BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Generalized Energy Based Models - Michael Arbel - UCL
DTSTART:20201030T131500Z
DTEND:20201030T141500Z
UID:TALK153616@talks.cam.ac.uk
CONTACT:Francisco Vargas
DESCRIPTION:*Paper:*\n\nTalk is based on "this":https://arxiv.org/pdf/2003
 .05033.pdf paper.\n\n*Abstract:*\n\nWe introduce the Generalized Energy Ba
 sed Model (GEBM) for generative modelling. These models combine two  train
 ed components: a base distribution (generally an implicit model)\, which c
 an learn the support of data with low intrinsic dimension in a high dimens
 ional space\; and an energy function\, to refine the probability mass on t
 he learned support. Both the energy function and base jointly constitute t
 he final model\, unlike GANs\, which retain only the base distribution (th
 e "generator"). GEBMs are trained by alternating between learning the ener
 gy and the base. We show that both training stages are well-defined: the e
 nergy is learned by maximising a generalized likelihood\, and the resultin
 g energy-based loss provides informative gradients for learning the base. 
 Samples from the posterior on the latent space of the trained model can be
  obtained via MCMC\, thus finding regions in this space that produce bette
 r quality samples. Empirically\, the GEBM samples on image-generation task
 s are of much better quality than those from the learned generator alone\,
  indicating that all else being equal\, the GEBM will outperform a GAN of 
 the same complexity. When using normalizing flows as base measures\, GEBMs
  succeed on density modelling tasks\, returning comparable performance to 
 direct maximum likelihood of the same networks.\n\n*About the Speaker:*\n\
 nTBC\n\n*Website:* https://michaelarbel.github.io/\n\nPart of ML@CL Semina
 r Series focusing on early career researchers in topics relevant to machin
 e learning and statistics.
LOCATION: https://dtudk.zoom.us/j/64671492015?pwd=ZkZyd1BpSHNvR3JQNDlWbXZ1
 aG9CUT09
END:VEVENT
END:VCALENDAR