Dortmund 2021 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 2: AKPIK II: Deep Learning
AKPIK 2.1: Vortrag
Mittwoch, 17. März 2021, 16:00–16:15, AKPIKa
Demonstrating learned tree reconstruction with graph neural networks — James Kahn2, Oskar Taubert2, •Ilias Tsaklidis1, Markus Götz2, Giulio Dujany3, Tobias Boeckh1, Florian Bernlochner1, Pablo Goldenzweig2, Isabelle Ripp-Baudot3, Lea Reuter2, and Arthur Thaller3 for the Belle II collaboration — 1Physikalisches Institut der Rheinischen Friedrich-Wilhelms-Universität Bonn — 2Karlsruher Institut für Technologie (KIT) — 3Hubert Curien Multi-disciplinary Institute, Strasbourg (IPHC)
Hierarchical tree data structures are commonly used in a variety of domains to express chronological interactions. Within particle physics, decay trees need to be reconstructed using only information from the final state particles (FSPs) that reach the detector. The FSPs are represented as leaf nodes in a decay tree graph and they are used to reconstruct the intermediate nodes up to the level of the root. Established methods that retrieve the structural information of the tree, often require domain-specific knowledge to narrow down the combinatorial phasespace, mainly due to the combinatorial explosion in scenarios with many FSPs. In this work, inspired by the usage of a tagging algorithm for full event reconstruction in Belle II, we propose a method of encoding the whole tree structure into leaf-node relations, using a Lowest Common Ancestor based matrix representation. We demonstrate this method using an attention-based transformer network as baseline and a Neural Relational Inference graph network.