Quark Gluon Jet Discrimination with Weakly Supervised Learning

TL;DR

This paper explores weakly supervised deep learning methods to distinguish quark and gluon jets in high energy physics, reducing reliance on detailed Monte Carlo simulations and enabling direct application to collision data.

Contribution

It demonstrates the effectiveness of weakly supervised learning for quark-gluon jet discrimination using simulated data and compares three different machine learning approaches.

Findings

Weakly supervised learning can effectively discriminate quark and gluon jets.

Performance comparison of CNN, RNN, and BDT methods shows their relative strengths.

Realistic simulated samples are sufficient for training in this context.

Abstract

Deep learning techniques are currently being investigated for high energy physics experiments, to tackle a wide range of problems, with quark and gluon discrimination becoming a benchmark for new algorithms. One weakness is the traditional reliance on Monte Carlo simulations, which may not be well modelled at the detail required by deep learning algorithms. The weakly supervised learning paradigm gives an alternate route to classification, by using samples with different quark--gluon proportions instead of fully labeled samples. The paradigm has, therefore, huge potential for particle physics classification problems as these weakly supervised learning methods can be applied directly to collision data. In this study, we show that realistically simulated samples of dijet and Z+jet events can be used to discriminate between quark and gluon jets by using weakly supervised learning. We implement and compare the performance of weakly supervised learning for quark--gluon jet classification using three different machine learning methods: the jet image-based convolutional neural network, the particle-based recurrent neural network and the feature-based boosted decision tree.