Bias Correction of Learned Generative Models using Likelihood-Free Importance Weighting

Abstract

A learned generative model often produces biased statistics relative to the underlying data distribution. A standard technique to correct this bias is importance sampling, where samples from the model are weighted by the likelihood ratio under model and true distributions. When the likelihood ratio is unknown, it can be estimated by training a probabilistic classifier to distinguish samples from the two distributions. We employ this likelihood-free importance weighting method to correct for the bias in generative models. We find that this technique consistently improves standard goodness-of-fit metrics for evaluating the sample quality of state-of-the-art deep generative models, suggesting reduced bias. Finally, we demonstrate its utility on representative applications in a) data augmentation for classification using generative adversarial networks, and b) model-based policy evaluation using off-policy data.

NeurIPS 2019

References in corpus (15)

• Auto-Encoding Variational Bayes
• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium
• Spectral Normalization for Generative Adversarial Networks
• Stochastic Backpropagation and Approximate Inference in Deep Generative Models
• On Calibration of Modern Neural Networks
• NICE: Non-linear Independent Components Estimation
• Data Augmentation Generative Adversarial Networks
• Markov Chain Monte Carlo and Variational Inference: Bridging the Gap
• Variational Approaches for Auto-Encoding Generative Adversarial Networks
• Learning in Implicit Generative Models
• Continuous State-Space Models for Optimal Sepsis Treatment - a Deep Reinforcement Learning Approach
• Bias and Generalization in Deep Generative Models: An Empirical Study
• Comparison of Maximum Likelihood and GAN-based training of Real NVPs
• Towards GAN Benchmarks Which Require Generalization
• Boosted Generative Models

Cited by in corpus (28)

• NVAE: A Deep Hierarchical Variational Autoencoder
• Pretrained Transformers as Universal Computation Engines
• Your GAN is Secretly an Energy-based Model and You Should use Discriminator Driven Latent Sampling
• DermGAN: Synthetic Generation of Clinical Skin Images with Pathology
• Energy-Based Reranking: Improving Neural Machine Translation Using Energy-Based Models
• Subsampling Generative Adversarial Networks: Density Ratio Estimation in Feature Space with Softplus Loss
• Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation
• Residual Energy-Based Models for Text
• Autofocused oracles for model-based design
• Residual Energy-Based Models for Text Generation
• Fair Generative Modeling via Weak Supervision
• Self-Diagnosing GAN: Diagnosing Underrepresented Samples in Generative Adversarial Networks
• Training Deep Energy-Based Models with f-Divergence Minimization
• Improving GAN Training with Probability Ratio Clipping and Sample Reweighting
• Discriminator Contrastive Divergence: Semi-Amortized Generative Modeling by Exploring Energy of the Discriminator
• A Contrastive Learning Approach for Training Variational Autoencoder Priors
• Telescoping Density-Ratio Estimation
• Generalized Energy Based Models
• Measuring Fairness in Generative Models
• Unsupervised Calibration under Covariate Shift
• Inclusive GAN: Improving Data and Minority Coverage in Generative Models
• Refining Deep Generative Models via Discriminator Gradient Flow
• Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence
• Enhancing Stability in Training Conditional Generative Adversarial Networks via Selective Data Matching
• Disentangling Fine-Tuning from Pre-Training in Visual Captioning with Hybrid Markov Logic
• Adaptation and Generalization for Unknown Sensitive Factors of Variations
• Disentangled Unsupervised Image Translation via Restricted Information Flow
• Bridging the Gap Between $f$-GANs and Wasserstein GANs