
Dr. Michael Brückner: "Probabilistic Programming is NUTS"
01. Nov.
Im Rahmen des Forschungskolloquiums Wirtschaftsinformatik und Data Science referiert Dr. Michael Brückner von Amazon über "Probabilistic Programming is NUTS".
Datum und Ort: 1. November 2018 12:15 Uhr C12.010
Abstract:
I will first give a brief introduction in probabilistic programming to motivate the need for scalable probabilistic inference methods. Then, I will explain the general idea of sampling-based probabilistic inference, in particular focusing on Markov Chain Monte Carlo (MCMC) methods. After presenting the Metropolis-Hastings MCMC algorithm, and giving an intuition why it is so slow, I will introduce Hamiltonian Monte Carlo and the No-U-turn-Sampler (NUTS). At the end of the talk you should understand what Hamiltonian Monte Carlo has in common with this skateboarder.