Key Takeaways

• Takeaway 1. Random sampling is a surprisingly strong baseline for active learning in chemistry — always benchmark against it before claiming that a fancy query strategy helps.

• Takeaway 2. Uncertainty sampling selects the points the model is least confident about, but high uncertainty near the training distribution boundary is not the same as informativeness — diversity-promoting methods (core-set, BADGE) often do better.

• Takeaway 3. BALD (Bayesian Active Learning by Disagreement) maximises the mutual information between model parameters and predictions; it is theoretically principled but computationally expensive for large pools.

• Takeaway 4. In chemistry, the cost of labelling (DFT calculation, synthesis, assay) dwarfs the cost of the model. Even a modest reduction in the number of required experiments translates directly to saved resources.

Further Reading & Resources