AI (artificial intelligence) is significantly changing the way we solve inverse problems in astrophysics.
In radio interferometry, the detection of radio sources and their classification require taking into account numerous effects such as non-Gaussian noise, incomplete sampling of Fourier space, and the need to construct a sufficient data set for the training. The difficulty increases when the source to be reconstructed evolves over time. Such examples of temporal variations are found in various inverse problems in astrophysics such as transient objects (supernovae, fast radio burst, etc.). ARGOS is a pilot study for a radio interferometer that will perform real-time continuous wide-field observations in centimetre wavelengths. The combination of a wide field of view with high sensitivity will allow ARGOS to detect transient sources that vary on timescales shorter than one second. ARGOS will be able to detect thousands of fast radio bursts per year. These events will need to be accurately differentiated from other transient sources detected by ARGOS, such as supernovae, gamma-ray bursts, white dwarfs, neutron stars, blazars, etc. Given the short time scales of some of these transient events and the need for quick follow up, ARGOS will require state-of-the-art classification solutions employing cutting-edge machine learning architectures. This thesis consists of developing innovative tools from machine learning to solve image reconstruction and source classification problems.