PacGAN: The Power of Two Samples in Generative Adversarial Networks (2017-12-12T00:00:00.000000Z)

TL;DR

It is shown that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process, and numerical experiments suggests that packing provides significant improvements in practice as well.

Abstract

Generative adversarial networks (GANs) are innovative techniques for learning generative models of complex data distributions from samples. Despite remarkable 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 mitigate mode collapse. We propose a principled approach to handle mode collapse called 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, 1953)—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.

Authors

Sewoong Oh

3 papers

Zinan Lin

3 papers

G. Fanti

2 papers

TL;DR

Abstract

Authors

References77 items

The Inductive Bias of Restricted f-GANs

Glow: Generative Flow with Invertible 1x1 Convolutions

On GANs and GMMs

Attention-based Graph Neural Network for Semi-supervised Learning

Spectral Normalization for Generative Adversarial Networks

AmbientGAN: Generative models from lossy measurements

Demystifying MMD GANs

The Robust Manifold Defense: Adversarial Training using Generative Models

Theoretical limitations of Encoder-Decoder GAN architectures

A Classification-Based Perspective on GAN Distributions

Understanding GANs: the LQG Setting

Progressive Growing of GANs for Improved Quality, Stability, and Variation

Phase space sampling and operator confidence with generative adversarial networks

Dual Discriminator Generative Adversarial Nets

f-GANs in an Information Geometric Nutshell

Towards Understanding the Dynamics of Generative Adversarial Networks

Do GANs actually learn the distribution? An empirical study

GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

Dualing GANs

Gradient descent GAN optimization is locally stable

The Numerics of GANs

Stabilizing Training of Generative Adversarial Networks through Regularization

MMD GAN: Towards Deeper Understanding of Moment Matching Network

VEEGAN: Reducing Mode Collapse in GANs using Implicit Variational Learning

Approximation and Convergence Properties of Generative Adversarial Learning

Improved Training of Wasserstein GANs

Deep Sets

Compressed Sensing using Generative Models

Generalization and Equilibrium in Generative Adversarial Nets (GANs)

Wasserstein GAN

AdaGAN: Boosting Generative Models

NIPS 2016 Tutorial: Generative Adversarial Networks

Image-to-Image Translation with Conditional Adversarial Networks

A Connection between Generative Adversarial Networks, Inverse Reinforcement Learning, and Energy-Based Models

Mode Regularized Generative Adversarial Networks

On the Quantitative Analysis of Decoder-Based Generative Models

Unrolled Generative Adversarial Networks

Generative Models and Model Criticism via Optimized Maximum Mean Discrepancy

SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

Semi-Supervised Classification with Graph Convolutional Networks

Generating Videos with Scene Dynamics

Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering

Improved Techniques for Training GANs

Adversarially Learned Inference

f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization

Adversarial Feature Learning

Generative Adversarial Text to Image Synthesis

Secure Multi-party Differential Privacy

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

A note on the evaluation of generative models

Bridges: a uniquely flexible HPC resource for new communities and data analytics

Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Adam: A Method for Stochastic Optimization

Deep Learning Face Attributes in the Wild

XSEDE: Accelerating Scientific Discovery

Extremal Mechanisms for Local Differential Privacy

Auto-Encoding Variational Bayes

The Composition Theorem for Differential Privacy

Distributed Representations of Words and Phrases and their Compositionality

Hilbert Space Embeddings and Metrics on Probability Measures

Children’s understanding of the relationship between addition and subtraction

An Extremal Inequality Motivated by Multiterminal Information-Theoretic Problems

A simple proof of the entropy-power inequality

An Extremal Inequality Motivated by Multiterminal Information Theoretic Problems

A Proof of the Fisher Information Inequality via a Data Processing Argument

Information theoretic inequalities

Determinant inequalities via information theory

Some Inequalities Satisfied by the Quantities of Information of Fisher and Shannon

Equivalent Comparisons of Experiments

GENERATIVE ADVERSARIAL NETS

Bayesian GANs

A Practical Guide to Training Restricted Boltzmann Machines

Learning Multiple Layers of Features from Tiny Images

The mnist database of handwritten digits