Generative Models: A Novel Tool to Generate Statistically Realistic Astrophysical Data
I am a postdoctoral researcher at the University of Nottingham, where I do research focusing on two major themes: testing theories of gravity using astrophysical data and applying machine learning techniques in astrophysics/cosmology. When it comes to machine learning, my key interest is in how to use deep neural network-based algorithms to emulate cosmological/astrophysical data.
In this talk I will present my recent work on generating realistic cosmological/astrophysical data using generative models. Generative models refer to a class of machine learning algorithms that are able to generate statistically realistic mock data by learning from a training dataset. From ultra-realistic human faces to novel works of art, generative models have been been used for solving a wide variety of problems. In the context of astrophysics, machine learning algorithms such as generative adversarial networks (GANs) or variational autoencoders (VAEs) offer a novel way for generating unseen and statistically realistic weak lensing and cosmic web data in a matter of seconds. In this talk I will introduce generative models, offer a variety of sample applications and discuss strategies to train such models efficiently. Finally, a theoretic overview of the training process will be given.