J/A+A/650/A100 Decomposition of Galactic sky with autoencoders (Milosevic+ 2021)
Bayesian decomposition of the Galactic multi-frequency sky using probabilistic
autoencoders.
Milosevic S., Frank P., Leike R.H., Mueller A., Ensslin T.A.
<Astron. Astrophys. 650, A100 (2021)>
=2021A&A...650A.100M 2021A&A...650A.100M (SIMBAD/NED BibCode)
ADC_Keywords: Milky Way ; Interstellar medium
Keywords: methods: data analysis - methods: statistical -
techniques: image processing - Galaxy: general - ISM: structure
Abstract:
All-sky observations show both Galactic and non-Galactic diffuse
emission, for example from interstellar matter or the cosmic microwave
background (CMB). The decomposition of the emission into different
underlying radiative components is an important signal reconstruction
problem. We aim to reconstruct radiative all-sky components using
spectral data, without incorporating knowledge about physical or
spatial correlations. We built a self-instructing algorithm based on
variational autoencoders following three steps: (1) We stated a
forward model describing how the data set is generated from a smaller
set of features, (2) we used Bayes' theorem to derive a posterior
probability distribution, and (3) used variational inference and
statistical independence of the features to approximate the posterior.
From this, we derived a loss function and optimized it with neural
networks. The resulting algorithm contains a quadratic error norm with
a self-adaptive variance estimate to minimize the number of
hyperparameters. We trained our algorithm on independent pixel
vectors, each vector representing the spectral information of the same
pixel in 35 Galactic all-sky maps ranging from the radio to the
gamma-ray regime. The algorithm calculates a compressed representation
of the input data. We find the feature maps derived in the algorithm's
latent space show spatial structures that can be associated with
all-sky representations of known astrophysical components. Our
resulting feature maps encode (1) the dense interstellar medium (ISM),
(2) the hot and dilute regions of the ISM, and (3) the CMB, without
being informed about these components a priori. We conclude that
Bayesian signal reconstruction with independent Gaussian latent space
statistics is sufficient to reconstruct the dense and the dilute ISM,
as well as the CMB, from spectral correlations only. The approximation
of the posterior can be performed computationally efficient using
variational inference and neural networks, making them a suitable
approach to probabilistic data analysis.
Description:
The images show compressed Galactic all-sky maps as calculated by a
neural network architecture called autoencoder. The autoencoder is
trained to calculate a compressed representation of 39 Galactic
all-sky maps compiled by Mueller et al. (2018A&A...620A..64M 2018A&A...620A..64M),
covering frequencies from the radio to the gamma-ray regime. The
results show the posterior mean and variance maps of the compression,
which encode (1) the dense interstellar medium (ISM), (2) the hot and
dilute regions of the ISM, and (3) the CMB. The images are in HEALPix
format (Gorski et al., 2005ApJ...622..759G 2005ApJ...622..759G) with NSIDE=128.
File Summary:
--------------------------------------------------------------------------------
FileName Lrecl Records Explanations
--------------------------------------------------------------------------------
ReadMe 80 . This file
list.dat 100 6 *Information on fits file HDU
fits/* . 1 fits file of feature maps
--------------------------------------------------------------------------------
Note on list.dat: The fits data contains six HDUs, each representing the
data for one feature map. The name of the feature map is stored in the
header card "FEATURE", for example "FEAT C VAR", which denotes the
data for the posterior variance map of feature C, and "FEAT C MEAN"
means the mean posterior of feature C.
-------------------------------------------------------------------------------
Byte-by-byte Description of file: list.dat
--------------------------------------------------------------------------------
Bytes Format Units Label Explanations
--------------------------------------------------------------------------------
1- 6 I6 --- Nx Number of pixels along X-axis
8- 13 I6 --- Ny Number of pixels along Y-axis
15 I1 --- HDU [1/6] HDU number
17- 20 I4 Kibyte size Size of FITS file
22- 38 A17 --- FileName Name of FITS file, in subdirectory fits
40-100 A61 --- Title Title of the FITS file
--------------------------------------------------------------------------------
Acknowledgements:
Sara Milosevic, Sara.mia.milosevic(at)googlemail.com
(End) Patricia Vannier [CDS] 29-Mar-2021