Evaluation of Mode Collapse in Generative Adversarial Networks

Sayeri Lala, Maha Shady, Anastasiya Belyaeva, Molei Liu

Abstract:

Though widely used in imaging tasks, General
Adversarial Networks suffer from mode collapse. Various novel
Generative Adversarial Network architectures have been proposed to address the issue. However, there exists little work
systematically comparing their performance. This work compares
state of the art Generative Adversarial Network architectures, in
particular AdaGAN, VEEGAN, Wasserstein GAN, and Unrolled
GAN, on datasets with different distributions, include synthetic
and real datasets and with respect to several commonly used
metrics for quantifying mode collapse. Our findings suggest
that AdaGAN performs consistently better among the GANs on
nearly all datasets and Wassertstein GAN performs poorly on the
datasets. Our results also suggest that one metric is not sufficient
to quantify mode collapse for GANs as the metrics do not give
consistent results