Black box variational inference (BBVI) with reparameterization gradients triggered the exploration of divergence measures other than the Kullback-Leibler (KL) divergence, such as alpha divergences. These divergences can be tuned to be more mass-covering (preventing overfitting in complex models), but are also often harder to optimize using Monte-Carlo gradients. In this paper, we view BBVI with generalized divergences as a form of biased importance sampling. The choice of divergence determines a bias-variance tradeoff between the tightness of the bound (low bias) and the variance of its gradient estimators. Drawing on variational perturbation theory of statistical physics, we use these insights to construct a new variational bound which is tighter than the KL bound and more mass covering. Compared to alpha-divergences, its reparameterization gradients have a lower variance. We show in several experiments on Gaussian Processes and Variational Autoencoders that the resulting posterior covariances are closer to the true posterior and lead to higher likelihoods on held-out data.

本文针对黑盒变分推断方法提出了基于广义差异度量的偏置重要性采样方法，并通过变分扰动理论构建了一族新的变分上界，对多个模型进行实验得出新的变分上界在数据拟合上更优

摄动黑盒变分推理