Explain EDL degradation on transformer-based architectures

Investigate why Evidential Deep Learning exhibits degraded performance on transformer-based architectures, particularly ViT3D on the NoduleMNIST3D dataset, and determine how to improve its stability in these settings.

Background

The paper evaluates Monte Carlo Dropout and Evidential Deep Learning across convolutional and transformer-based three-dimensional backbones for lung nodule classification. Although uncertainty-driven training generally preserves classification performance and improves calibration, EDL performs substantially worse on some attention-based architectures. The clearest example is ViT3D on NoduleMNIST3D, where EDL achieves an AUC of 0.50 and an F1 score of 0.70, compared with 0.83 and 0.83, respectively, for the baseline.

The authors identify the cause of this degradation as unresolved and suggest that the interaction between the Dirichlet evidence parameterisation, attention-based architectures, and lower-resolution input volumes requires further examination. Understanding this behavior is necessary for reliable application of evidential uncertainty estimation beyond convolutional backbones.