Samuel Spencer

New deep learning analyses are a promising new method of background rejection and event reconstruction for Imaging Atmospheric Cherenkov Telescopes (IACTs), particularly in the context of the next generation Cherenkov
Telescope Array (CTA). This is because they allow for sensitive analysis of complete camera images at high speed. Unlike other fields of astrophysics, where deep learning is being used to characterise astronomical sources,
deep learning use in IACT astronomy is comparatively unusual in that the analysis targets are Extensive Air Showers (EAS) in Earth’s atmosphere. As such, we have access to large datasets of highly complex Monte Carlo
simulations of both the air shower particle physics and our detectors. However, this in turn leads to a highly non-trivial domain gap problem when attempting to apply deep learning methods trained on simulations to real
data. We will present state-of-the-art results displaying the combined effects of custom simulations, optimisation and graph-based network architectures to attack this problem.

Figure 1 shows our detection of the Crab Nebula using a ConvLSTM2D event classifier and VERITAS, at 13.8 sigma. This was only possible after performing Bayesian optimisation of the classifier hyperparameters, a further example of which we show in Figure 2. Given the sensitivity of these classifiers to small image details from Night Sky Background, we are now investigating graph-based Chebyshev networks as an alternative, an example of the input to this is shown in Figure 3, and initial classifier ROC curves are shown in Figure 4.