Cosmological and astrophysical data are becoming sufficiently large and complex to soon prove intractable by traditional statistical techniques. Unravelling the twin mysteries of dark energy and dark matter will require new data analysis methods capable of extracting knowledge from upcoming data streams. The recent rise of machine learning to prominence in the physical sciences promises to deliver the necessary tools — but questions remain about the scalability, interpretability and trustworthiness of the approach. 

 

After an accessible introduction to machine learning in cosmology, I will present recent advances in simulation-based inference techniques for fast, scalable inference with guaranteed coverage. I will discuss a general, statistically principled solution to the ubiquitous problem of covariate shift in supervised learning. While the case studies will be of cosmological nature, the methodology is entirely general and applicable to many other scientific problems.