Title: Enhancing the utility of deep learning applications to LHC phenomenology with QCD priors
Understanding the behaviour of deep-learning algorithms within perturbative Quantum Chromodynamics (QCD) is essential to connect with the various short-distance physics happening in the collisions at the Large Hadron Collider. In this talk, I'll describe the inductive biases that Convolutional Neural Networks (CNNs) presume about the data and their relation to perturbative QCD, which already indicate their suitability in distinguishing different radiation patterns. As a case study, I'll talk about constraining the invisible branching ratio of the Higgs boson produced via the vector boson fusion channel with CNNs.
Although promising, event analysis with low-level data is challenging in hadronic environments due to various non-perturbative effects. Deep-learning algorithms are sensitive to these effects at the outset. To reduce this susceptibility, I'll describe an infrared and collinear safe Graph Neural Network (GNN) algorithm and its application as a jet classifier and anomaly detector. This ensures that the features extracted by such GNNs relate directly to the hard interaction at the collision on top of its already favourable inductive biases compared to CNNs.
Zoom Meeting ID: 948 7183 3595