On the convergence properties of GAN training

By Lars M. Mescheder. 2018

86 citations recorded by the source corpus.

Subfields: Generative Models, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Generative Models

In Optimization & Training

Citations in the corpus

High-Resolution Image Synthesis with Latent Diffusion Models (2021, background): “cult to optimize [2, 27, 53] and struggle to capture the full data distribution [54].”
Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling (2018, methodology): “We produce CelebAHQ-256 samples (at full 8-bit resolution) that are of similar visual ﬁdelity to those produced with methods such as GANs that lack however an intrinsic measure of generalization (Mescheder, 2018; Karras et al., 2017).”
Negative Momentum for Improved Game Dynamics (2018, background): “Min-Max Bilinear Game [Figure 3] In our first set of experiments, we showcase the effect of negative momentum in a bi-dimensional min-max optimization setup (3) where φ,θ ∈ R and M = 1 (Mescheder, 2018).”
A Variational Inequality Perspective on Generative Adversarial Networks (2018, methodology): “The bilinear objective has been extensively used (Goodfellow, 2016; Mescheder, 2018; Yadav et al., 2018) to highlight the difficulties of gradient descent for saddle point optimization.”
The Mechanics of n-Player Differentiable Games (2018, background): “Nagarajan & Kolter (2017) prove convergence for some algorithms, but under very strong assumptions Mescheder (2018).”
The Unusual Effectiveness of Averaging in GAN Training (2018, background): “e prior. For extended results refer to the Appendix. 4.1 Illustrative Examples 4.1.1 Bilinear Game In order to understand the effects of averaging, we present results on a bilinear game min max ˚˚ [Mescheder, 2018]. In this game, the true data distribution is an impulse function at point 0. Both gen”
pi-GAN: Periodic Implicit Generative Adversarial Networks for 3D-Aware Image Synthesis (2020, unknown): “We use the non-saturating GAN loss with R1 regularization [36]:”
DeepLandscape: Adversarial Modeling of Landscape Video (2020, methodology): “The model is trained within standard GAN approach with non-saturating loss [9] with R1 regularization [30] as in the original StyleGAN paper.”
Towards Photographic Image Manipulation with Balanced Growing of Generative Autoencoders (2019, background): “Many recent improvements improve the stability and robustness of the training process by suggesting new loss functions [1], regularization methods [29, 31, 28], multi-resolution training [17, 39], architectures [18], or combinations of these.”
A Variational Inequality Perspective on Generative Adversarial Nets (2018, methodology): “The bilinear objective has been extensively used (Goodfellow, 2016; Mescheder, 2018; Yadav et al., 2018) to highlight the difficulties of gradient descent for saddle point optimization.”