PHYSTAT informal review: Machine-learning based unfolding

by Anja Butter (Centre National de la Recherche Scientifique (FR)), Mikael Kuusela (Carnegie Mellon University (US))

Wednesday 15 Apr 2026, 15:30 → 17:00 Europe/Zurich

News: The event was postponed from October 22.

 

This is a PHYSTAT Informal Review event*. Today Anja Butter (physicist) together with Mikael Kuusela (statistician) will review the topic "Machine-learning based unfolding".  

Agenda:

• 3.30 pm Opening:
• 3.30 pm Physicists Presentation (20'+10')
• 4 pm Statisticians Presentation (20'+10')
• 4.30 pm General Discussion and Closing (30')

 

Abstract: 

Unfolding is the problem of inferring parton or particle-level distributions from observations smeared by detector effects. This is a challenging, ill-posed inverse problem with a long history in experimental particle physics. Traditionally, unfolding is done using explicitly regularized solutions for distributions discretized using histogram bins. In recent years, a large number of new machine learning-based unfolding methods have been proposed. Among their many benefits, these methods are typically binning-free, enable unfolding over high-dimensional spaces, and usually require only simulation-based access to the forward operator.

In this informal review, physicist Anja Butter and statistician Mikael Kuusela will share their perspectives on ML-based unfolding. Butter will first introduce the unfolding problem and review the landscape of ML-based unfolding methods, including their successes and challenges in real-world unfolding scenarios. Kuusela will then provide an overview of the statistical foundations of these methods, discuss their regularization properties, and highlight some caveats from a statistician's perspective.

*PHYSTAT informal reviews: In this virtual format, a Tandem consisting of a physicist and a statistician will review a statistical method introduced by one of the parties or a general critical analysis topic from the Physicist's and Statistician's perspectives. The virtual events comprise: two 20+10 min. complementary presentations followed by ~30 minutes of general discussion.