BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:BSU Seminar: &quot\;Sparse Hamiltonian Flows (or: Bayesian Coreset
 s Without all the Fuss)&quot\; - Prof Trevor Campbell\, The University of 
 British Columbia 
DTSTART:20220426T130000Z
DTEND:20220426T140000Z
UID:TALK172058@talks.cam.ac.uk
CONTACT:Alison Quenault
DESCRIPTION:Bayesian inference provides a coherent approach to learning fr
 om data and uncertainty assessment in complex\, expressive statistical mod
 els. However\, algorithms for performing inference have not yet caught up 
 to the deluge of data in modern applications. One approach---Bayesian core
 sets---involves replacing the large dataset with a small\, weighted\, repr
 esentative subset of data during inference. The coreset is designed to cap
 ture the information from the full dataset\, but be much less computationa
 lly expensive to store in memory and iterate over. Although the methodolog
 y is sound in principle\, efficiently constructing such a coreset in pract
 ice remains a significant challenge: current methods tend to be complicate
 d to implement\, slow\, require a secondary inference step after coreset c
 onstruction\, and do not enable model selection. In this talk\, I will int
 roduce a new method---sparse Hamiltonian flows---that addresses all of the
 se challenges. The method involves first subsampling the data uniformly\, 
 and then optimizing a Hamiltonian flow parametrized by coreset weights and
  including periodic momentum quasi-refreshment steps. I will present theor
 etical results demonstrating that the method enables an exponential compre
 ssion of the dataset in representative models\, and that the quasi-refresh
 ment steps reduce the KL divergence to the target. Real and synthetic expe
 riments demonstrate that sparse Hamiltonian flows provide accurate posteri
 or approximations with significantly reduced runtime compared with competi
 ng dynamical-system-based inference methods.\n\nThis talk will be based on
  two papers that are available online as preprints:\nChen et al\, "Bayesia
 n coresets via sparse Hamiltonian flows\," https://arxiv.org/abs/2203.0572
 3\nNaik et al\, "Fast Bayesian coresets via subsampling and quasi-Newton r
 efinement\," https://arxiv.org/abs/2203.09675\n
LOCATION:This will be a virtual seminar. FREE registration: https://us02we
 b.zoom.us/meeting/register/tZ0rc-2vpz4qHNxdJ4Cm4hyLEn4BVOG5YcIt 
END:VEVENT
END:VCALENDAR