How to measure uncertainty in uncertainty sampling for active learning

Abstract Various strategies for active learning have been proposed in the machine literature. In uncertainty sampling, which is among most popular approaches, learner sequentially queries label of those instances its current prediction maximally uncertain. The predictions as well measures used to quantify degree uncertainty, such entropy, are traditionally a probabilistic nature. Yet, alternative approaches capturing learning, alongside with corresponding measures, recent years. particular, some these seek distinguish different sources and separate types reducible (epistemic) irreducible (aleatoric) part total prediction. goal this paper elaborate on usefulness compare their performance learning. To end, we instantiate sampling analyze properties thus obtained, them an experimental study.