Do GANs actually learn the distribution? An empirical study

Abstract

Do GANS (Generative Adversarial Nets) actually learn the target distribution? The foundational paper of (Goodfellow et al 2014) suggested they do, if they were given sufficiently large deep nets, sample size, and computation time. A recent theoretical analysis in Arora et al (to appear at ICML 2017) raised doubts whether the same holds when discriminator has finite size. It showed that the training objective can approach its optimum value even if the generated distribution has very low support ---in other words, the training objective is unable to prevent mode collapse. The current note reports experiments suggesting that such problems are not merely theoretical. It presents empirical evidence that well-known GANs approaches do learn distributions of fairly low support, and thus presumably are not learning the target distribution. The main technical contribution is a new proposed test, based upon the famous birthday paradox, for estimating the support size of the generated distribution.

References in corpus (7)

• Auto-Encoding Variational Bayes
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop
• Adversarial Feature Learning
• Adversarially Learned Inference
• Improved generator objectives for GANs
• Stacked Generative Adversarial Networks

Cited by in corpus (55)

• Progressive Growing of GANs for Improved Quality, Stability, and Variation
• Generating Natural Adversarial Examples
• Improved Precision and Recall Metric for Assessing Generative Models
• Using GANs for Sharing Networked Time Series Data: Challenges, Initial Promise, and Open Questions
• Demystifying MMD GANs
• Ten Years of Generative Adversarial Nets (GANs): A survey of the state-of-the-art
• Classification Accuracy Score for Conditional Generative Models
• Simulation of electron-proton scattering events by a Feature-Augmented and Transformed Generative Adversarial Network (FAT-GAN)
• Implicit Maximum Likelihood Estimation
• Is Generator Conditioning Causally Related to GAN Performance?
• A Convex Duality Framework for GANs
• How Well Generative Adversarial Networks Learn Distributions
• Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields
• Mode Collapse and Regularity of Optimal Transportation Maps
• A unified framework for 21cm tomography sample generation and parameter inference with Progressively Growing GANs
• A black box for dark sector physics: Predicting dark matter annihilation feedback with conditional GANs
• GANGs: Generative Adversarial Network Games
• Imperfect ImaGANation: Implications of GANs Exacerbating Biases on Facial Data Augmentation and Snapchat Selfie Lenses
• Theoretical limitations of Encoder-Decoder GAN architectures
• Learning stochastic object models from medical imaging measurements by use of advanced ambient generative adversarial networks
• A New Semi-supervised Learning Benchmark for Classifying View and Diagnosing Aortic Stenosis from Echocardiograms
• Effect of The Latent Structure on Clustering with GANs
• Lipizzaner: A System That Scales Robust Generative Adversarial Network Training
• Investigating Under and Overfitting in Wasserstein Generative Adversarial Networks
• Compressed Sensing MRI Reconstruction Regularized by VAEs with Structured Image Covariance
• Understanding the Effectiveness of Lipschitz-Continuity in Generative Adversarial Nets
• Interpreting Deep Learning: The Machine Learning Rorschach Test?
• Interaction Matters: A Note on Non-asymptotic Local Convergence of Generative Adversarial Networks
• DeepMasterPrints: Generating MasterPrints for Dictionary Attacks via Latent Variable Evolution
• DADA: Deep Adversarial Data Augmentation for Extremely Low Data Regime Classification
• Train simultaneously, generalize better: Stability of gradient-based minimax learners
• Super-Resolution via Conditional Implicit Maximum Likelihood Estimation
• Reciprocal Adversarial Learning via Characteristic Functions
• Red-GAN: Attacking class imbalance via conditioned generation. Yet another perspective on medical image synthesis for skin lesion dermoscopy and brain tumor MRI
• RecurJac: An Efficient Recursive Algorithm for Bounding Jacobian Matrix of Neural Networks and Its Applications
• Orthogonal Wasserstein GANs
• Lessons Learned from the Training of GANs on Artificial Datasets
• Learning Shared Semantic Space with Correlation Alignment for Cross-modal Event Retrieval
• Learning Generative Models of Tissue Organization with Supervised GANs
• Stable Rank Normalization for Improved Generalization in Neural Networks and GANs
• Generative Models for Security: Attacks, Defenses, and Opportunities
• Capturing the diversity of biological tuning curves using generative adversarial networks
• Detecting GAN generated errors
• Conditional Generative Modeling via Learning the Latent Space
• Implicit Manifold Learning on Generative Adversarial Networks
• To Regularize or Not To Regularize? The Bias Variance Trade-off in Regularized AEs
• Graph-based Discriminators: Sample Complexity and Expressiveness
• Radio Halo Detection in MWA Data using Deep Neural Networks and Generative Data Augmentation
• Generating new pictures in complex datasets with a simple neural network
• SetGAN: Improving the stability and diversity of generative models through a permutation invariant architecture
• Bringing Balance to Hand Shape Classification: Mitigating Data Imbalance Through Generative Models
• Training Generative Networks with general Optimal Transport distances
• High Diversity Attribute Guided Face Generation with GANs
• GILBO: One Metric to Measure Them All
• The Implicit Metropolis-Hastings Algorithm