Heidelberg 2022 – wissenschaftliches Programm

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Deep Learning

AKPIK 4.5: Vortrag

Donnerstag, 24. März 2022, 17:15–17:30, AKPIK-H13

A Recurrent Neural Network for Radio Imaging — •Stefan Fröse and Kevin Schmidt — Astroparticle Physics WG Elsässer, TU Dortmund University, Germany

In radio astronomy, an array of correlated antennas, called a radio interferometer, is used to produce high-resolution images of the sky. The measurements take place in the complex Fourier space due to the pairwise correlation of antennas. Therefore, the amount of information to receive from such an array is restricted by the number of antennas. The resulting spatial dirty map of these measurements will be cleaned using a Neural Network. The architecture for this network is based on a Recurrent Neural Network (RNN). RNNs can be used to extract information from sequential data, like text or speech. In the context of inverse problems the RNN can be derived directly from a maximum a posteriori approach. Furthermore the iterative behaviour of the network can be exploited to construct a CLEAN-like network to reconstruct a map of the sky. This results in the so-called RIM architecture published by Patrick Putzky & Max Welling (arxiv:1706.04008). The Neural Network is able to clean given dirty maps for simulated radio images and also shows convergence for the EHT dataset of M87.