Generalization and Equilibrium in Generative Adversarial Nets (GANs)

Abstract

We show that training of generative adversarial network (GAN) may not have good generalization properties; e.g., training may appear successful but the trained distribution may be far from target distribution in standard metrics. However, generalization does occur for a weaker metric called neural net distance. It is also shown that an approximate pure equilibrium exists in the discriminator/generator game for a special class of generators with natural training objectives when generator capacity and training set sizes are moderate. This existence of equilibrium inspires MIX+GAN protocol, which can be combined with any existing GAN training, and empirically shown to improve some of them.

This is an updated version of an ICML'17 paper with the same title. The main difference is that in the ICML'17 version the pure equilibrium result was only proved for Wasserstein GAN. In the current version the result applies to most reasonable training objectives. In particular, Theorem 4.3 now applies to both original GAN and Wasserstein GAN

Cited by in corpus (86)

• Generative Adversarial Networks: An Overview
• How Generative Adversarial Networks and Their Variants Work: An Overview
• Maximum Density Divergence for Domain Adaptation
• GAN-Leaks: A Taxonomy of Membership Inference Attacks against Generative Models
• Demystifying MMD GANs
• Integrating Expert Knowledge with Domain Adaptation for Unsupervised Fault Diagnosis
• Prescribed Generative Adversarial Networks
• A Selective Overview of Deep Learning
• Approximability of Discriminators Implies Diversity in GANs
• FLAMBE: Structural Complexity and Representation Learning of Low Rank MDPs
• Wasserstein Generative Adversarial Uncertainty Quantification in Physics-Informed Neural Networks
• Small-GAN: Speeding Up GAN Training Using Core-sets
• Generalization in Generative Adversarial Networks: A Novel Perspective from Privacy Protection
• GANs May Have No Nash Equilibria
• Minimax Distribution Estimation in Wasserstein Distance
• On Catastrophic Forgetting and Mode Collapse in Generative Adversarial Networks
• Error Bounds of Imitating Policies and Environments
• Vicinal and categorical domain adaptation
• Generative Latent Flow
• Some Theoretical Insights into Wasserstein GANs
• An error analysis of generative adversarial networks for learning distributions
• Imperfect ImaGANation: Implications of GANs Exacerbating Biases on Facial Data Augmentation and Snapchat Selfie Lenses
• Implicit competitive regularization in GANs
• This Person (Probably) Exists. Identity Membership Attacks Against GAN Generated Faces
• Generalization Properties of Optimal Transport GANs with Latent Distribution Learning
• SGD Learns One-Layer Networks in WGANs
• Deep Extrapolation for Attribute-Enhanced Generation
• Private Post-GAN Boosting
• Investigating Under and Overfitting in Wasserstein Generative Adversarial Networks
• Learning from a Complementary-label Source Domain: Theory and Algorithms
• On Computation and Generalization of Generative Adversarial Imitation Learning
• Limit Distribution Theory for the Smooth 1-Wasserstein Distance with Applications
• Contrastively Smoothed Class Alignment for Unsupervised Domain Adaptation
• A mean-field analysis of two-player zero-sum games
• Capacity Bounded Differential Privacy
• Asymptotic Guarantees for Generative Modeling Based on the Smooth Wasserstein Distance
• How Does GAN-based Semi-supervised Learning Work?
• Quantifying and Improving Transferability in Domain Generalization
• Decision-Aware Conditional GANs for Time Series Data
• On Characterizing GAN Convergence Through Proximal Duality Gap
• CDE-GAN: Cooperative Dual Evolution Based Generative Adversarial Network
• Generalized Energy Based Models
• Cycle-consistent Conditional Adversarial Transfer Networks
• Approximating Lipschitz continuous functions with GroupSort neural networks
• Image Hashing by Minimizing Discrete Component-wise Wasserstein Distance
• Augmentation-Interpolative AutoEncoders for Unsupervised Few-Shot Image Generation
• A study of semi-supervised speaker diarization system using gan mixture model
• GANs with Variational Entropy Regularizers: Applications in Mitigating the Mode-Collapse Issue
• Reconstruction and Membership Inference Attacks against Generative Models
• Forward Super-Resolution: How Can GANs Learn Hierarchical Generative Models for Real-World Distributions
• Limiting Behaviors of Nonconvex-Nonconcave Minimax Optimization via Continuous-Time Systems
• Reciprocal Adversarial Learning via Characteristic Functions
• Lessons Learned from the Training of GANs on Artificial Datasets
• Convergence and Sample Complexity of SGD in GANs
• Training Wasserstein GANs without gradient penalties
• Outlier-Robust Optimal Transport: Duality, Structure, and Statistical Analysis
• On Predicting Generalization using GANs
• Tractable Density Estimation on Learned Manifolds with Conformal Embedding Flows
• Towards A Multi-agent System for Online Hate Speech Detection
• DO-GAN: A Double Oracle Framework for Generative Adversarial Networks
• GANs with Conditional Independence Graphs: On Subadditivity of Probability Divergences
• Alleviation of Gradient Exploding in GANs: Fake Can Be Real
• Collaborative Sampling in Generative Adversarial Networks
• Risk Bounds for Unsupervised Cross-Domain Mapping with IPMs
• Channel-Recurrent Autoencoding for Image Modeling
• ChainGAN: A sequential approach to GANs
• Domain Partitioning Network
• MaskAAE: Latent space optimization for Adversarial Auto-Encoders
• Online Kernel based Generative Adversarial Networks
• On the Performance Analysis of the Adversarial System Variant Approximation Method to Quantify Process Model Generalization
• Tessellated Wasserstein Auto-Encoders
• Making Method of Moments Great Again? -- How can GANs learn distributions
• Unified cross-modality feature disentangler for unsupervised multi-domain MRI abdomen organs segmentation
• On the Anomalous Generalization of GANs
• MMCGAN: Generative Adversarial Network with Explicit Manifold Prior
• Towards GANs' Approximation Ability
• Toward a Generalization Metric for Deep Generative Models
• Statistical Mechanical Analysis of Neural Network Pruning
• Improving the expressiveness of neural vocoding with non-affine Normalizing Flows
• Improving Generative Adversarial Networks with Local Coordinate Coding
• Neural Estimation of Statistical Divergences
• Separation Results between Fixed-Kernel and Feature-Learning Probability Metrics
• A Wasserstein Minimax Framework for Mixed Linear Regression
• Lifelong Mixture of Variational Autoencoders
• Continual Learning of Generative Models with Limited Data: From Wasserstein-1 Barycenter to Adaptive Coalescence
• Host-Pathongen Co-evolution Inspired Algorithm Enables Robust GAN Training