Operator Variational Inference

Abstract

Variational inference is an umbrella term for algorithms which cast Bayesian inference as optimization. Classically, variational inference uses the Kullback-Leibler divergence to define the optimization. Though this divergence has been widely used, the resultant posterior approximation can suffer from undesirable statistical properties. To address this, we reexamine variational inference from its roots as an optimization problem. We use operators, or functions of functions, to design variational objectives. As one example, we design a variational objective with a Langevin-Stein operator. We develop a black box algorithm, operator variational inference (OPVI), for optimizing any operator objective. Importantly, operators enable us to make explicit the statistical and computational tradeoffs for variational inference. We can characterize different properties of variational objectives, such as objectives that admit data subsampling---allowing inference to scale to massive data---as well as objectives that admit variational programs---a rich class of posterior approximations that does not require a tractable density. We illustrate the benefits of OPVI on a mixture model and a generative model of images.

Appears in Neural Information Processing Systems, 2016

References in corpus (6)

• Adam: A Method for Stochastic Optimization
• Auto-Encoding Variational Bayes
• Stochastic Backpropagation and Approximate Inference in Deep Generative Models
• Expectation Propagation for approximate Bayesian inference
• Auxiliary Deep Generative Models
• Automated Variational Inference in Probabilistic Programming

Cited by in corpus (32)

• Causal Effect Inference with Deep Latent-Variable Models
• Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks
• Hierarchical Implicit Models and Likelihood-Free Variational Inference
• Monte Carlo Gradient Estimation in Machine Learning
• Measuring Sample Quality with Kernels
• Generalized Variational Inference: Three arguments for deriving new Posteriors
• Gradient Estimators for Implicit Models
• A Spectral Approach to Gradient Estimation for Implicit Distributions
• Simple, Distributed, and Accelerated Probabilistic Programming
• Learning the Stein Discrepancy for Training and Evaluating Energy-Based Models without Sampling
• An introduction to variational inference in Geophysical inverse problems
• Advances in Variational Inference
• Approximate Inference with Amortised MCMC
• Adversarial Variational Optimization of Non-Differentiable Simulators
• Minimum Stein Discrepancy Estimators
• Stein Neural Sampler
• Sliced Kernelized Stein Discrepancy
• Truncated Variational Expectation Maximization
• Posterior inference unchained with EL_2O
• Two Methods For Wild Variational Inference
• Wasserstein Variational Inference
• Generate High Resolution Images With Generative Variational Autoencoder
• Kernelized Complete Conditional Stein Discrepancy
• GO Gradient for Expectation-Based Objectives
• Generalized Bayesian Filtering via Sequential Monte Carlo
• Importance Sampled Stochastic Optimization for Variational Inference
• Doubly Semi-Implicit Variational Inference
• Measure Transport with Kernel Stein Discrepancy
• Adversarial Variational Bayes Methods for Tweedie Compound Poisson Mixed Models
• End-To-End Dilated Variational Autoencoder with Bottleneck Discriminative Loss for Sound Morphing -- A Preliminary Study
• Bayesian Sparsification Methods for Deep Complex-valued Networks
• Adversarial Stein Training for Graph Energy Models