Neural networks and separation of Cosmic Microwave Background and astrophysical signals in sky maps

Abstract

The Independent Component Analysis (ICA) algorithm is implemented as a neural network for separating signals of different origin in astrophysical sky maps. Due to its self-organizing capability, it works without prior assumptions on the signals, neither on their frequency scaling, nor on the signal maps themselves; instead, it learns directly from the input data how to separate the physical components, making use of their statistical independence. To test the capabilities of this approach, we apply the ICA algorithm on sky patches, taken from simulations and observations, at the microwave frequencies, that are going to be deeply explored in a few years on the whole sky, by the Microwave Anisotropy Probe (MAP) and by the {\sc Planck} Surveyor Satellite. The maps are at the frequencies of the Low Frequency Instrument (LFI) aboard the {\sc Planck} satellite (30, 44, 70 and 100 GHz), and contain simulated astrophysical radio sources, Cosmic Microwave Background (CMB) radiation, and Galactic diffuse emissions from thermal dust and synchrotron. We show that the ICA algorithm is able to recover each signal, with precision going from 10% for the Galactic components to percent for CMB; radio sources are almost completely recovered down to a flux limit corresponding to $0.7σ_{CMB}$, where $σ_{CMB}$ is the rms level of CMB fluctuations. The signal recovering possesses equal quality on all the scales larger then the pixel size. In addition, we show that the frequency scalings of the input signals can be partially inferred from the ICA outputs, at the percent precision for the dominant components, radio sources and CMB.