Heidelberg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
T: Fachverband Teilchenphysik
T 26: Data Analysis, Information Technology and Artificial Intelligence
T 26.2: Vortrag
Montag, 21. März 2022, 16:30–16:45, T-H39
Performance Studies of the Conditional Attention Deep Sets b-Tagger for the ATLAS Experiment — •Alexander Froch1, Manuel Guth2, and Andrea Knue1 — 1Albert-Ludwigs-Universität Freiburg — 2Université de Genève
The identification of jets containing b-hadrons, called b-tagging, is crucial for most analyses performed
at the ATLAS experiment. Several new multivariate techniques have been developed for this purpose. One of these
is the Deep-Impact-Parameter-Sets (DIPS) tagger.
The DIPS tagger is a deep neural network based on the Deep-Sets architecture. It
uses track information of the particles inside the clustered jets for
classification. It is part of a new generation of tagging algorithms currently developed in
ATLAS. DIPS itself can distinguish between different jet origins, like light, charm or bottom jets.
Although DIPS already outperforms the currently recommended RNNIP tagger, its high-pT performance can still be improved further.
While the number of fragmentation tracks increases rapidly with pT, less heavy-flavour tracks are being reconstructed at high pT.
Therefore, it is more difficult for this kind of network to filter the most important tracks.
To further enhance the tagging capabilities of DIPS and fix the issues arising in the higher pT region, DIPS will be extended
with an attention mechanism conditioned on jet kinematics. This new version is the Conditional Attention Deep Sets (CADS) tagger.
The new CADS tagger will be discussed and its performance will be compared to the current best DIPS model.