Layers of complexity: the importance of realism in neural network classifications of interacting galaxies
Connor Bottrell, None · Figshare · 2019
Visual classification is hard. It is time-consuming and ultimately subjective. Fortunately, there are solutions to these problems in the form of citizen science projects and machine learning methods. But far more challenging is identification and classification of interacting systems. When armed only with photometry, contaminants such as field stars and unfortuitously projected galaxies often lead to misclassifications (dependence on environment). Similarly, the classifcations rely sensitively on the identification of low-surface brightness features such as tidal arms, streams, and shells (dependence on sky brightness and resolution). Simulations, on the other hand, offer foreknowledge of the stage of an interaction. Using a suite of 24 galaxy interactions covering a range of mass ratios and orbits with corresponding secular evolution runs, we train convolutional neural networks to classify galaxies into three categories: (iso) isolated; (p/i) pair/interacting; and (pm) post-merger. Our networks are trained on images with increasing levels of realism: (0) projected stellar mass maps; (1) full radiative transfer; (2) insertion into real images. We are then well-poised to answer unique questions. How important is the level of realism to the successful classification of interacting systems? How important is it to train the network with contaminating effects (inserting into real images) so that the network may adequately discriminate between bona-fide interactions and false positives? In this talk, I will discuss our answers to these questions.