Abstract

The technology that underpins machine learning – differentiable programming – is poised to revolutionise astronomy, making it possible for the first time to fit very high dimensional models: hierarchical models describing many objects; the sensitivity of millions of pixels in a detector; models of images or spectra with very many free parameters; or neural networks that represent physics we cannot easily solve in closed form. It also enables fundamental information-theoretic quantities like the Fisher information to be calculated, allowing for determination and optimization of the information content of an experiment. I will discuss how we apply this to the James Webb interferometer experiment, to provide a data-driven self-calibration of the telescope’s highest resolution mode and its difficult systematics; to design the Toliman Space Telescope to do high-precision, distortion-tolerant astrometry; and give an overview of related work on interferometry, transits and AGN reverberation mapping in our group.