Powered by Indico v3.0.3
ABSTRACT: Deep learning has transformed jet tagging, in bringing a leap to tagging performance and hence substantially improving the sensitivity of physics searches at the LHC. In seek of further enhancement, recent interests fall in experimenting with more advanced neural network architectures, or injecting physics knowledge into the design of the network. This talk focuses on the latter, with a discussion on how intrinsic physics symmetries play a vital role. In the first part of the talk, we introduce LorentzNet, an efficient model with a graph neural network (GNN) backbone that respects the Lorentz group symmetry of a jet. We showcase that the model improves the tagging performance over the previous state-of-the-art algorithms including ParticleNet, especially when trained only on a few thousand jets. In the second part, we investigate the core of such improvement. We use various experiments to reveal that “full Lorentz-symmetry preservation” serves as a strong inductive bias of jet physics tasks, and is the key ingredient in boosting the network performance. Based on the spirit, we propose two generalized patch structures for modifying a baseline network which exhibit an overall performance gain. The studies hint at a potential role of Lorentz-symmetric design in future network development for jet physics.
This talk is based on arXiv.2201.08187 and arXiv:2208.07814.
About the speaker: Congqiao Li is currently a doctoral researcher at Peking University and a member of CMS Collaboration. He has been working on high-energy physics experimental analyses to measure the Higgs boson Yukawa coupling with charm quarks. He contributed to the collaboration in offline computing of event generators, calibration of heavy flavour jet tagging algorithms, and promoting the use of new advances in deep learning. His current interest includes designing modern physics analysis with use of advanced deep learning insights.
Zoom Link: https://ihep-ac-cn.zoom.us/j/82686085014?pwd=eDBzYVFjVzd6ZmhYaGhQMDArRkdmZz09
Zoon ID: 82686085014
Passwd: 841771