On gradient regularizers for MMD GANs

Abstract

We propose a principled method for gradient-based regularization of the critic of GAN-like models trained by adversarially optimizing the kernel of a Maximum Mean Discrepancy (MMD). We show that controlling the gradient of the critic is vital to having a sensible loss function, and devise a method to enforce exact, analytical gradient constraints at no additional cost compared to existing approximate techniques based on additive regularizers. The new loss function is provably continuous, and experiments show that it stabilizes and accelerates training, giving image generation models that outperform state-of-the art methods on $160\times 160$ CelebA and $64\times 64$ unconditional ImageNet.

arXiv commentCode at https://github.com/MichaelArbel/Scaled-MMD-GAN

References in corpus (15)

• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium
• Spectral Normalization for Generative Adversarial Networks
• Improved Training of Wasserstein GANs
• Which Training Methods for GANs do actually Converge?
• BEGAN: Boundary Equilibrium Generative Adversarial Networks
• MMD GAN: Towards Deeper Understanding of Moment Matching Network
• Generative Moment Matching Networks
• The Cramer Distance as a Solution to Biased Wasserstein Gradients
• An empirical study on evaluation metrics of generative adversarial networks
• A Note on the Inception Score
• Stabilizing Training of Generative Adversarial Networks through Regularization
• Training generative neural networks via Maximum Mean Discrepancy optimization
• Generative Models and Model Criticism via Optimized Maximum Mean Discrepancy
• Towards the Automatic Anime Characters Creation with Generative Adversarial Networks
• A Test of Relative Similarity For Model Selection in Generative Models

Cited by in corpus (21)

• Deep Knockoffs
• COT-GAN: Generating Sequential Data via Causal Optimal Transport
• How Well Generative Adversarial Networks Learn Distributions
• ICE-BeeM: Identifiable Conditional Energy-Based Deep Models Based on Nonlinear ICA
• A case for new neural network smoothness constraints
• Sliced Kernelized Stein Discrepancy
• Learning deep kernels for exponential family densities
• Two Sample Testing in High Dimension via Maximum Mean Discrepancy
• Estimating Barycenters of Measures in High Dimensions
• Interaction Matters: A Note on Non-asymptotic Local Convergence of Generative Adversarial Networks
• Two-sample Testing Using Deep Learning
• Blindness of score-based methods to isolated components and mixing proportions
• Generalized Energy Based Models
• Kernel-Guided Training of Implicit Generative Models with Stability Guarantees
• Learning Manifold Implicitly via Explicit Heat-Kernel Learning
• Double cycle-consistent generative adversarial network for unsupervised conditional generation
• KernelNet: A Data-Dependent Kernel Parameterization for Deep Generative Modeling
• Amortised Learning by Wake-Sleep
• Adversarially Robust Kernel Smoothing
• Towards Accurate Knowledge Transfer via Target-awareness Representation Disentanglement
• The Geometric Occam's Razor Implicit in Deep Learning