SMuK 2023 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 63: ML Methods III

T 63.3: Talk

Wednesday, March 22, 2023, 16:20–16:35, HSZ/0405

Uncertainty aware training — Markus Klute, •Artur Monsch, Günter Quast, Lars Sowa, and Roger Wolf — Institute of Experimental Particle Physics (ETP), Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany

As physics experiments continue their measurements, with the LHC Run-3 and the future High-Luminosity LHC as notable examples, the amount of data is steadily increasing. These continued measurements will lead to reduced statistical uncertianties of many analyses, emphasizing the importance of systematic uncertainties in analysis results. This talk presents a machine-learning (ML)-based data analysis strategy to obtain an optimal test statistic minimizing analysis-specific statistical and systematic uncertainties. To achieve this the training objective for the neural network is modified to take systematic variations into account, leading to an overall uncertainty reduction on the analysis objective. The method will be demonstrated on a simple example using pseudo data and on a reduced CMS dataset used for an ML-based analysis of the observed Higgs boson in the di-τ final state with the goal of differential measurements of Higgs boson production, with the CMS experiment.