Detecting gravitational lenses using machine learning: exploring interpretability and sensitivity to rare lensing configurations

Forthcoming large imaging surveys such as Euclid and the Vera Rubin Observatory Legacy Survey of Space and Time are expected to find more than 10(5) strong gravitational lens systems, including many rare and exotic populations such as compound lenses, but these 10(5) systems will be interspersed among much larger catalogues of similar to 10(9) galaxies. This volume of data is too much for visual inspection by volunteers alone to be feasible and gravitational lenses will only appear in a small fraction of these data which could cause a large amount of false positives. Machine learning is the obvious alternative but the algorithms' internal workings are not obviously interpretable, so their selection functions are opaque and it is not clear whether they would select against important rare populations. We design, build, and train several convolutional neural networks (CNNs) to identify strong gravitational lenses using VIS, Y, J, and H bands of simulated data, with F1 scores between 0.83 and 0.91 on 100 000 test set images. We demonstrate for the first time that such CNNs do not select against compound lenses, obtaining recall scores as high as 76 per cent for compound arcs and 52 per cent for double rings. We verify this performance using Hubble Space Telescope and Hyper Suprime-Cam data of all known compound lens systems. Finally, we explore for the first time the interpretability of these CNNs using Deep Dream, Guided Grad-CAM, and by exploring the kernels of the convolutional layers, to illuminate why CNNs succeed in compound lens selection.