Stein Variational Inference: Approximate Learning and Inference with Stein's MethodMachine learning and statistics are essentially about matching observed data and mathematical models (typically probabilistic models). Hence, mathematical tools that measure and optimize notions of discrepancies between data and distributions play a central role in all aspects of statistical learning, including estimating parameters and models from empirical data, simulating and reasoning with given models (e.g., Bayesian posterior inference), and evaluating the goodnessoffit between data and models (evaluation); see Figure 1 left. However, as the scale and complexity of the probabilistic models that empower modern AI/ML/statistical problems grow rapidly and dramatically, the classical discrepancy tools, many of which center around Kullback Leibler (KL) divergence, become computationally infeasible. Stein variational inference, consisting of kernelized Stein discrepancy (KSD) and Stein variational gradient descent (SVGD), is a collection of computational devices for discrepancy evaluation and optimization for intractable models. It draws key insights from Stein's method due to Charles M. Stein, which is a set of remarkably powerful theoretical techniques originally developed for proving approximation and limit theorems in probability theory. Stein variational methods turn Stein’s method into practical computational tools that can be used to handle complex data and models in computational statistics and machine learning. Figure 1. Left: Probabilistic learning, inference, evaluation tasks can be viewed as evaluating or minimizing discrepancies between data and distributions. Kernelized Stein discrepancy (KSD) provides a unified tool useful for all the three tasks on unnormalized distributions. Right: Key theoretical concepts in Stein variational inference (lower panel), which can be viewed as a "kernelized'' counterpart of the classical Wasserstein gradient flow theory of Langevin Monte Carlo (upper panel). Kernelized Stein DiscrepancyKernelized Stein discrepancy (KSD), based on combining the classical Stein discrepancy with reproducing kernel Hilbert space (RKHS), allows us to access the compatibility between empirical data and probabilistic distributions, and provides a powerful tool for developing algorithms for model evaluation (goodnessoffit test), as well as learning and inference in general. Unlike the traditional divergence measures (such as KL, Chisquare divergence), KSD does not require to evaluate the normalization constant of the distribution, and can be applied even for the intractable, unnormalized distributions widely used in modern machine learning and statistics.
Stein Variational Gradient DescentStein variational gradient descent (SVGD) is a nonparametric variational inference algorithm that mixes the advantages of variational inference, Monte Carlo, quasi Monte Carl and gradient based optimization, based on exploiting an interesting connection between Stein discrepancy and KL divergence. It is a deterministic sampling algorithm that evolves a set of interacting particles to form sample efficient particle approximation to the given distribution. It is both 1) a gradient flow of KL divergence w.r.t. a ‘‘kernelized’’ Wasserstein metric on the space of distributions, and 2) a numerical quadrature method which arranges a set of points to match the expectation of a set of basis functions induced from Stein operator.
Slides/NotesProbabilistic Learning and Inference Using Stein's Method [slides, slides] A Kernelized Stein Discrepancy for Goodnessoffit Tests and Model Evaluation [ICML2016 slides] Short notes: Papers
