Traditional analysis of cosmological Large Scale Structures assumes that the likelihood is Gaussian, which unfortunately is not optimal and may even be biased in the non linear regime. Simulation Based Inference (SBI), also known as Likelihood Free Inference or Implicit Likelihood Inference, directly models the Bayesian optimal posterior, but typically needs a large number of simulations to converge. In this talk, I will first summarize several widely-used neural-network-based SBI methods, and then introduce Neural Quantile Estimation (NQE), a new approach for SBI which achieves state-of-the-art performance on a variety of benchmark problems. I will also present some preliminary results on the application of NQE to the optimal analysis of 2-dim weak lensing maps.
BIO
He Jia (贾赫) is a PhD Candidate at Department of Astrophysical Sciences, Princeton University, who obtained his BS in Physics at Peking University in 2020. He is broadly interested in Black Holes, Cosmology and Machine Learning, with a focus on how to constrain fundamental physics with astrophysical observations. Currently, he is studying what we can learn from Event Horizon Telescope images, photon ring autocorrelations, as well as future space-based VLBI observations of Super Massive Black Holes. He also works on extracting information from cosmological Large Scale Structures with Simulation Based Inference.
Host: Cheng Zhao
