Karlsruhe 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 96: Data, AI, Computing 7 (uncertainties, likelihoods)
T 96.3: Talk
Thursday, March 7, 2024, 16:30–16:45, Geb. 30.33: MTI
Sharing AI-based Searches with Classifier Surrogates — •Sebastian Bieringer1, Gregor Kasieczka1, Jan Kieseler2, and Mathias Trabs3 — 1Universität Hamburg, Institut für Experimentalphysik, Luruper Chaussee 149, 22761 Hamburg, Germany — 2Karlsruhe Institute of Technology, Institute of Experimental Particle Physics, 76131 Karlsruhe, Germany — 3Karlsruhe Institute of Technology, Institute of Stochastics, 76131 Karlsruhe, Germany
In recent years, Neural Network-based classification has been used to improve evaluations at collider experiments. While this strategy proofs to be hugely successful, the underlying models are not commonly shared with the public as they are based on experiment-internal simulation. We propose to cater a generative model, a Classifier Surrogate, sampling the classification output from high-level jet information, with every Neural Network-based evaluation to enable further tests on the evaluation. Continuous Normalizing Flows are one suitable generative architecture that can be efficiently trained using Conditional Flow Matching and easily extended by Bayesian uncertainties to indicate the unknown inputs to the user. For a top-tagging example, we demonstrate the application of Flows in combination with uncertainty estimation through either inference of a mean-field Gaussian weight posterior, or Monte Carlo sampling network weights.
Keywords: Bayesian Deep Learning; Generative Networks; Tagging; Classification