Trustworthy AI at Runtime: Out-of-Distribution and Robustness in Open Worlds

12^th May 2026, 13:00 Brodie Tower Room 106
Changshun Wu
Université Grenoble Alpes

Abstract

Modern deep learning systems often perform well in controlled training environments, but their reliability degrades after deployment in open-world settings, where two fundamental gaps emerge: an epistemic gap that leads to “I don’t know” failures on out-of-distribution inputs, and a stability gap that makes predictions brittle under perturbations. In this talk, I present a runtime perspective on trustworthy AI that addresses these two gaps through deployment-time mechanisms rather than training-time redesign. For the epistemic gap, I first introduce a lightweight runtime monitoring framework for out-of-distribution detection that requires no modification to model architectures or training procedures, yet provides effective detection and filtering of unreliable predictions in real-world autonomous driving scenarios. I then present a complementary approach that moves beyond external detection by introducing a mitigation strategy for overconfident predictions, which improves robustness to distribution shift and is compatible with standard vision architectures including YOLO, Faster R-CNN, and DETR. For the stability gap, I turn to the problem of prediction robustness under perturbations. We take a first step toward extending randomized smoothing beyond its classical classification setting into generative models, from a runtime perspective, to explore its role in improving predictive stability. All these results highlight runtime interventions as a useful perspective for studying trustworthiness in open-world deployment, including monitoring, mitigation, and certification.

Biography

Changshun Wu is currently a researcher-engineer at the University of Grenoble Alpes and the Verimag Laboratory in France. His research focuses on developing safe and interpretable AI systems by combining interpretability methods, formal verification for safety guarantees, robustness evaluation, and runtime monitoring for detecting anomalies and out-of-distribution inputs.