The integration of artificial intelligence (AI) into basic science is experiencing unprecedented momentum. However, specific challenges must be met: 

• any measurement or prediction must be accompanied by a precise confidence interval ;
• measurements are based on a wide range of data, each with inherent uncertainties and potential correlations;
• confidence in measurement must be communicated to peers;
• the complexity of an experiment lies somewhere between a junkyard game and an autonomous car: detailed, expensive and imperfect simulators are available;
• each experimental set-up is a unique, complex device, producing semi-structured data;
• costly data in the petabyte range.

These challenges have been met in the past for extraordinarily complex measurements where the role of AI was not major (for example, the discoveries of the Higgs boson or gravitational waves). On the other hand, fundamental science has a decades-old culture of Big Data.

Themes addressed

The aim of this workshop is to bring together experts in basic sciences, computer science and statistics to discuss the topic of uncertainties, which has been broken down into several themes: 

Monday

Opening session, Quantifying uncertainty, predicting uncertainty

Tuesday

Morning: Simulation-based inference
Afternoon: Explainable AI, Trustworthy AI

Wednesday morning

Data frugal approaches, data-centric AI, reference data and challenges, Fair Universe hackathon 

Thursday

Morning: unfolding (or de-biasing, de-flouting) 
Afternoon: Uncertainty control in generative models

Friday

Morning: architectures (Adversarial, Bayesian, ... )
Afternoon: closing session

Confirmed speakers

Luca Biferale (U Roma 2)
Jérôme Bobin (CEA)
Tommaso Dorigo (U Padova)
Gregor Kasiezcka (U Hamburg)
Mikael Kuusela (Carnegie Mellon)
Gilles Louppe (U Liège)
Mark Neubauer (U Illinois)