Generalized Regularized Evidential Deep Learning Models: Theory and Comprehensive Evaluation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Evidential deep learning (EDL) models, grounded in Subjective Logic, enhance deterministic neural networks with uncertainty quantification.
• Current EDL models face challenges due to non-negative evidence constraints, leading to activation-dependent learning-freeze behavior and diminished gradients in low-evidence regions.

Purpose of the Study:

• To theoretically analyze the activation-dependent learning-freeze behavior in EDL models.
• To develop novel activation functions and regularizers for consistent evidence updates across different activation regimes.

Main Methods:

• Theoretical characterization of learning-freeze behavior in EDL models.
• Design of a generalized family of activation functions and corresponding evidential regularizers.
• Empirical validation on benchmark image classification, few-shot learning, and blind face restoration tasks.

Main Results:

• The study theoretically explains the learning-freeze phenomenon in EDL models.
• The proposed generalized regularized evidential models demonstrate consistent evidence updates.
• Experiments show the effectiveness of the new approach across diverse computer vision tasks.

Conclusions:

• The developed activation functions and regularizers provide a robust alternative for EDL models.
• This work addresses key limitations, enabling more reliable uncertainty quantification in deep learning.
• The findings pave the way for improved performance in AI systems requiring nuanced uncertainty estimation.