Dr Adrian Bayer (Princeton) works on the intersection of astrophysics and machine learning. Recent interests include: creating multi-probe cosmological simulations for galaxy, lensing, and CMB surveys; performing optimal and interpretable cosmological inference by sampling high-dimensional parameter spaces to reconstruct the Universe's initial conditions and using simulation-based inference; and developing statistical techniques to find astronomical signals in large and noisy spaces, such as for super massive black hole binaries and exoplanets.
Title: What’s the Likelihood? Field-Level Cosmology x Detecting Signals in Large and Noisy Places
Abstract: During this era of Big Data, astrophysicists are analysing vast parameter spaces to perform inference or to detect signals supporting new theories. A ubiquitous question that arises in this process is, “What’s the likelihood?“. I will discuss various ways likelihoods are used, and what one needs to keep in mind while using them, by considering two case studies: 1) cosmological field-level inference and 2) detecting supermassive black hole binary mergers.
First, I will motivate field-level inference as an optimal approach for extracting information from cosmic structure and reconstructing the initial conditions of the Universe. I will discuss different methods of field-level inference, ranging from explicit likelihood analyses using differentiable forward modeling to implicit likelihood analyses using neural networks. I will then highlight the exciting potential for using these methods to improve BAO constraints with DESI. Then, I will discuss searches for rare objects, such as supermassive black hole binaries, presenting novel data-driven methods to correct traditional likelihood analyses for the look-elsewhere effect and for unknown noise properties in the data, showing that a careful treatment of the likelihood can lead to vastly different conclusions.