ON the tradeoff between mode collapse and sample quality in generative adversarial networks

Sudarshan Adiga, Mohamed Adel Attia, Wei Ting Chang, Ravi Tandon

Research output: Chapter in Book/Report/Conference proceedingConference contribution

22 Scopus citations

Abstract

Generative Adversarial Networks (GAN) are used to generate synthetic samples while closely following the underlying distribution of a real data set. While GANs have recently gained significant popularity, they often suffer from the mode collapse problem, where the generated samples lack diversity. Moreover, some approaches that attempt to resolve the model collapse problem do not necessarily yield high quality synthetic samples. In this paper, we propose two novel performance metrics, namely mode-collapse divergence (MCD) which quantifies the extent of mode collapse for a GAN architecture. Second, we propose the metric Generative Quality Score (GQS), which measures the quality of generated samples. We present a comprehensive study of the performance of various GAN architectures proposed in the literature through the lens of MCD and GQS, for three different data sets, namely MNIST, Fashion MNIST and CIFAR-10.

Original languageEnglish (US)
Title of host publication2018 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2018 - Proceedings
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1184-1188
Number of pages5
ISBN (Electronic)9781728112954
DOIs
StatePublished - Jul 2 2018
Event2018 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2018 - Anaheim, United States
Duration: Nov 26 2018Nov 29 2018

Publication series

Name2018 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2018 - Proceedings

Conference

Conference2018 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2018
Country/TerritoryUnited States
CityAnaheim
Period11/26/1811/29/18

Keywords

  • Generative Adversarial Networks
  • Mode Collapse
  • Performance Metrics

ASJC Scopus subject areas

  • Information Systems
  • Signal Processing

Fingerprint

Dive into the research topics of 'ON the tradeoff between mode collapse and sample quality in generative adversarial networks'. Together they form a unique fingerprint.

Cite this