Machine learning for inverse problems

Inverse problems play a crucial role in science and engineering. They arise because theoretical models are typically developed in the causal direction of the data generating process, mapping from a physical system to observable data. Data analysis generally aims to solve the inverse direction: starting from observed data, the goal is to infer underlying properties that could have given rise to the data. Therefore, inference often corresponds to inversion of a forward model. In this talk, I will explain how machine learning can solve such inverse problems in cases where conventional inference is intractable or prohibitively expensive. I will show examples from several scientific applications, focusing in particular on a line of work in gravitational wave astronomy.