Coffee talks
Friday 17/05/2024 @ 11:30, Sala riunioni quarto piano e on-line (meet.google.com/sue-bwvk-axf)
Michele Delli Veneri (Napoli), "Deep Focus: a Deep Learning Metalerner for ALMA imaging"
In this contribution, we will present Deep Focus, a Deep Learning (DL) based metalearner, tasked with solving both the deconvolution and source detection problems in radio-interferometric data cubes. Deep Focus leverages HPC clusters to explore the space of DL architectures in search for those with the highest performances. Uncertainties on detected sources are produced through majority voting. Deep Focus is trained and validated on mock data generated through ALMASim, our open-source ALMA simulator built to facilitate the development of ML algorithms for ALMA data imaging. Building on CASA Simulator, ALMASim generates simulations of galactic and extragalactic continuum and emission lines as observed across ALMA configurations. ALMASim creates sky model cubes (noiseless sky), dirty ALMA cubes (realistic observations), and injected source properties (flux, position, and morphological characteristics) directly sampling the ALMA archive. It employs real ancillary info, and metadata (as antenna setup, weather, band) and leverages parallel computing for efficient large-scale data simulation. Deep Focus performances are compared with CLEAN on real and mock data, showing improvements in both accuracy and time of reconstruction.