Possibilistic Predictive Uncertainty for Deep Learning

The article presents Dirichlet-approximated possibilistic posterior predictions (DAPPr), a framework for modeling epistemic uncertainty in deep learning that utilizes possibility theory. DAPPr introduces a possibilistic posterior over parameters and employs supremum operators for projection to the prediction space, resulting in a training objective with closed-form solutions. This approach demonstrates competitive or superior performance in uncertainty quantification compared to existing second-order predictors, offering a balance of principled derivation and computational efficiency, which is crucial for practitioners addressing overconfidence in neural network predictions.