Speaker: Justin Domke is an associate professor at the University of Massachusetts, Amherst. He studies probabilistic machine learning. He recently received the NSF CAREER award. Previously, he was a researcher at National ICT Australia and a PhD student at the University of Maryland, College Park.
Title: Advances in black-box variational inference
Abstract: Variational inference (VI) shows impressive results on particular models, but Markov chain Monte Carlo (MCMC) remains the champion for "black-box" inference on generic user-defined models. This talk with describe three advances to push VI closer to this goal. The first is a strategy to automatically find a gradient estimator with the best trade-off of computational time against variance. The second is a method to make data subsampling possible in hierarchical models, where the number of latent variables grows with the size of the data. The third is a method to integrate an MCMC procedure into VI to deliver more accurate bound.
If you would like to attend, please email email@example.com for the Zoom details.