| 摘要: |
| 产生式对抗网络（generative adversarial networks，简称GANs）可以生成逼真的图像，因此最近被广泛研究.值得注意的是，概率图生成对抗网络（graphical-GAN）将贝叶斯网络引入产生式对抗网络框架，以无监督的方式学习到数据的隐藏结构.提出了条件概率图生成对抗网络（conditional graphical-GAN），它可以在弱监督环境下，利用粗粒度监督信息来学习到更精细而复杂的结构.条件概率图生成对抗网络的推理和学习遵循与graphical-GAN类似的方法.提出了条件概率图生成对抗网络的两个实例.条件高斯混合模型（conditional Gaussian mixture GAN，简称cGMGAN）可以在给出粗粒度标签的情况下从混合数据中学习细粒度聚类.条件状态空间模型（conditional state space GAN，简称cSSGAN）可以在给定对象标签的情况下学习具有多个对象的视频的动态过程. |
| 关键词: 深度生成模型 产生式对抗网络 概率图模型 弱监督学习 条件模型 |
| DOI：10.13328/j.cnki.jos.005924 |
| 分类号:TP181 |
| 基金项目:国家自然科学基金（61620106010，61621136008）；博士后创新人才拔尖计划 |
|
| Conditional Graphical Generative Adversarial Networks |
|
LI Chong-Xuan, ZHU Jun, ZHANG Bo
|
|
Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China
|
| Abstract: |
| Generative adversarial networks (GANs) have been promise on generating realistic images and hence have been studied widely. Notably, graphical generative adversarial networks (graphical-GAN) introduce Bayesian networks to the GAN framework to learn the underlying structures of data in an unsupervised manner. This study proposes a conditional version of graphical-GAN, which can leverage coarse side information to enhance the graphical-GAN and learn finer and more complex structures, in weakly-supervised learning settings. The inference and learning of conditional graphical-GAN follows a similar protocol to graphical-GAN. Two instances of conditional graphical-GAN are presented. The conditional Gaussian mixture GAN can learn fine clusters from mixture data given a coarse label. The conditional state space GAN can learn the dynamics of videos with multiple objects given the labels of the objects.. |
| Key words: deep generative model generative adversarial network graphical model weakly-supervised learning conditional model |
