Grigor Bezirganyan

Modern AI models are remarkably accurate, and remarkably overconfident. A model that mislabels a tumor, mistranslates a legal clause, or misreads a road sign will often do so with the same confidence as its correct predictions. As AI moves deeper into healthcare, autonomous systems, and production pipelines, the relevant question is no longer only "what does the model predict?" but "how much should we trust this prediction?"

This talk introduces Uncertainty Quantification (UQ) in AI through a practical and accessible lens. We will begin with why UQ matters today, drawing on real failure modes from healthcare diagnostics and from my current work in translation quality estimation, where knowing when a model is unsure is often more valuable than the prediction itself.

From there, we will build intuition for how UQ actually works. The talk covers the core ideas behind Evidential Deep Learning and Subjective Logic, frameworks that allow models to express not just a prediction but also their degree of belief, disbelief, and uncertainty. We will see how this differs from traditional softmax confidence, and why that difference matters in deployment.

The second half focuses on multimodal UQ, where models combine inputs from multiple sources such as images, text, sensors, and audio. Multimodality introduces new challenges: what happens when modalities disagree? Standard fusion methods tend to average these conflicts away, hiding information that is often critical. I will present my research on conflict-aware multimodal UQ, using visualizations and results to show how explicitly modeling disagreement between modalities leads to more honest and more useful uncertainty estimates. The talk closes with a broader discussion on the future of trustworthy AI in the context of multimodal agents.