Evidence Regularization for Multimodal Deep Evidential Regression

Zhimin Shao, Weibei DOU

Keywords: Uncertainty estimation, Multimodal, Deep learning, Regression

Abstract: Uncertainty estimation is crucial in cost-sensitive areas, especially in the medical field, where multimodal information is common and effective. Existing studies have found the zero-confidence issue in unimodal settings, while the analysis in multimodal scenarios is lacking. In this work, we introduce the confidence paradox, where unimodal uncertainty is high but decreases after fusion, and present evidence regularization to tackle this issue. Initial results on the cubic and CT slice datasets show reduced root mean squared errors and improved detection of out-of-distribution samples, improving predictive reliability and training stability.

Submission Number: 88