BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Towards Better Understanding of Likelihood-based Generative Models
DTSTART:20190703T101500
DTEND:20190703T121500
DTSTAMP:20260924T102305Z
UID:04ceeed57dfc8cd46681edbf39ce96d943cda1aa8b256e33ec39de04
CATEGORIES:Conferences - Seminars
DESCRIPTION:Mladen Dimovski\nEDIC candidacy exam\nExam president: Prof. Ol
 ivier Lévêque\nThesis advisor: Prof. Patrick Thiran\nCo-examiner: Prof. 
 François Fleuret\n\nAbstract\nLikelihood-based modes are an important cla
 ss of generative models that explicitly define a model density and optimiz
 e its parameters by maximizing the likelihood of the observed data.  Alth
 ough the maximum likelihood approach is the method of choice for tradition
 al density estimation\, its usage as a training objective for approximatin
 g arbitrary probability distributions in high dimension raises some critic
 al questions.\nIn this talk\, we present and analyze two classes of widely
  influential likelihood-based models used for generative modelling\, the f
 low-based models and the variational auto-encoders. We examine their poten
 tial\, their intrinsic limitations and the extent to which these can be ov
 ercome. Finally\, we discuss a work that proposes a novel method for gener
 ative model evaluation\, an important aspect of research in the domain.\n\
 nBackground papers\nDiagnosing and Enhancing VAE Models\, by Bin Dai\, Da
 vid Wip. International Conference on Learning Representations\, 2019.\nMa
 sked Autoregressive Flow for Density Estimation\, by George Papamakarios 
 et al. Advances in Neural Information Processing Systems\, 2017.\nAssessi
 ng Generative Models via Precision and Recall\, by Mehdi S. M. Sajjadi et
  al. Advances in Neural Information Processing Systems\, 2018.
LOCATION:BC 329 https://plan.epfl.ch/?room==BC%20329
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR