Implicit Likelihood Inference: From Emulators to Cosmological Constraints

Date:

The Implicit Likelihood Inference (ILI) working group develops and applies machine-learning techniques for cosmological inference in settings where the likelihood cannot be written down explicitly. Instead, the posterior distribution is directly learned from forward simulations, making it easy to incorporate non-trivial masks, selection effects or scale-cuts, provided the simulations are treated in the same way as the data. Deaglan Bartlett will outline the group’s efforts to scale these techniques for galaxy clustering, which is challenging given the large volume yet high resolution needed for the forward simulations. Bartlett will discuss recent developments in the tools that make this kind of large-scale inference tractable, including sensitivity analyses that diagnose which aspects of a simulation drive constraining power, summary statistics that combine explicit and learned compression, emulators which produce fast yet high-fidelity forward simulations, and validation checks to ensure the group’s results are robust to a variety of types of model-misspecification. Bartlett will give an overview of progress on the LtU-ILI software framework, which provides standardized access to a range of simulation-based inference methods and underpins several of the collaboration’s flagship analyses of survey data. Together, this work aims to make implicit inference both more trustworthy and more efficient as we move toward survey-scale analyses with real data.

Watch the talk here