English

Variational Inference via $\chi$-Upper Bound Minimization

Machine Learning 2017-11-15 v4 Machine Learning Computation Methodology

Abstract

Variational inference (VI) is widely used as an efficient alternative to Markov chain Monte Carlo. It posits a family of approximating distributions qq and finds the closest member to the exact posterior pp. Closeness is usually measured via a divergence D(qp)D(q || p) from qq to pp. While successful, this approach also has problems. Notably, it typically leads to underestimation of the posterior variance. In this paper we propose CHIVI, a black-box variational inference algorithm that minimizes Dχ(pq)D_{\chi}(p || q), the χ\chi-divergence from pp to qq. CHIVI minimizes an upper bound of the model evidence, which we term the χ\chi upper bound (CUBO). Minimizing the CUBO leads to improved posterior uncertainty, and it can also be used with the classical VI lower bound (ELBO) to provide a sandwich estimate of the model evidence. We study CHIVI on three models: probit regression, Gaussian process classification, and a Cox process model of basketball plays. When compared to expectation propagation and classical VI, CHIVI produces better error rates and more accurate estimates of posterior variance.

Keywords

Cite

@article{arxiv.1611.00328,
  title  = {Variational Inference via $\chi$-Upper Bound Minimization},
  author = {Adji B. Dieng and Dustin Tran and Rajesh Ranganath and John Paisley and David M. Blei},
  journal= {arXiv preprint arXiv:1611.00328},
  year   = {2017}
}

Comments

Neural Information Processing Systems, 2017

R2 v1 2026-06-22T16:38:59.027Z