Conformal Prediction and Uncertainty Wrapper: What Statistical Guarantees Can You Get for Uncertainty Quantification in Machine Learning?

With the increasing use of Artificial Intelligence (AI), the dependability of AI-based software components becomes a key factor, especially in the context of safety-critical applications. However, as current AI-based models are data-driven, there is an inherent uncertainty associated with their outcomes. Some in-model uncertainty quantification (UQ) approaches integrate techniques during model construction to obtain information about the uncertainties during inference, e.g., deep ensembles, but do not provide probabilistic guarantees. Two model-agnostic UQ approaches that both provide probabilistic guarantees are conformal prediction (CP), and uncertainty wrappers (UWs). Yet, they differentiate in the type of quantifications they provide. CP provides sets or regions containing the intended outcome with a given probability, UWs provide uncertainty estimates for point predictions. To investigate how well they perform compared to each other and a baseline in-model UQ approach, we provide a side-by-side comparison based on their key characteristics. Additionally, we introduce an approach combining UWs with CP. The UQ approaches are benchmarked with respect to point uncertainty estimates, and to prediction sets. Regarding point uncertainty estimates, the UW shows the best reliability as CP was not designed for this task. For the task of providing prediction sets, the combined approach of UWs with CP outperforms the other approaches with respect to adaptivity and conditional coverage.