Flow-GAN: Combining Maximum Likelihood and Adversarial Learning in Generative Models

Abstract

Adversarial learning of probabilistic models has recently emerged as a promising alternative to maximum likelihood. Implicit models such as generative adversarial networks (GAN) often generate better samples compared to explicit models trained by maximum likelihood. Yet, GANs sidestep the characterization of an explicit density which makes quantitative evaluations challenging. To bridge this gap, we propose Flow-GANs, a generative adversarial network for which we can perform exact likelihood evaluation, thus supporting both adversarial and maximum likelihood training. When trained adversarially, Flow-GANs generate high-quality samples but attain extremely poor log-likelihood scores, inferior even to a mixture model memorizing the training data; the opposite is true when trained by maximum likelihood. Results on MNIST and CIFAR-10 demonstrate that hybrid training can attain high held-out likelihoods while retaining visual fidelity in the generated samples.

AAAI 2018

Cited by in corpus (42)

• An Introduction to Variational Autoencoders
• Normalizing Flows: An Introduction and Review of Current Methods
• Normalizing Flows for Probabilistic Modeling and Inference
• Improved Precision and Recall Metric for Assessing Generative Models
• Guided Image Generation with Conditional Invertible Neural Networks
• Glow: Generative Flow with Invertible 1x1 Convolutions
• Constructing Unrestricted Adversarial Examples with Generative Models
• Roundtrip: A Deep Generative Neural Density Estimator
• Deep Generative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models
• PointFlow: 3D Point Cloud Generation with Continuous Normalizing Flows
• Prescribed Generative Adversarial Networks
• CaSPR: Learning Canonical Spatiotemporal Point Cloud Representations
• Approximability of Discriminators Implies Diversity in GANs
• The Information Autoencoding Family: A Lagrangian Perspective on Latent Variable Generative Models
• All You Need is a Good Functional Prior for Bayesian Deep Learning
• Understanding and Mitigating Exploding Inverses in Invertible Neural Networks
• Fair Generative Modeling via Weak Supervision
• Robust Variational Autoencoder
• Tail of Distribution GAN (TailGAN): Generative-Adversarial-Network-Based Boundary Formation
• Understanding the Limitations of Conditional Generative Models
• Globally Injective ReLU Networks
• Coupling the reduced-order model and the generative model for an importance sampling estimator
• Neural Density Estimation and Likelihood-free Inference
• Training Generative Reversible Networks
• Approaching the Limit of Image Rescaling via Flow Guidance
• Sliced Iterative Normalizing Flows
• VAE-KRnet and its applications to variational Bayes
• Invertible Zero-Shot Recognition Flows
• Tractable Density Estimation on Learned Manifolds with Conformal Embedding Flows
• Variational Dynamic Mixtures
• Semi-Implicit Generative Model
• A Forest from the Trees: Generation through Neighborhoods
• FIS-GAN: GAN with Flow-based Importance Sampling
• Leveraging GPT-2 for Classifying Spam Reviews with Limited Labeled Data via Adversarial Training
• Autoencoder-based Representation Learning from Heterogeneous Multivariate Time Series Data of Mechatronic Systems
• Augmented KRnet for density estimation and approximation
• FoCL: Feature-Oriented Continual Learning for Generative Models
• Trust the Critics: Generatorless and Multipurpose WGANs with Initial Convergence Guarantees
• Variational Hetero-Encoder Randomized GANs for Joint Image-Text Modeling
• EBMs Trained with Maximum Likelihood are Generator Models Trained with a Self-adverserial Loss
• Testing for Typicality with Respect to an Ensemble of Learned Distributions
• Probabilistic Autoencoder using Fisher Information