Credal Set

Credal sets represent uncertainty by encompassing a set of probability distributions, rather than a single one, offering a robust framework for handling both aleatoric (inherent randomness) and epistemic (knowledge-based) uncertainty. Current research focuses on developing methods for learning and applying credal sets in various machine learning contexts, including model averaging, conformal prediction, and multi-armed bandit problems, often leveraging techniques from optimal transport and geometric deep learning. This approach improves uncertainty quantification and prediction accuracy, particularly in challenging scenarios like out-of-distribution detection and multi-label classification, with implications for reliable and explainable AI.