Accelerating Bayesian Inference with Deep Learning

Bayesian inference in cosmology often incurs high computational costs. One needs Monte-Carlo methods to sample from complex posterior distributions and estimate their normalizing constants, the so-called Bayesian evidence. Frequently, the number of model evaluations required is of order million and larger. In this talk, I will discuss a novel algorithm for Bayesian inference that combines Importance Nested Sampling with ensemble deep learning. I show that this new algorithm can significantly outperform existing Bayesian inference codes on various problems, ranging from synthetics likelihoods to real-world applications, in terms of computational cost and accuracy. I also present an open-source Python implementation of this algorithm called nautilus and its application in cosmology.