evidential-learning

This paper introduces Evidential Adversarial Training (EV-AT), a method that improves the robustness-uncertainty trade-off in classifiers by combining an evidence-based loss with robust evidence alignment, achieving state-of-the-art results on selective classification benchmarks.