Parts | Days | Selection | Search | Updates | Downloads | Help

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Neural Networks II

AKPIK 4.4: Talk

Wednesday, March 22, 2023, 16:30–16:45, ZEU/0118

Gamma Source Detection using Deep Multitask Networks and Noisy Label Learning — •Lukas Pfahler — TU Dortmund University, Artificial Intelligence Group, Dortmund, Germany

Machine learning has been established as an effective tool for data analysis in modern high energy particle experiments. For the FACT telescope, we solve three supervised learning tasks - gamma-hadron separation, energy estimation, and origin estimation - using simulated training data and manual feature extraction. We outline how we can replace the manual feature engineering currently applied with a learned representation trained with multitask supervision. Our approach will train a shared representation that can solve all three prediction tasks with specialized prediction networks build on top of the shared representation. Furthermore, we look into an alternative source of supervision that reduces the burden of simulating training data by using real telescope recordings. We rely on the concept of noisy labels and introduce a novel method for learning under label noise where only one noise rate is known. We show how gamma-hadron separation can be framed in this setting and illustrate that the method allows us to train accurate classifiers.