Sequential Monte Carlo for Inclusive KL Minimization in Amortized Variational Inference

arXiv:2403.10610v1 Announce Type: new
Abstract: For training an encoder network to perform amortized variational inference, the Kullback-Leibler (KL) divergence from the exact posterior to its approximation, known as the inclusive or forward KL, is an increasingly popular choice of variational objective due to the mass-covering property of its minimizer. However, minimizing this objective is challenging. A popular existing approach, Reweighted Wake-Sleep (RWS), suffers from heavily biased gradients and a circular pathology that results in highly concentrated variational distributions. As an …

abstract approximation arxiv cs.lg divergence encoder inference network popular posterior property stat.ml training type