Bayesian Deep Learning for Large Scale Structure

Abstract

Modern machine learning methods based on high-dimensional Bayesian inference can complement observations and play an important role in large scale structure cosmology. Through years of improving observation techniques and computational and modeling capabilities, we are getting better at estimating and measuring the values of the cosmological parameters that describe the large scale structure of our universe. In this study, we present the results of the first-ever application of Bayesian deep learning for large scale cosmological structure, which is an important step to make machine learning more rigorous for cosmology. Working with the Quijote simulations, we apply MultiSWAG (Multi-modal Stochastic Weight Averaging Gaussian) based Convolutional Neural Networks (CNNs) to volumetric representations of dark matter simulations in order to predict the values of five cosmological parameters, thus extracting the maximal amount of information from cosmological datasets.