Speaker
Mirko Bunse
(Artificial Intelligence Unit, Computer Science VIII, TU Dortmund University)
Description
Imaging atmospheric Cherenkov telescopes (IACTs) typically require simulations to obtain labeled training data for reconstruction tasks. For the specific task of gamma hadron classification, we show that no simulations are needed if the direction of each event is employed as a so-called "noisy label". Machine learning research on the theory of class-conditional label noise provides us with formal proofs that accurate classifiers can indeed be learnt from labels of this kind. Our experiments with data from the FACT telescope demonstrate that this approach can outperform state-of-the-art classification pipelines.
