Differentiable Annealed Importance Sampling Minimizes The Symmetrized Kullback-Leibler Divergence Between Initial and Target Distribution

TL;DR

This paper demonstrates that differentiable annealed importance sampling (DAIS) optimizes the symmetrized KL divergence between initial and target distributions, acting as a variational inference method that improves uncertainty estimation.

Contribution

It shows that DAIS minimizes the symmetrized KL divergence in the limit, framing it as a novel variational inference approach with practical benefits.

Findings

DAIS often yields more accurate uncertainty estimates than traditional VI.

Empirical evaluation on synthetic and real data supports DAIS's effectiveness.

DAIS outperforms importance weighted VI and score climbing in uncertainty quantification.

Abstract

Differentiable annealed importance sampling (DAIS), proposed by Geffner & Domke (2021) and Zhang et al. (2021), allows optimizing over the initial distribution of AIS. In this paper, we show that, in the limit of many transitions, DAIS minimizes the symmetrized Kullback-Leibler divergence between the initial and target distribution. Thus, DAIS can be seen as a form of variational inference (VI) as its initial distribution is a parametric fit to an intractable target distribution. We empirically evaluate the usefulness of the initial distribution as a variational distribution on synthetic and real-world data, observing that it often provides more accurate uncertainty estimates than VI (optimizing the reverse KL divergence), importance weighted VI, and Markovian score climbing (optimizing the forward KL divergence).