1st – 3rd September 2025, EURECOM, France
ProbNum25 is an international conference on Probabilistic Numerics, methods for statistically solving numerical problems (optimisation, integration, solving differential equations) and probabilistically quantifying the numerical errors as computational uncertainties. Probabilistic Numerics make numerical algorithms faster, more reliable, and easier to design and use. They are developed and used in machine learning, artificial intelligence, scientific simulation and computational statistics. More details can be found below.
ProbNum25 welcomes researchers and practitioners interested in methods, theory and applications of Probabilistic Numerics for an open exchange of ideas.
Venue, Dates and Registration
ProbNum25 will be held at EURECOM near Nice in southern France from 1-3 September 2025 (Details). The event is open to all. We expect to open registrations on 14th May 2025.
Publications as PMLR Proceedings
ProbNum25 calls for papers that will be published as Proceedings of Machine Learning Research (PMLR). The submission deadline is currently scheduled for 5th March 2025 (Details).
Is ProbNum 2025 for me?
A probabilistic numerical method is an algorithm that quantifies errors arising from the finite nature of computation. In this context, the process of computation is often interpreted in the language of probabilistic (or Bayesian) inference, where computation is treated as a source of information, much like data in statistics and machine learning. Probabilistic numerical methods are thus also called computation aware. While such algorithms may in some cases use random numbers, stochasticity is a concept complementary to probabilistic inference: Probabilistic numerical methods are not necessarily stochastic numerical methods, and many stochastic numerical methods are not computation-aware. Instead, probabilistic numerical methods typically return probability measures, e.g. parametrized through sufficient statistics or moments. Beyond the basic goal of quantifying computational uncertainty, the value of probabilistic functionality is that it uses the same mathematical concepts and code functionality as statistics and machine learning, which simplifies overarching goals like inference on parameters or latent forces in scientific simulation (aka. inverse problems), joint inference on a physical system from both mechanistic and empirical data (aka. data assimilation), the adaptive control of algorithmic cost, or more generally the use of numerical methods inside of machine learning systems. In this sense, probabilistic numerical methods provide a native algorithmic formalism for AI, ML, and Statistics.
ProbNum25 aims to be an inclusive, enjoyable venue for the discussion, dissemination and promotion of such ideas. We welcome submissions from researchers in academia and industry, from theory to applications. We also invite interested participants who want to learn more about the field without submitting their own yet to attend. Dedicated tutorials will provide introductions and overviews of multiple areas.
Past Meetings
ProbNum25 is the latest edition of a series of events on Probabilistic Numerics since 2012. Past events can be found here.
Organisers
- Motonobu Kanagawa (EURECOM, France)
- Jon Cockayne (University of Southampton, UK)
- Alexandra Gessner (Astrazeneca, Spain)
- Philipp Hennig (University of Tuebingen, Germany)
Contact Information
Any questions can be sent to the organisers above, specifically as an email to motonobu.kanagawa@eurecom.fr with a title containing „ProbNum25.’’