Deep Continuum Suppression with Predictive Uncertainties at the Belle II Experiment

The SuperKEKB collider in Japan accelerates
electrons and positrons which collide at an energy of 11 GeV aiming to produce B meson
pairs. To analyse the properties of a B meson, it gets recombined from its decay daughters.
Since there are a lot of background events that mimic the signature of a B meson, it is
important to suppress these events. Such a suppression is called Continuum Suppression. This work aims to further improve the Deep Continuum Suppression, which is perfromed by a Multilayer Perceptron (MLP), by three points: Firstly, an MLP needs a fixed order for input particles. Therefore, this thesis
presents a reliable, self-attention-based input mechanism which allows for invariance under
the particle order. Secondly, to guarantee certainty about the prediction of deep learning models, there is a
high interest in models with predictive uncertainties. This is addressed using the concept
of Deep Ensembles to predict uncertainties of continuum classifications.
Finally, the use of vertex information for the training of a model leads to a bias in
certain analysis variables. Since this could lead to falsification of further studies, a Distance
Correlation is used to decorrelate the Continuum Suppression model from third variables.