We will have a special cosmolunch tomorrow. Johannes Lange ( https://johannesulf.github.io/ ) will visit us. The title and abstract of his talk is below.
Please sign up https://docs.google.com/spreadsheets/d/1ykItYZSGeTDJnbeJzOemvCDf0jMeuu3aZ5Pp_LA5YlQ/edit ?usp=sharing if you want to chat with him.
A New High-Efficiency Bayesian Sampler for Astronomy
In Bayesian statistics, we need Monte-Carlo methods to sample from complex posterior distributions and estimate their normalizing constants, the so-called Bayesian evidence. I introduce a novel approach to boost the efficiency of the Importance Nested Sampling (INS) technique using neuralnetwork regression. Unlike rejection-based sampling methods such as traditional nested sampling (NS) or Markov Chain Monte Carlo (MCMC) algorithms, importance sampling techniques can use all likelihood evaluations for posterior and evidence estimation. However, for importance sampling to be efficient, one needs a proposal distribution that closely mimics the target distribution. I show how one can combine INS with deep learning to accomplish this task. I then introduce nautilus (https://github.com/johannesulf/nautilus ), a reference open-source Python implementation of this technique for Bayesian posterior and evidence estimation. I compare nautilus against popular NS and MCMC packages on a variety of synthetic problems and real-world applications in exoplanet detection, galaxy SED fitting and cosmology. I show that nautilus delivers highly accurate results and that it needs substantially fewer likelihood evaluations than existing samplers.