Performance studies with Normalizing Flow transformations

A Normalizing Flow is a generative model used to explicitly learn an underlying PDF of given data. Besides generating new data, a major advantage of Normalizing Flows lies in its mathematical structure which allows performing exact density estimation. For HEP this allows successful simulation of particle collisions or simulation of the interaction with detectors, but also anomaly detection by density estimation.

For effective utilization of these Normalizing Flows in HEP analyses, it is important to understand the fundamental properties of different architectures and their performance as a function of data dimensionality.

In this thesis, three different transformations (RNVP, MAF and CSF) are compared using metrics, and their properties on a two dimensional dataset are investigated.