Multimodal learning holds the potential to transform machine learning by uncovering dependencies and correlations between different data sources. However, current systems often fail to fully exploit these modalities due to the phenomenon of modality competition, where multiple modalities compete for learning resources, resulting in some being underutilized. This seminar will delve into the challenges of modality competition, exploring its characteristics and potential sources, and discuss various proposed solutions, along with their limitations, to achieve balanced training across modalities and leverage their combination to enhance overall model performance.
Konstantinos Kontras originates from Heraklion, Crete, in Greece. He obtained a diploma in Electrical and Computer Engineering in 2018 at the University of Patras in Greece. In 2020, he obtained a postgraduate degree in AI engineering at KU Leuven. In 2025, he received his PhD in the Biomed lab of the STADIUS Center of the Electrical Engineering Department at KU Leuven, where he now continues as a PostDoc. His research focused on enhancing multimodal fusion by refining model architectures and training objectives for deep learning networks.
