Epistemic uncertainty in AI for science

Abstract

Although artificial intelligence (AI) has recently achieved astonishing results, its inability to deal with uncertainty severely limits its future applications. Current AI cannot confidently make predictions robust enough to stand the test of data generated by processes different (even by tiny details, as shown by ‘adversarial’ results) from those seen at training time. While recognising this issue under different names (e.g. ‘overfitting’ or 'domain adaptation'), traditional machine learning seems unable to address it in non-incremental ways. The issue is all the more felt in the rising field of AI for science and engineering, which has seen incredible successes in the past few years.

Our "epistemic AI" concept re-imagines AI from the foundations, by properly treating the "epistemic" uncertainty stemming from a machine's partial knowledge of the world. The aim is to create a new learning paradigm designed to provide worst-case guarantees on its predictions, thanks to a proper modelling of real-world uncertainties. As we discuss in this talk, epistemic learning can be extended to neural operators networks, which can solve entire classes of systems of partial differential equations, providing a step change in the way AI is used to model scientific and engineering problems.

Bio

Fabio Cuzzolin was born in Jesolo, Italy. He received the laurea degree magna cum laude from the University of Padova, Italy, in 1997 and a Ph.D. degree from the same institution in 2001, with a thesis entitled “Visions of a generalized probability theory”. He was a researcher with Politecnico di Milano in Milan, Italy, and a postdoc with the UCLA Vision Lab at the University of California at Los Angeles, led by Prof Stefano Soatto. He later joined the Perception team at INRIA Rhone-Alpes, Grenoble, as a Marie Curie fellow. He joined Prof Phil Torr's vision group at the Department of Computing of Oxford Brookes University in September 2008, and has been a Professor of Artificial Intelligence since January 2016. In 2012 he funded the Visual Artificial Intelligence Laboratory (VAIL), currently funded by the European Union, URKI, Innovate UK, the British Council and the Leverhulme Trust for around £3M. Since 2024 he is the inaugural Director of the Institute for AI, Data Analysis and Systems (AIDAS) at Oxford Brookes University, and the university's academic liaison with the Turing University Network.

Fabio is a world leader in the field of imprecise probabilities and random set theory, to which he contributed an original geometric approach. His Lab's research spans artificial intelligence, machine learning, computer vision, surgical robotics, autonomous driving, AI for healthcare as well as uncertainty theory. The team is pioneering frontier topics such as machine theory of mind, epistemic artificial intelligence, evolving and universal AI, neural operators under uncertainty, neurosymbolic reasoning and continual semi-supervised learning. He is the Coordinator of the H2020 FET Open project "Epistemic AI" (E-pi), and was Scientific Officer for the H2020 project 779813-SARAS (Smart Autonomous Robot Assistant Surgeon). Fabio is the author of circa 160 publications, published or under review, including 5 books and around 45 journal papers and book chapters, with a total impact factor of around 300. His work won a Best Paper Award at PRICAI’08, a Best Poster at ISIPTA’11 and the 2012 MLVR Summer School, a nomination for Best Paper and an Outstanding Reviewer Award at BMVC’12, a Best Paper award at the IJCAI 2022 AI4AD workshop and recently the Best Student Paper prize at IJCLR 2022.