Monte Carlo Seminar| Bob Carpenter| How does Stan work?

Orgazined by Online Monte Carlo Seminar [sites.google.com/view/monte-carlo-seminar/] Speaker: Bob Carpenter (Flatiron Institute) Title: How does Stan work? Abstract: In this talk, I'll describe the Stan probabilistic programming language and statistical inference stack in technical detail. I will start with an overview of Stan's language from a user's perspective with examples of use. Then I'll drill down into how the programming language works at a technical level, focusing on reverse-mode automatic differentiation, transforms for constrained parameters, and posterior predictive quantities. Next, I'll describe how the differentiable log density defined by a Stan program can be used for Bayesian inference. I'll describe the details of Stan's multinomial no-U-turn sampler (NUTS) and its adaptation, as well as those of its variational inference (Pathfinder and ADVI), and Laplace approximation. I'll conclude by covering some of the packages that depend on Stan in a fundamental way, such as BridgeStan, brms, and Prophet, which collectively have far more users than Stan itself. #hamiltonian #montecarlo #mcmc