Presenter: Dimitrios Kollias, Queen Mary University of London
Date: 07 January 2026

The seminar will outline a research agenda for robust and trustworthy multimodal AI, centered on advances in representation learning, multi-objective optimization, causality and fairness under distribution shift. We argue that progress in multimodal learning requires moving beyond empirical performance toward models that explicitly account for structure, dependencies and invariances across modalities and tasks. I will discuss unifying principles for learning aligned and interpretable representations, handling conflicting objectives and partial supervision, and managing trade-offs between utility, robustness and fairness, with particular emphasis on the interplay between domain generalization and algorithmic fairness. These ideas are empirically validated in the context of behavior understanding and analysis, affective computing and digital humans, illustrating how principled multimodal learning can enable reliable and deployable intelligent systems.

Biography

Dimitrios Kollias is an Associate Professor in multimodal AI at Queen Mary University of London, UK. He serves as a consultant, academic advisor, and/or scientific collaborator to multiple companies, including Toyota Motor Europe, Hume AI, GRNET, WPP, and Remark AI. He is a Responsible AI Mentor for Digital Catapult’s High Growth AI Accelerator (Innovate UK BridgeAI programme) and a Mentor in the IEEE UK & Ireland Women in Engineering (WIE) Connect Mentorship Programme. He is a Fellow of the Higher Education Academy, holds a Postgraduate Certificate, and is a member of IEEE, BMVA, AAAI, ACM, as well as an associate member of IARP.