Machine learning approach to the detection of point sources in maps of the CMB temperature anisotropies

We propose a machine learning approach to the blind detection of extragalactic point sources on maps of the temperature anisotropies of the cosmic microwave background. Using realistic simulations of the microwave sky as seen by Planck, we train a convolutional neural network (CNN) that solves source detection as an image segmentation problem. We divide the sky into regions of progressively increasing Galactic foreground intensity and independently train specialized CNNs for each region. This strategy leads to promising levels of completeness and reliability, with our CNN substantially outperforming traditional detection methods like the matched filter in regions close to the Galactic plane.