论文链接：https://arxiv.org/pdf/1804.02086.pdf
Deep latent-variable models learn representations of high-dimensional data in an
unsupervised manner.（无监督方式）
disentangle statistically independent axes of variation （变量独立统计的核心）by introducing
modifications to the standard objective function（修正标准目标方程解耦概率统计独立的变量？）
assume a simple diagonal Gaussian prior（假设高斯先验）
not able to reliably disentangle discrete factors of variation(不能解耦离散变量的特征)
two-level hierarchical objective
control relative degree of statistical independence
between blocks of variables and individual variables within blocks
explicitly (明确的)represent the trade-offs
between mutual information（数据和表示间的互信息） between data and representation, KL divergence between representation and prior（先验和表示之间的散度）, and coverage of the support of the empirical data distribution.（经验数据分布支持的覆盖）
The relationship between x and z is described by a conditional probability distribution
pθ(x|z) parameterized by a deep neural network. 使用神经网络表示likelihood
where qφ(z|x) and pθ(x|z) represent probabilistic encoders and decoders respectively.（q和p分别代表表示编码和解码的概率）
deep generative models often provide high-fidelity reconstructions（深度生成模型经常提供高保真的重建）
In contrast to classical methods such as principal components or factor analysis（主成分或因子分析）, individual dimensions of z（z的维度） don’t necessarily encode any particular semantically meaningful variation in x.（对x不起作用）
perturbations of an individual dimension of the latent code z （隐编码维度的扰动）perturb the corresponding x in an interpretable manner.（影响对应x的表示方式）
unsupervised approaches （无监督的方式）that modify the VAE objective （修正目标VAE）to induce disentangled representations（减少解耦表示）——>β-VAE
modify the VAE objective by adding, removing, or altering the weight of individual terms（通过增加、移除、变更独立目标的权重）
based on a two-level hierarchical decomposition （两级封层分解）of the VAE objective
control the relative levels of statistical independence（统计独立性） between groups of variables and for individual variables in the same group.（组变量和同组内的个体变量）
we induce statistical independence（引入统计独立性） by minimizing the total correlation (TC)（最小化总的联系）, a generalization of the mutual information to more than two variables.（超过两个变量生成的互信息）
we reinterpret the standard VAE objective as a KL divergence between a generative model and its corresponding inference model.（在生成模型和它相关的推理模型间）
Inspection of the learned representations （对已经学习的表示的调查）confirms that our objective uncovers interpretable features（发现解释特征） in an unsupervised setting, and quantitative metrics （非监督设定和定量指标）demonstrate improvement over related methods.
an implicit goal （含蓄目标？）in learning disentangled representations（解耦表示） that is now considered explicitly（明确的）