Dortmund 2021 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 71: Data analysis, Information technology III

T 71.4: Vortrag

Mittwoch, 17. März 2021, 16:45–17:00, Tu

AI-safety for jet flavour tagging at the CMS experiment — Xavier Coubez^1,2, Nikolas Frediani¹, Spandan Mondal¹, Andrzej Novak¹, Alexander Schmidt¹, and •Annika Stein¹ — ¹RWTH Aachen University, Germany — ²Brown University, USA

Besides traditional Machine Learning techniques, Deep Learning has gained popularity in High Energy Physics in general. At the CMS experiment in particular, jet identification algorithms use Deep Neural Networks to classify the quark flavour from which the jet originates. The tagger learns to discriminate between heavy flavour quarks and light quarks.

The aim of AI safety studies is to test how susceptible neural networks are when mismodeling occurs in the simulation or when adversarial attacks are applied to the input data. Subtle mismodelings could be invisible to typical validation methods. In this talk, several methods to manipulate the input data and the impact on the performance of the tagging algorithm will be shown.