New Plug-in Loss Framework Simplifies Evidential Deep Learning Uncertainty Estimation

Evidential Deep Learning (EDL) offers single-pass uncertainty estimation by modeling class probabilities via Dirichlet distributions, but its expected objectives are complex and hard to analyze. A new paper from TU Munich presents a simplified framework: plug-in losses, which approximate the EDL empirical risk by evaluating a standard loss (e.g., cross-entropy) at the predicted Dirichlet mean. Under mild assumptions, the authors prove the approximation error decays as the total evidence grows, for a broad class of loss functions including mean-squared error and cross-entropy. This analysis provides theoretical justification for using the softmax classifier within uncertainty estimation — a connection previously lacking in EDL.

Empirical validation on the Google Speech Commands dataset shows that the plug-in loss approach achieves predictive accuracy and selective prediction performance comparable to classical EDL, while using much simpler training pipelines (standard losses and optimizers). This is the first work to achieve coverage-accuracy trade-offs for speech recognition through EDL, making uncertainty estimation more accessible for real-world sensor systems. The framework bridges the gap between standard classification and evidential approaches, potentially enabling broader adoption of uncertainty-aware models in production environments.