What can Deep Learning help in addressing Astrophysical Challenges
Rapid progress in data science promises to deeply transform how we extract meaning from complex data sets. New powerful analysis and simulation techniques such as deep learning (DL) are quickly emerging and impacting all data-based sciences. Astrophysics is no exception, as illustrated by some of our work: Ravanbakhsh et al. (2016) were the first to estimate cosmological parameters from 3D dark matter simulations directly achieving ~3-5 times stronger constraints than power-spectrum alone; He et al. (2018) used a convolutional neural network to generate large scale structure dark matter simulation million times faster than our traditional simulation techniques; Yip et al. 2019 used U-net to predict the position of all types of galaxies given a dark matter density map, by-passing the need of large volume full hydrodynamic simulation of the Universe. Zamudio-Fernandez et al. 2019 used a generative adversarial network to produce new neutral hydrogen simulation with Illustris as a training set. All of these examples illustrate how deep learning can be used to tackle astrophysical challenges, and all of them point to an end goal of extracting the maximum amount of information from the observations of our Universe by directly comparing our theoretical understanding in the form of simulated Universes to the observables. In this talk, I will discuss this vision and how deep learning techniques help our quest in understanding the Universe.
|Friday, 6 November 2020
|Université de Montréal