Merging deep learning with physical models, for the analysis of modern cosmological surveys
The upcoming generation of cosmological surveys such as LSST will aim to map the Universe in great detail and on an unprecedented scale. This of course implies new and outstanding challenges at all levels of the scientific analysis, from pixel level data reduction to cosmological inference. In this talk, I will illustrate how recent advances in deep learning and associated automatic differentiation frameworks, can help us tackle these challenges and rethink our approach to data analysis for cosmological surveys. We will see how at the pixel level, combining physical models of the instrument (which account for noise/PSF) with deep generative models (which account for complex galaxy morphologies) can allow us to solve a number of astronomical inverse problems ranging from deconvolution to deblending galaxy images. At the intermediate level of estimating gravitational lensing maps, I will present our recent work on combining analytic theoretical priors with simulation-driven deep learning priors, in order to solve the challenging dark matter mass-mapping inversion problem, and sample from its full posterior. Finally, at the cosmological analysis level, I will present our efforts to implement N-body simulations directly in TensorFlow, opening the door to a range of novel and efficient inference techniques, and allowing for fast hybrid physical/ml simulations.
|Date: ||Thursday, 11 February 2021|
|Where: ||Université de Montréal|
|Contact: ||Laurence Perrault Levasseur|