Abstract:
While our understanding of the Universe's composition and formation has
advanced in the last two decades, fundamental questions persist. The
nature of the majority of the Universe, comprising 27% Dark Matter and
68% Dark Energy, remains unknown. To unveil these mysteries, astronomers
are constructing ambitious telescopes and sky surveys. LSST and Euclid
will soon come online, surveying vast portions of the sky with
unprecedented depth.
To deal with this volume of data, machine learning methods are planned
as integral parts of the analysis. I will discuss areas where machine
learning is useful for this work. My group is studying how to translate
machine learning methods from industry and computer science contexts to
cosmology. One approach we have taken is to use probabilistic neural
networks to combine the predictive power of deep neural networks with
the statistical framework of Bayesian inference. I will show that this
approach can give us good predictions with accurate uncertainties for
the distances (redshift) to galaxies using their brightness and images
at different wavelengths. I will also discuss whether we can teach a
neural network to understand galaxy evolution and how we could assess it.