Exploring the Cosmos with Deep Learning

Explore the cutting-edge applications of Deep Learning in modern Cosmology through this illuminating talk by Dr. Francois Lanusse at the Alan Turing Institute. Delve into the challenges of understanding dark matter and dark energy, which comprise 95% of the Universe, and discover how new cosmological surveys are mapping the cosmos on an unprecedented scale. Learn about innovative Deep Learning techniques addressing critical issues in cosmological data analysis, from automated detection of rare astronomical objects like strong gravitational lenses using deep residual networks to the creation of realistic galaxy images with deep generative models. Gain insights into the use of graph convolutional networks for modeling galaxy properties along the cosmic web in large-scale simulations. Covering topics such as the Large Synoptic Survey Telescope, Conditional Variational AutoEncoders, and spectral graph convolutions, this comprehensive presentation offers a fascinating glimpse into the intersection of artificial intelligence and astrophysics, showcasing how these advanced technologies are revolutionizing our understanding of the Universe.

Intro
the ACDM view of the Universe
the Large Synoptic Survey Telescope
what does it look like?
the challenge for modern surveys
example of application: gravitational time delays
automated lens searches: Ring Finder (Gavazzi et al. 2014)
Conventional Convolutional Neural Network
Bottleneck residual units
Pre-activation ResNet
I performance on simulations
Euclid strong lens finding challenge
impact of galaxy morphology
Auto-Encoding Variational Bayes (Kingma & Welling 2014)
recognition model and variational lower bound
the reparameterization trick
Conditional Variational AutoEncoder (CVAE)
testing the conditional generation
morphological statistics
take away message
intrinsic alignment of galaxies
hydrodynamical simulations
inpainting intrinsic aligments on Nbody simulations
spectral theory on graphs
spectral graph convolutions
Graph Convolutional Networks (Kipf & Welling 2017)
results on the galaxy inpainting problem
conclusion