A Simple and Explainable Method for Uncertainty Estimation using Attribute Prototype Networks

Claudius Zelenka, Andrea Göhring, Daniyal Kazempour, Maximilian Hünemörder, Lars Schmarje, Peer Kröger

Abstract: Deep learning’s utility in applications like medical diagnosis, autonomous driving, and natural language processing often hinges on the accurate estimation of uncertainty. Yet, conventional methods for uncertainty estimation face challenges, including high computational cost, difficulties with scalability, or poor interpretability. This paper presents a novel approach to uncertainty estimation using Attribute Prototype Networks (APNs), a method designed for learning robust and interpretable data representations. By leveraging prototype similarity scores, we propose a straightforward way to quantify the uncertainty of predictions, providing explainability and introducing a new technique for detecting out-of-distribution samples based on the distance to the nearest prototype. Our experiments demonstrate that this method offers valuable uncertainty information across several datasets. Our research opens up a new avenue for uncertainty estimation in deep learning, providing a simpler and more explainable solution.