Speaker
Description
Neural Simulation-Based Inference (NSBI) is a family of emerging techniques that enable statistical inference with high-dimensional data, even when exact likelihoods are analytically intractable. These techniques use deep learning to build likelihood- or posterior-based inference models directly from simulations. By reducing reliance on hand-crafted, low-dimensional summary observables, NSBI can retain more information and improve sensitivity in precision measurements and searches, as demonstrated across several scientific domains.
This talk will introduce the basics of NSBI and review recent applications in ATLAS and CMS, while focusing on a key practical challenge for full-scale adoption to these experiments: the scalable treatment of O(10^2) nuisance parameters encoding systematic uncertainties. Starting from a fully non-parametric approach to NSBI commonly used across scientific domains, the talk will build towards domain adapted semi-parametric models that make NSBI scalable for LHC-style experimental models. It will show how these approaches address sparse coverage of nuisance parameter space common in ATLAS and CMS statistical models, keeping training and inference tractable at the LHC scale, while retaining robustness essential to publishing a full measurement.