A broad class of unsupervised deep learning methods such as Generative Adversarial Networks (GANs) involve training of overparameterized models where the number of parameters of the model exceeds a certain threshold. Indeed, most successful GANs used in practice are trained using overparameterized generator and discriminator networks, both in terms of depth and width. A large body of work in supervised learning have shown the importance of model overparameterization in the convergence of the gradient descent (GD) to globally optimal solutions. In contrast, the unsupervised setting and GANs in particular involve non-convex concave mini-max optimization problems that are often trained using Gradient Descent/Ascent (GDA). The role and benefits of model overparameterization in the convergence of GDA to a global saddle point in non-convex concave problems is far less understood. In this work, we present a comprehensive analysis of the importance of model overparameterization in GANs both theoretically and empirically. We theoretically show that in an overparameterized GAN model with a 1-layer neural network generator and a linear discriminator, GDA converges to a global saddle point of the underlying non-convex concave min-max problem. To the best of our knowledge, this is the first result for global convergence of GDA in such settings.more »
On the Limitations of First-Order Approximation in GAN Dynamics
While Generative Adversarial Networks (GANs) have demonstrated promising performance on multiple vision tasks, their learning dynamics are not yet well understood, both in theory and in practice. To address this issue, we study GAN dynamics in a simple yet rich parametric model that exhibits several of the common problematic convergence behaviors such as vanishing gradients, mode collapse, and diverging or oscillatory behavior. In spite of the non-convex nature of our model, we are able to perform a rigorous theoretical analysis of its convergence behavior. Our analysis reveals an interesting dichotomy: a GAN with an optimal discriminator provably converges, while first order approximations of the discriminator steps lead to unstable GAN dynamics and mode collapse. Our result suggests that using first order discriminator steps (the de-facto standard in most existing GAN setups) might be one of the factors that makes GAN training challenging in practice.
- Publication Date:
- NSF-PAR ID:
- Journal Name:
- Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden
- Sponsoring Org:
- National Science Foundation
More Like this
Generative adversarial networks (GANs) are innovative techniques for learning generative models of complex data distributions from samples. Despite remarkable recent improvements in generating realistic images, one of their major shortcomings is the fact that in practice, they tend to produce samples with little diversity, even when trained on diverse datasets. This phenomenon, known as mode collapse, has been the main focus of several recent advances in GANs. Yet there is little understanding of why mode collapse happens and why recently proposed approaches are able to mitigate mode collapse. We propose a principled approach to handling mode collapse, which we call packing. The main idea is to modify the discriminator to make decisions based on multiple samples from the same class, either real or artificially generated. We borrow analysis tools from binary hypothesis testing—in particular the seminal result of Blackwell —to prove a fundamental connection between packing and mode collapse. We show that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process. Numerical experiments on benchmark datasets suggests that packing provides significant improvements in practice as well.
Video summarization aims to simplify large-scale video browsing by generating con- cise, short summaries that diver from but well represent the original video. Due to the scarcity of video annotations, recent progress for video summarization concentrates on unsupervised methods, among which the GAN-based methods are most prevalent. This type of methods includes a summarizer and a discriminator. The summarized video from the summarizer will be assumed as the final output, only if the video reconstructed from this summary cannot be discriminated from the original one by the discriminator. The primary problems of this GAN-based methods are two-folds. First, the summarized video in this way is a subset of original video with low redundancy and contains high priority events/entities. This summarization criterion is not enough. Second, the training of the GAN framework is not stable. This paper proposes a novel Entity–relationship Aware video summarization method (ERA) to address the above problems. To be more spe- cific, we introduce an Adversarial Spatio-Temporal network to construct the relationship among entities, which we think should also be given high priority in the summarization. The GAN training problem is solved by introducing the Wasserstein GAN and two newly proposed video-patch/score-sum losses. In addition, the score-summore »
Generative adversarial networks (GANs) are powerful tools for learning generative models. In practice, the training may suffer from lack of convergence. GANs are commonly viewed as a two-player zero-sum game between two neural networks. Here, we leverage this game theoretic view to study the convergence behavior of the training process. Inspired by the fictitious play learning process, a novel training method, referred to as Fictitious GAN, is introduced. Fictitious GAN trains the deep neural networks using a mixture of historical models. Specifically, the discriminator (resp. generator) is updated according to the best-response to the mixture outputs from a sequence of previously trained generators (resp. discriminators). It is shown that Fictitious GAN can effectively resolve some convergence issues that cannot be resolved by the standard training approach. It is proved that asymptotically the average of the generator outputs has the same distribution as the data samples.
In this paper, we present a simple approach to train Generative Adversarial Networks (GANs) in order to avoid a mode collapse issue. Implicit models such as GANs tend to generate better samples compared to explicit models that are trained on tractable data likelihood. However, GANs overlook the explicit data density characteristics which leads to undesirable quantitative evaluations and mode collapse. To bridge this gap, we propose a hybrid generative adversarial network (HGAN) for which we can enforce data density estimation via an autoregressive model and support both adversarial and likelihood framework in a joint training manner which diversify the estimated density in order to cover different modes. We propose to use an adversarial network to transfer knowledge from an autoregressive model (teacher) to the generator (student) of a GAN model. A novel deep architecture within the GAN formulation is developed to adversarially distill the autoregressive model information in addition to simple GAN training approach. We conduct extensive experiments on real-world datasets (i.e., MNIST, CIFAR-10, STL-10) to demonstrate the effectiveness of the proposed HGAN under qualitative and quantitative evaluations. The experimental results show the superiority and competitiveness of our method compared to the baselines.