Understanding Mutual Information and its Use in InfoGAN

Understanding Mutual Information and its Use in InfoGAN
复制标题

了解互信息及其在 InfoGAN 中的使用

DOI:
--
复制
发表时间:
2016
期刊:
--
影响因子:
--
通讯作者:
Andrew Drozdov
Andrew Drozdov
中科院分区:
--
文献类型:
--
作者:
Katrina Evtimova;Andrew Drozdov

文献摘要

参考文献

被引文献

相似文献

可解释的变量在生成模型中很有用。生成对抗网络(GAN)是其输入中灵活的生成模型。最大化GAN(INFOGAN)的信息将发电机的输出与其输入的组件相关联,称为潜在代码。通过强制将输出绑定到此输入组件,我们可以控制输出表示的某些属性。众所周知,当在gan中共同训练鉴别器和发电机时,很难找到NASH平衡。我们发现了使用Infogan生成图像的一些成功且失败的配置。
Interpretable variables are useful in generative models. Generative Adversarial Networks (GANs) are generative models that are flexible in their input. The Information Maximizing GAN (InfoGAN) ties the output of the generator to a component of its input called the latent codes. By forcing the output to be tied to this input component, we can control some properties of the output representation. It is notoriously difficult to find the Nash equilibrium when jointly training the discriminator and generator in a GAN. We uncover some successful and unsuccessful configurations for generating images using InfoGAN.
Exflib 信息
DOI: --
发表时间: --
期刊:
影响因子: --
作者:
通讯作者: --