Reproduce “Evidential Uncertainty Quantification: A Variance-Based Perspective” on OpenResearch

About this paper

Ruxiao Duan, Brian Caffo, Harrison X. Bai, Haris I. Sair, Craig Jones

This research from Johns Hopkins University introduces a variance-based approach for uncertainty quantification in Evidential Deep Learning (EDL) classification, enabling the estimation of class-level uncertainties and novel between-class correlations. The proposed method demonstrates competitive classification accuracy in Active Domain Adaptation (ADA) while providing more granular insights into model confidence and confusion patterns.

Key insights

• Variance-based uncertainty quantification, traditionally used in Evidential Deep Learning (EDL) regression, can be successfully adapted to EDL classification by using the Law of Total Covariance.
• The proposed framework provides fine-grained class-level uncertainties (total, aleatoric, epistemic) and novel between-class correlations, offering deeper insights into model confidence and confusion patterns than prior entropy-based methods.
• A simultaneous certainty and uncertainty sampling strategy improves Active Domain Adaptation performance by strategically incorporating both highly confident pseudo-labeled samples and uncertain samples for human annotation.

Summary by alphaXiv.