BALSA: Active Learning for Regression by Distribution Disagreement
A pool-based active-learning method for regression models that output full predictive distributions (normalizing flows, Gaussian neural networks). BALSA extends the BALD acquisition rule by measuring disagreement between Monte-Carlo-dropout-sampled predictive distributions directly, using KL-divergence or Earth Mover's (Wasserstein) distance rather than collapsing them to scalar entropy or standard deviation. This better isolates epistemic from aleatoric uncertainty and yields state-of-the-art acquisition performance across four real-world regression datasets.
BAYESIAN ACTIVE LEARNING BY DISTRIBUTION DISAGREEMENT Thorben Werner ∗
Tackles pool-based Active Learning for regression, where measuring model uncertainty is harder than in classification. Using models with full predictive distributions (Gaussian neural networks and co…