How to GAN Higher Jet Resolution

Abstract

QCD-jets at the LHC are described by simple physics principles. We show how superresolution generative networks can learn the underlying structures and use them to improve the resolution of jet images. We test this approach on massless QCD-jets and on fat top-jets and find that the network reproduces their main features even without training on pure samples. In addition, we show how a slim network architecture can be constructed once we have control of the full network performance.

What's Anomalous in LHC Jets?

“…[74], and some of them with the specific goal of enhancing the sensitivity of a given anomaly detection method. In general, we also expect such preprocessings to affect the sensitivity to specific physics signals [17,82]. The first choice, crucial for any neural network, is how we define the p T -information per jet constituent.…”

“…Both standard Model and anomalous features in jets occur at very different p T -scales; hard decays lead to features at p T 1 GeV, while jets with a modified parton showering are sensitive to GeV-scale constituents, similar to quark-gluon tagging. This means the standard choice biases a classification or anomaly detection technique towards features at high p T [3,82]. This explains why autoencoders tag jets with higher complexity more easily, if complexity or structure is usually assumed to affect the harder features of the jet.…”

“…The reweightings also change the density of the jets in physics space, and given that we define anomalous jets as those in the low density regions, they physically alter what jets are anomalous. For an optimal network training it might eventually be beneficial to provide the network with two reweighted inputs, one focusing on soft pixels and one focusing on hard pixels [82].…”

Style-based quantum generative adversarial networks for Monte Carlo events

“…In particular, the introduction of GAN models in HEP Monte Carlo simulation has been discussed extensively in the last years, see Refs. [48][49][50][51][52][53][54]. In this work, we consider the possibility to use a qGAN model in a data augmentation context, where the model is trained with a small amount of input samples and it learns how to sample the underlying distribution.…”

Targeting Multi-Loop Integrals with Neural Networks

“…Along the established LHC simulation chain, machine learning has already shown great promise when it comes to faster and more precise predictions [3]. This includes phase space integration [4,5], phase space sampling [6][7][8][9], amplitude evaluation [10][11][12][13], event subtraction [14], event unweighting [15,16], parton showering [17][18][19][20], parton densities [21,22] or particle flow descriptions [23,24]. Full neural network-based event generators [25][26][27][28][29][30] can be used to invert the simulation chain and unfold detector effects as well as QCD jet radiation [31][32][33].…”