Speaker
Description
The success of large observatories such as the IceCube neutrino telescope is highly dependent on the accuracy of their reconstruction algorithms. In IceCube, traditional likelihood-based methods are limited by the lookup tables used for calculating the event hypotheses, since their complexity requires them to be simplified. Promising results have recently been obtained with Event-Generator, a generative neural network that can replace such tables and lead to an improvement in reconstruction performance since it does not require simplification. The success of this neural network lies in its design, which, unlike most deep learning applications, is able to explicitly exploit the information domain of IceCube event generation, such as symmetries and detector properties. In this talk, Event-Generator will be introduced and the current status and future plans for its implementation in IceCube-Gen2 will be presented.
