Exploring the Universality of Hadronic Jet Classification

Kingman Cheung1¹, Yi-Lun Chung^1*, Shih-Chieh Hsu², Benjamin Nachman³

¹Department of Physics, National Tsing Hua University, Hsinchu, Taiwan
²Department of Physics, University of Washington, Seattle, Washington, USA
³Physics Division, Lawrence Berkeley National Laboratory, Berkeley, CA, USA

* Presenter:Yi-Lun Chung, email:s107022801@m107.nthu.edu.tw

Deep learning is becoming the basis of a variety of techniques for data analysis in High Energy Physics, especially for disentangling signal from background events. These techniques provide extraordinary performance by exploiting subtle correlations in high-dimensional datasets such as those present in hadronic final states. Parton Shower Monte Carlo (PSMC) algorithms are the dominant contribution to the high dimensionality of simulated datasets. Differences among PSMC programs result in systematic uncertainties in deep learning analyses. To explore these two-point uncertainties and the impact on deep neural networks (DNNs), we have surveyed several popular PSMC models to train DNN classifiers. By fixing the testing generator, we can see if the DNNs have learned to use the same information, even if the extent to which that information is expressed varies between training datasets. We show that if the test target is fixed and DNNs are varied, high precision can be obtained at chosen working points. These results suggest that DNNs can learn universal properties of hadronic jets and be insensitive to fragmentation models.

Keywords: Machine Learning, Uncertainty, Classification