Prediction-Oriented Bayesian Active Learning
This addresses a problem for machine learning practitioners by offering an improved drop-in replacement for active learning, though it is incremental as it builds on existing information-theoretic approaches.
The paper tackles the suboptimality of information-theoretic active learning methods like BALD for predictive performance by proposing the expected predictive information gain (EPIG) acquisition function, which focuses on information gain in prediction space, and shows that EPIG leads to stronger predictive performance compared to BALD across various datasets and models.
Information-theoretic approaches to active learning have traditionally focused on maximising the information gathered about the model parameters, most commonly by optimising the BALD score. We highlight that this can be suboptimal from the perspective of predictive performance. For example, BALD lacks a notion of an input distribution and so is prone to prioritise data of limited relevance. To address this we propose the expected predictive information gain (EPIG), an acquisition function that measures information gain in the space of predictions rather than parameters. We find that using EPIG leads to stronger predictive performance compared with BALD across a range of datasets and models, and thus provides an appealing drop-in replacement.