Revisiting Essential and Nonessential Settings of Evidential Deep Learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Uncertainty Quantification

Background:

• Evidential Deep Learning (EDL) is a prominent method for estimating predictive uncertainty.
• EDL utilizes subjective logic to model class probability distributions via Dirichlet concentration parameters.
• Current EDL implementations contain nonessential settings that may hinder performance.

Purpose of the Study:

• To propose Re-EDL, a simplified and more effective variant of Evidential Deep Learning.
• To address limitations in EDL's model construction and optimization processes.
• To improve the reliability and accuracy of uncertainty estimation in deep learning models.

Main Methods:

• Re-EDL treats the prior weight parameter as an adjustable hyperparameter.
• It optimizes the expectation of the Dirichlet PDF directly.
• Variance-minimizing and KL-divergence regularization terms are deprecated.

Main Results:

• Re-EDL demonstrates improved performance compared to standard EDL.
• The method achieves state-of-the-art results on extensive experiments.
• Simplifying EDL enhances the balance between evidence proportion and magnitude.

Conclusions:

• Re-EDL offers a more effective approach to uncertainty estimation by refining EDL.
• The proposed simplifications lead to more reliable and accurate predictions.
• The adjusted methodology retains the core strengths of subjective logic in EDL.