Ming-Feng Ho (Graduate Institute of Astrophysics, National Taiwan University; Institute of Chinese Literature and Philosophy, Academia Sinica; Institute of Astrophysics and Astronomy, Academia Sinica); Lung-Yih, Chiang (Institute of Astrophysics and Astronomy, Academia Sinica); Chi-Hung Yan (Institute of Astrophysics and Astronomy, Academia Sinica)
Cosmic microwave background (CMB) temperature anisotropies encode the history of the universe, which manifest itself in the angular power spectrum. Differences in the power spectra even from small patches expose details such as gravitational lensing, Doppler boosting effects, or also unknown effect. We design a filter and construct a convolutional feature map on the CMB sky to represent the stretching and contraction effects on small areas of ESA Planck CMB map. On the other hand, we apply a method to allow convolutional neural network (CNN) building on spheres. Via the hierarchical grid of HEALPix, we can build CNN directly on every pixel of a sphere without any interpolation. The combination of CNN and HEALPix can help astronomers to train convolutional feature maps on full-sky data.