Neural Probabilistic Modeling for Astrophysics and Galaxy Evolution
Statistical modeling in modern astrophysics and cosmology frequently involves simplified analytic models which fail to capture the underlying complexity of the problem at hand. And though traditional machine learning methods have proved useful, they scale poorly with dataset size and will struggle to make optimal use of the vast quantities of data soon to be produced by near-future surveys. Recently, neural networks have provided a powerful new foothold for probabilistic modeling in the presence of large data volumes by acting as expressive universal function approximators. In this thesis, I consider two applications of neural networks: (i) I use a convolutional neural network with an adversarial regularizing loss to deblend superimposed galaxy images. I focus on two-component blends and show that a model trained with a combination of supervised pixel-wise and regularizing adversarial losses provides high fidelity deblended images. And (ii) I use normalizing flows, a neural density estimation technique, to model the distribution over intrinsic quasar continua near Lyman-alpha given the redward spectrum. I then constrain the timeline of the Epoch of Reionization by measuring the neutral fraction of hydrogen in the spectrum of two z>7 quasars. I also describe ongoing work on recovering the local density field in the vicinity of distant galaxies with attention-based graph neural networks and likelihood free inference for astrophysical simulator models with intractable likelihoods.