Type: Poster
Session: Posters (Monday & Tuesday)
Author: Biswajit Biswas

This work will also be presented as a flash talk in the Machine Learning and Artificial Intelligence session.

Abstract: In upcoming surveys like LSST, we expect more than two-thirds of the galaxies to be affected by blending - the overlapping of neighboring sources. To impose strong constraints on cosmological parameters with probes like weak lensing, we need to control systematics, and blending is one of the major contributors. Commonly used methods for solving the inverse problem of source separation, so-called deblending, either fail to capture the diverse morphologies of galaxies or are too slow to analyze billions of galaxies. To overcome these challenges, we developed a deep learning-based algorithm called MADNESS that deblends galaxies from a field by finding the maximum-a-posteriori solution using deep generative models like variational autoencoders and normalizing flows. I will present the methodology of our algorithm and evaluate its performance. I will compare the results against state-of-the-art techniques including the scarlet deblender, using flux reconstruction, shapes and morphology, and color as a metric, and show that MADNESS is able to obtain significantly better results.