- Feb. 24, 2020, 2:00 pm
- Curia II
- Aleksandra Ćiprijanović, Mathematical Institute of the Serbian Academy of Sciences and Arts
Abstract: Distinguishing between merging and non-merging galaxies in observations can sometimes be a very slow and difficult process. Today, with the availability of large-scale simulations, we have the ability to learn from large samples of labeled images of merging galaxies and transfer that knowledge to search for these objects in the observed data. The task of distinguishing between merging and non-merging galaxies in simulated images can be well performed by convolutional neural networks (CNNs). We investigate this approach for the first time at high redshifts (i.e., z=2). For this task we use images from the Illustris-1 cosmological simulation and apply observational and experimental noise that mimics that from the Hubble Space Telescope. The classification accuracy of our DeepMerge CNN is between 76-79% and it outperforms even a Random Forest classifier, which was shown to be superior to conventional statistical methods (Concentration, Asymmetry, the Gini, M_20 statistics etc.). We also investigate the selection effects of the classifier with respect to merger state and star formation rate, and we extract Grad-CAMs to further understand the process of classification by DeepMerge CNN.