Using a Conditional Generative Adversarial Network to Control the Statistical Characteristics of Generated Images for IACT Data Analysis

Abstract

Using a cGAN for IACT Data AnalysisJulia Dubenskaya et al.

Ultra-fast generation of air shower images for Imaging Air Cherenkov Telescopes using Generative Adversarial Networks

“…The coordinate is defined as the angle between the 𝑥-axis on the camera and the major axis of the ellipse and, therefore, features six maxima due to the hexagonal camera design of FlashCam. Combined with the previous observations, this finding indicates that many different shower geometries are generated by the algorithm, excluding mode collapsing present in previous approaches [23].…”

“…Overall, the pixel occupancies of the simulated and generated data sets are very similar and show the same behavior of a smaller occupancy at the edges of the camera and a larger occupancy in the central area of the camera. Combined with the finding of the image quality, these findings show that our generative algorithm does not suffer from mode collapsing, as present in previous works [22,23]. The three vertically outermost pixels in the simulated image, as well as the mount pixels, i.e., the three triangularly-arranged pixels in the center of the camera, are not simulated in its current design.…”

“…As an initial step, we will concentrate on a simplified approach, which is commonly pursued in event generation using generative models that combine both parts in a single model. The computational requirements for the next-generation flagship in gamma astronomy, the Cherenkov Telescope Array (CTA) [21], led to the first promising studies of fast, memory-efficient, and sustainable generation of IACT images using generative models [22][23][24][25] based on conditional GANs [26] or Wasserstein GANs [27]. Despite showing encouraging results in generating the high-level image parameters directly [28], none of them was, so far, able to provide high-resolution IACT images with high quality, including a comprehensive and convincing analysis of the fidelity of the generated images, e.g., by studying Hillas parameters and their correlations.…”