Characterizing and Avoiding Problematic Global Optima of Variational Autoencoders

Abstract

Yacoby Pan Doshi-Velez φ such that g φ (x) = η(x); we denote the variational distributions g φ (x) by q φ (z|x). Thus, maximization of the ELBO can be expressed :Posterior Matching (PM) Objective

Failure Modes of Variational Autoencoders and Their Effects on Downstream Tasks

“…In contrast, there has been little work to characterize pathologies at the global optima of the MFG-VAEs training objective. [52] shows that, when the decoder's capacity is restricted, posterior collapse and the mismatch between aggregated posterior and prior can occur as global optima of the training objective. In contrast to existing work, we focus on global optima of the MFG-VAE objective in fully general settings: with fully flexible generative and inference models, as well as with and without learned observation noise.…”

“…[33,47]). Recent work [52] attributes a number of these pathologies to properties of the training objective; in particular, the objective may compromise learning a good generative model in order to learn a good inference model -in other words, the inference model over-regularizes the generative model. While this pathology has been noted in literature [4,53,6], no prior work has characterizes the conditions under which the MFG-VAE objective compromises learning a good generative model in order to learn a good inference model; more worrisomely, no prior work has related MFG-VAE pathologies with the performance of MFG-VAEs on downstream tasks.…”