“HIGAN: Cosmic Neural Hydrogen with Generative Adversarial Networks” NeurIPS Workshop
Paper by CDS alumni to be presented at NeurIPS Workshop Machine Learning and Physical Sciences
A paper entitled, “HIGAN: Cosmic Neural Hydrogen with Generative Adversarial Networks” authored by CDS Master of Data Science alumni Atakan Okan and Juan Zamudio-Fernandez, has been accepted to the NeurIPS 2019 workshop Machine Learning and Physical Sciences. The aim of this workshop, according to the team, is to “bring together computer scientists, mathematicians and physical scientists who are interested in applying machine learning to various outstanding physical problems, including in inverse problems; approximating physical processes; understanding what a learned model really represents; and connecting tools and insights from the physical sciences to the study of machine learning models.”
This project began as a capstone in collaboration with the Flatiron Institute — Center for Computational Astrophysics, which Juan and Atakan later continued as an Independent research project course. The need for this project arose from a computationally and financially expensive simulation called IllustrisTNG, a simulation that only generated one big cube.
“We tried different generative models (especially generative adversarial networks — GANs) and architectures to generate cubes that are with similar structures,” the team told us. “After months of trialing and optimization, Wasserstein GANs and its varied architectures started to outperform the state of the art astrophysical model called Halo occupation distribution (HOD).” They performed a comprehensive literature review, and implemented and replicated state-of-the-art models to see which one worked best.
“It was great to learn about astrophysics and problems that academics face,” the team says. “It was beneficial for us to define the problem outside of classic data science problem types and come up with a solution that addressed it.” He continues, “There were a lot of realizations about the current work done in the field where differences in the dataset can affect the results of the tried-and-verified architectures/models’ performance. Our dataset was extremely skewed and slowed our progress quite a lot.”
The workshop will be held at the 33rd Conference on Neural Information Processing Systems (NeurIPS) on December 14th, 2019.
By Mary Oliver
