Learning from Simulators: Neural Networks for Inverse Problems

Abstract

A simulator lets us ask, "What would we observe if we knew the inputs?" In science, we usually face the reverse problem: given what we observed, which inputs could have produced it? I'll introduce neural simulation-based inference and show how a network can learn a distribution over plausible answers from simulated examples, even when the likelihood is too costly to evaluate. I'll show how we combined a graph neural network with a normalizing flow to infer Ultra-high-energy cosmic-ray arrival directions. Then I'll turn to a question that matters beyond physics; how do we test whether we can trust a model's uncertainty estimates?

Biography

Oscar Macias is an assistant professor in the Department of Physics and Astronomy at San Francisco State University. He studies the high-energy universe through gamma rays, neutrinos, and cosmic rays, and develops machine learning methods to interpret the signals they leave in detectors. His research is supported by the US National Science Foundation. He earned his Ph.D. in physics at the University of Canterbury in New Zealand and was a GRAPPA Prize Fellow at the University of Amsterdam before joining SFSU.