Thursday

As next-gen radio surveys come online, data analysis pipelines will need to adapt to the vastly larger data volumes and increased sensitivity of instruments. A key part of this preparation requires realistic simulations of large numbers of unique sources in order to guide decision-making and pre-emptively improve analysis procedures. Deep generative models have been successful for producing high resolution synthetic images in many domains, including astrophysics, and are one potential solution to simulating different types of astronomical observation. In this talk, I will discuss the advantages of generating synthetic radio sources using generative adversarial networks (GANs) and in particular how these synthetic data can be leveraged to help constrain radio galaxy classification in the regime of many more unlabelled data points than labelled, relevant to the case of new sky surveys from the SKA and its precursor instruments. I will argue that the given use case may lead to some unintuitive considerations when evaluating such synthetic data-sets and that conventional metrics for image "quality" may be misleading.