Making uncertainty estimates simpler without losing accuracy

Evidential Deep Learning estimates uncertainty by predicting Dirichlet distributions over class probabilities, but the math is complex and hard to implement. This paper replaces the complicated Dirichlet objective with simpler plug-in losses (like cross-entropy) evaluated at the mean, proving the approximation error shrinks reliably. Testing on speech recognition shows the simplified approach matches classical EDL's accuracy and uncertainty calibration while using standard training pipelines.