Knowledge from Uncertainty in Evidential Deep Learning

TL;DR

This paper investigates the evidential signal in Evidential Deep Learning (EDL), revealing its connection to misclassification bias and how it differs from other Dirichlet-based methods, especially in large language models.

Contribution

It provides empirical and theoretical analysis of EDL's evidential signal, highlighting its origins and differences from related approaches, and discusses the impact of training with out-of-distribution samples.

Findings

EDL's evidential signal correlates with misclassification bias.

Differences in loss functions affect uncertainty coupling.

Training with out-of-distribution samples influences EDL's uncertainty estimates.

Abstract

This work reveals an evidential signal that emerges from the uncertainty value in Evidential Deep Learning (EDL). EDL is one example of a class of uncertainty-aware deep learning approaches designed to provide confidence (or epistemic uncertainty) about the current test sample. In particular for computer vision and bidirectional encoder large language models, the `evidential signal' arising from the Dirichlet strength in EDL can, in some cases, discriminate between classes, which is particularly strong when using large language models. We hypothesise that the KL regularisation term causes EDL to couple aleatoric and epistemic uncertainty. In this paper, we empirically investigate the correlations between misclassification and evaluated uncertainty, and show that EDL's `evidential signal' is due to misclassification bias. We critically evaluate EDL with other Dirichlet-based approaches, namely Generative Evidential Neural Networks (EDL-GEN) and Prior Networks, and show theoretically and empirically the differences between these loss functions. We conclude that EDL's coupling of uncertainty arises from these differences due to the use (or lack) of out-of-distribution samples during training.