guest ::
TY - THES AU - Stein, Annika TI - Novel jet flavour tagging algorithms exploiting adversarial deep learning techniques with efficient computing methods and preparation of open data for robustness studies PB - RWTH Aachen University VL - Dissertation CY - Aachen M1 - RWTH-2024-07840 SP - 1 Online-Ressource : Illustrationen PY - 2024 N1 - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University N1 - Dissertation, RWTH Aachen University, 2024 AB - Machine learning algorithms are an indispensable tool for science. Precision tests of the standard model of particle physics and searches for processes involving elementary particles are facilitated with novel reconstruction algorithms that exploit complex neural network architectures. Such applications however oftentimes rely on simulated processes, one example being the identification of the flavour of quarks or gluons initiating particle jets (jet flavour tagging). Besides other experimental sources of uncertainties, the efficiency uncertainties stemming from object identification involving neural networks can contribute significantly to final results, expressed for example as uncertainties in a signal strength. Tests with control regions reveal differences in performance between samples obtained through simulation and those from detector data, meaning that calibration is required. With this in mind, this thesis aims at providing not only efficient measures to mitigate this performance gap between data and simulation from the ground up (especially when the algorithm performs very well on simulation), but also derives concepts that assist in understanding why the proposed approaches work. Building up from early versions of adversarial attacks and defenses, a new algorithm, denoted Normed Gradient Method (NGM), is introduced and adapted for physics applications for the first time. This also marks the introduction of a state-of-the-art transformer architecture for small-radius jets for the CMS experiment. In combination with NGM, the currently best performance metrics for this task at CMS are achieved, improving over previous algorithms. The network maintains high performance even under exposition to systematic modifications of inputs. It is thus the first time an (adversarially) robust algorithm is introduced for the official reconstruction software of a high-energy particle detector. Event throughput improves compared to an algorithm that achieves slightly worse performance. The efficient integration was significantly facilitated by a novel software framework specifically developed for jet flavour tagging that is capable of performance studies with data, although the neural network is still in development (training) stage. The time-to-insight from neural network training to reliable performance studies is significantly reduced. This is possible, because the framework is built around only one data tier that serves multiple purposes to study the neural network. Besides the work within the context of one experiment, another focus is the preparation and utilization of CERN Open Data for robustness studies. This last part of this thesis is dedicated to the conversion of already available open datasets, which can only be used with experiment-specific software, into machine learning-friendly formats. The result is the first open dataset that allows small-radius jet flavour tagging studies with simulation and recorded detector data for a broader audience of data scientists that do not necessarily know how to operate the experiment software. As every experiment provides unique data, tools, and problem statements, the concept of a wrapper structure is introduced, which allows applying a core set of adversarial techniques in a toolbox to the different use cases. LB - PUB:(DE-HGF)11 DO - DOI:10.18154/RWTH-2024-07840 UR - https://publications.rwth-aachen.de/record/991721 ER -