SMuK 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Neural Networks II

AKPIK 4.2: Vortrag

Mittwoch, 22. März 2023, 16:00–16:15, ZEU/0118

Uncertainty estimations for deep learning-based imaging — •Felix Geyer, Arne Poggenpohl, and Kevin Schmidt — Astroparticle Physics WG Elsässer, TU Dortmund University, Germany

Radio interferometry is used to monitor and observe distant astronomical sources and objects with high resolution. Especially Very Long Baseline Interferometry (VLBI) allows for achieving the highest resolutions by combining the data of multiple telescopes. This results in an effective diameter corresponding to the greatest distance between two telescopes. The taken data consists of visibilities in Fourier space, which depend on the baselines between the telescopes. Because the distribution of these baselines is sparse, the sample of visibilities is incomplete. After transforming this sample to spatial space, this so-called "dirty image" is inadequate for physical inference and analyses.

In traditional methods, the image then undergoes an elongated and mostly manually performed cleaning process in order to remove background artifacts and restore the original source distribution. Contrary, a new and fast approach to reconstructing missing data reasonably is using neural networks. As an additional advantage, these networks can also be used to estimate the uncertainty of the prediction. This is done by not only predicting the mean value of the pixels but also the standard deviation by feeding the input and the prediction to a separate network. All of this is part of our framework called radionets, which is another focus of this talk.