Double Descent and Other Interpolation Phenomena in GANs

arXiv:2106.04003v2 Announce Type: replace
Abstract: We study overparameterization in generative adversarial networks (GANs) that can interpolate the training data. We show that overparameterization can improve generalization performance and accelerate the training process. We study the generalization error as a function of latent space dimension and identify two main behaviors, depending on the learning setting. First, we show that overparameterized generative models that learn distributions by minimizing a metric or $f$-divergence do not exhibit double descent in generalization errors; specifically, all …

abstract adversarial arxiv cs.lg data error function gans generative generative adversarial networks identify interpolation networks performance process show space study training training data type