https://arxiv.org/abs/1804.02086

Abstract

Deep latent-variable models learn representa-

tions of high-dimensional data in an unsuper-

vised manner. A number of recent efforts have

focused on learning representations that disen-

tangle statistically independent axes of varia-

tion, often by introducing suitable modifica-

tions of the objective function. We synthesize

this growing body of literature by formulating

a generalization of the evidence lower bound

that explicitly represents the trade-offs between

sparsity of the latent code, bijectivity of repre-

sentations, and coverage of the support of the

empirical data distribution. Our objective is

also suitable to learning hierarchical representa-

tions that disentangle blocks of variables whilst

allowing for some degree of correlations within

blocks. Experiments on a range of datasets

demonstrate that learned representations con-

tain interpretable features, are able to learn dis-

crete attributes, and generalize to unseen com-

binations of factors.

各种vae的比较！