A collaborative research project led by Dr Nafise Sadat Moosavi (School of Computer Science), in partnership with Dr Benjamin Heinzerling (RIKEN AIP) and Prof Kentaro Inui (Tohoku University & RIKEN AIP), has been awarded £215,560.00 (UK share £193,560.00) through the International Collaboration Awards 2024 (ISPF). The project, titled Geometric Representations of Uncertainty for Foundation Models, aims to address a critical challenge in artificial intelligence: ensuring that AI systems produce reliable and trustworthy outputs by improving how they quantify and manage uncertainty.
Foundation models, which power many of today’s AI systems for generating text, images, and multimodal content, often produce high-quality outputs that can appear entirely accurate - even when they are not. This overconfidence can lead to misuse in critical domains where precision is essential, such as healthcare and decision-making. Current approaches to uncertainty quantification (UQ) in AI systems are often limited, struggling with scalability and integration into complex neural networks.
To address these challenges, the team will develop a novel, multi-dimensional geometric approach to uncertainty modelling. By representing both epistemic and aleatoric uncertainties, their method aims to provide a comprehensive uncertainty profile for each model prediction. This will enhance the interpretability and robustness of AI systems while improving their ability to balance creativity with factual accuracy. The research is designed to generalise across various AI applications, making it a scalable and impactful solution.
The funding will support a three-year postdoctoral researcher, based at Sheffield and jointly supervised by the collaborators. By addressing a fundamental weakness in current AI systems, this research has the potential to set new standards for trustworthiness and transparency in artificial intelligence, enabling its safe and effective use in a wide range of applications.
Congratulations to Dr. Moosavi and the team on this remarkable achievement!