BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:On Different Distances Between Distributions and Generative Advers
 arial Networks - Martin Arjovsky
DTSTART:20170525T100000Z
DTEND:20170525T110000Z
UID:TALK72839@talks.cam.ac.uk
CONTACT:12852
DESCRIPTION:Generative adversarial networks (GANs) are notoriously difficu
 lt to train. At the core of it\, we show that these problems arise natural
 ly when trying to learn distributions whose support lie in low dimensional
  manifolds. We show how these problems are consequences of trying to optim
 ize the classical divergences (KL\, JSD\, etc) between our real and data d
 istribution\, and that these are symptoms of a more general phenomenon\, p
 ointing towards the inefficacy of the usual divergences in certain setting
 s. After that\, we bring into play the Wasserstein distance\, which we pro
 ve doesn't suffer from the same behaviour\, and provide a first step on an
  algorithm that tries to approximately optimize this distance.\n\n\nReleva
 nt papers:\n\n"Towards Principled Methods for Training Generative Adversar
 ial Networks":https://arxiv.org/abs/1701.04862\n\n"Wasserstein GAN":https:
 //arxiv.org/abs/1701.07875\n
LOCATION:JDB Seminar Room\, CUED
END:VEVENT
END:VCALENDAR