Karlsruhe 2024 – scientific programme

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

T: Fachverband Teilchenphysik

T 84: Methods in particle physics 5 (tagging)

T 84.5: Talk

Thursday, March 7, 2024, 17:00–17:15, Geb. 20.30: 2.066

Overview of adversarial studies for heavy flavour tagging — •Alexander Jung, Ming-Yan Lee, Uttiya Sarkar, Alexander Schmidt, Hendrik Schönen, Jan Schulz, and Ulrich Willemsen — III. Physikalisches Institut A, RWTH Aachen University, Germany

Neural networks have become indispensable in jet tagging algorithms. The ever-increasing performance in classifying jets comes with the disadvantage that these algorithms are susceptible to mismodeled input data. Networks are trained on simulated samples with a fixed detector setup. The real setup is not always constant, e.g. misalignment can occur or parts of the detector can fail. However, it would not be feasible to take these variations into account in the simulation, which means that mismodeling occurs "by design". In this contribution, we will look at how neural networks react to mismodeling, i.e. how robust they are against them and how their robustness can be improved.

Keywords: Neural Network; AI Safety; CMS; Heavy Flavour Tagging; Machine Learning