An Online Learning Approach to Generative Adversarial Networks

By Paulina Grnarova, Kfir Yehuda Levy, Aurélien Lucchi, Thomas Hofmann, Andreas Krause. 2017

92 citations recorded by the source corpus.

We consider the problem of training generative models with a Generative Adversarial Network (GAN). Although GANs can accurately model complex distributions, they are known to be difficult to train due to instabilities caused by a difficult minimax optimization problem. In this paper, we view the problem of training GANs as finding a mixed strategy in a zero-sum game. Building on ideas from online learning we propose a novel training method named Chekhov GAN 1 . On the theory side, we show that our method provably converges to an equilibrium for semi-shallow GAN architectures, i.e. architectures where the discriminator is a one layer network and the generator is arbitrary. On the practical side, we develop an efficient heuristic guided by our theoretical results, which we apply to commonly used deep GAN architectures. On several real world tasks our approach exhibits improved stability and performance compared to standard GAN training.

Subfields: Generative Models

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

Citations in the corpus

Self-Supervised GANs via Auxiliary Rotation Loss (2018, background): “This challenge has received a great deal of attention and explicit temporal dependencies have been proposed to improve training in this setting [2, 15, 17, 19].”
GANs Trained by a Two Time-Scale Update Rule Converge to a Nash Equilibrium (2017, background): “For special GAN variants, convergence can be proved under certain assumptions [9, 21], as can local stability [41].”
A Variational Inequality Perspective on Generative Adversarial Networks (2018, background): “Interesting non-convex results were proved, for a new notion of regret minimization, by Hazan et al. (2017) and in the context of online learning for GANs by Grnarova et al. (2018).”
InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning (2020, methodology): “Previous approaches to mitigate catastrophic forgetting in GANs include using forms of memory [8], [57], [58], such as checkpoint averaging.”
Solving Weakly-Convex-Weakly-Concave Saddle-Point Problems as Successive Strongly Monotone Variational Inequalities (2018, background): “However, their results are either asymptotic [22, 39, 49] or their analysis require strong assumptions of the problem [49, 33] (e.g., the problem is concave in the max variable).”
Minimax Theorem for Latent Games or: How I Learned to Stop Worrying about Mixed-Nash and Love Neural Nets (2020, background): “Many papers have adopted a game theoretic perspective to understand GANs and motivate the computation of mixed equilibria (Arora et al., 2017; Oliehoek et al., 2018; Hsieh et al., 2019; Grnarova et al., 2018).”
From Poincaré Recurrence to Convergence in Imperfect Information Games: Finding Equilibrium via Regularization (2020, background): “, 2019) or game theoretic algorithms (Oliehoek et al., 2017; Grnarova et al., 2018).”
Evaluating GANs via Duality (2018, unknown): “…have analyzed the convergence properties of alternating gradient-based techniques, most of them require the objective function to be convex-concave (Goodfellow et al., 2014; Nowozin et al., 2016) with the exception of Grnarova et al. (2017) who prove convergence for semi-concave zero-sum games.”
Skill Rating for Generative Models (2018, methodology): “We additionally trained a separate Chekhov GAN.”
On Gradient Descent Ascent for Nonconvex-Concave Minimax Problems (2019, background): “…emerged as a focus in optimization and machine learning [Namkoong and Duchi, 2016, Sinha et al., 2018, Rafique et al., 2018, Sanjabi et al., 2018, Grnarova et al., 2018, Lu et al., 2019, Nouiehed et al., 2019, Thekumparampil et al., 2019, Kong and Monteiro, 2019]; see Table 1 for a comprehensive…”
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications (2020, methodology): “GANs have also been used in other machine learning tasks [290] such as active learning [291], [292], online learning [293], ensemble learning [294], zero-shot learning [295], [296], and multi-task learning [297].”
How Generative Adversarial Networks and Its Variants Work : An Overview of GAN (2017, background): “There still remains open problems in theoretical aspects of GAN in whether GAN really converges and it can perfectly overcome mode collapse, as [31], [124], [125] discuss.”
Minmax Optimization: Stable Limit Points of Gradient Descent Ascent are Locally Optimal (2019, background): “All of these results require, however, access to oracles which solve the minimization and maximization problems separately, keeping the other variable ﬁxed and outputting a mixed Nash equilibrium [see, e.g., Feige et al., 2015, Chen et al., 2017, Grnarova et al., 2017, Gonen and Hazan, 2018].”
A Variational Inequality Perspective on Generative Adversarial Nets (2018, background): “A study of the continuous version of two player games has been done by Ratliff et al. (2016) and interesting non-convex results have been proved in the context of GANs by Grnarova et al. (2018).”
The History Began from AlexNet: A Comprehensive Survey on Deep Learning Approaches (2018, methodology): “GANs are used for online learning in which it is being trained for finding a mixed strategy in a zero-sum game which is named Checkov GAN 1[224].”
On the Limitations of First-Order Approximation in GAN Dynamics (2017, background): “The recent work (Grnarova et al., 2018) views GAN training through the lens of online learning.”
A Domain Agnostic Measure for Monitoring and Evaluating GANs (2018, background): “Some exceptions are [4, 9, 7, 12] but these works do not address the problem of evaluation.”