Implicit Likelihood Inference

People

Leads: Matthew Ho, Benjamin Wandelt

Members: Ana Maria Delgado, Anirban Bairagi, Ben Wandelt, Carolina Cuesta-Lazaro, ChangHoon Hahn, Chirag Modi, Christian Jespersen, Deaglan Bartlett, Justin Alsing, Leander Thiele, Lucas Makinen, Lucia Perez, Matthew Ho, Matthew Scoggins, Niall Jeffrey, Nicolas Chartier, Rosa Malandrino, Shirley Ho, Shivam Pandey, Xiaosheng Zhao,

Description

The Implicit Likelihood Inference (ILI) working group develops state-of-the-art techniques to perform observational inference for problems where the full, analytic likelihood cannot be easily described. This is achieved through the use of machine learning models, which learn the likelihood of observational data directly from simulations.

The ILI group also creates and maintains generalized software solutions to tackle the diverse physical challenges studied by the Learning the Universe (LtU) collaboration. This includes the development of the LtU-ILI framework (Ho et al., 2024), a highly configurable and user-friendly software package that provides access to multiple simulation-based inference techniques. The framework facilitates the comparison of different ILI methods on standardized test problems, making these cutting-edge tools more accessible to the broader scientific community. Additionally, the group explores high-dimensional inference challenges, including those complementary to the BORG group’s efforts to infer three-dimensional cosmological initial conditions (Legin et al., 2024).

ILI innovates on these techniques, developing new state-of-the-art tools for performing automated inference on astrophysical datasets. These include: optimal information extraction and compression, amortized estimation of Bayesian evidence (Jeffrey & Wandelt, 2024), information aggregation from set-based data (Makinen et al., 2023), field-level inference (de Santi Natalı́ S. M. et al., 2023), sensitivity tests (Modi et al., 2023), and machine learning interpretability tools (Ho et al., 2023). These works are developed out of necessity for the problems studied in the LtU collaboration, but are practically applicable to a wide range of scientific inference problems.

ILI also leads the flagship LtU cosmological survey analysis of the CMASS galaxy sample from the Sloan Digital Sky Survey (SDSS). Incorporating tools from the AFM, Robustness, and SynthObs groups, we have developed a modular simulation pipeline to rapidly generate mock galaxy catalogs for the CMASS sample. We then apply ILI techniques to infer cosmological parameters from real data. To date, this is the largest application of Implicit Likelihood Inference to a spectroscopic galaxy survey and the first to use accelerated forward models to emulate full N-body simulations.

Projects

• SDSS-CMASS Cosmological Analysis
• Cosmic Graphs: the Language of Large-Scale Structure
• Fishnets
• Hybrid Summary Statistics
• Information-Ordered Bottleneck
• LtU-ILI