Categorical Reparameterization with Gumbel-Softmax
Categorical Reparameterization with Gumbel-Softmax
复制标题
DOI：
--
复制
发表时间：
2016-11
期刊：
影响因子：
--
通讯作者：
Eric Jang;S. Gu;Ben Poole
中科院分区：
文献类型：
--
作者：
Eric Jang;S. Gu;Ben Poole
分类变量是表示世界上离散结构的自然选择。然而，由于无法通过样本反向传播，随机神经网络很少使用分类潜在变量。在这项工作中，我们提出了一个有效的梯度估计器，用一个新的Gumbel-Softmax分布的可微样本取代了来自分类分布的不可微样本。这种分布具有可以平滑退火成分类分布的基本性质。我们表明，我们的Gumbel-Softmax估计器在结构化输出预测和具有分类潜在变量的无监督生成建模任务上优于最先进的梯度估计器，并且在半监督分类上实现了很大的加速。
Categorical variables are a natural choice for representing discrete structure in the world. However, stochastic neural networks rarely use categorical latent variables due to the inability to backpropagate through samples. In this work, we present an efficient gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample from a novel Gumbel-Softmax distribution. This distribution has the essential property that it can be smoothly annealed into a categorical distribution. We show that our Gumbel-Softmax estimator outperforms state-of-the-art gradient estimators on structured output prediction and unsupervised generative modeling tasks with categorical latent variables, and enables large speedups on semi-supervised classification.