生成对抗网络[8]最近被引入作为一种训练生成模型的新方法。在这项工作中，我们引入了条件版本的生成对抗网络，它可以通过简单地将我们希望条件化的数据 y 输入到生成器和判别器中来构建。我们展示了该模型可以生成以类别标签为条件的 MNIST 数字。我们还说明了该模型如何用于学习多模态模型，并提供了图像标注应用的初步示例，其中我们展示了该方法如何生成不属于训练标签的描述性标签。
Generative Adversarial Nets [8] were recently introduced as a novel way to train generative models. In this work we introduce the conditional version of generative adversarial nets, which can be constructed by simply feeding the data, y, we wish to condition on to both the generator and discriminator. We show that this model can generate MNIST digits conditioned on class labels. We also illustrate how this model could be used to learn a multi-modal model, and provide preliminary examples of an application to image tagging in which we demonstrate how this approach can generate descriptive tags which are not part of training labels.
核心贡献 · Key contributions
提出条件生成对抗网络（cGAN），将条件数据同时输入生成器和判别器。 Introduces conditional generative adversarial nets (cGANs) by feeding conditioning data to both generator and discriminator.
展示基于类别标签的 MNIST 数字生成。 Demonstrates MNIST digit generation conditioned on class labels.
在 MIR Flickr 数据集上展示图像标注的多模态学习能力。 Shows multi-modal learning capability via image tagging on MIR Flickr dataset.
利用词嵌入生成训练标签中未出现的描述性标签。 Generates descriptive tags not present in training labels using word embeddings.
为概率生成模型中的一对多映射提供框架。 Provides a framework for one-to-many mapping in probabilistic generative models.
局限 · Limitations
初步结果：条件模型在 MNIST 上表现不如非条件对抗网络。 Preliminary results; conditional model underperforms non-conditional adversarial nets on MNIST.
图像标注实验使用固定的卷积和语言模型，未进行联合训练。 Image tagging experiments use fixed convolutional and language models without joint training.
每张图像仅使用单个标签，多标签集合生成留待未来工作。 Only single tags used per image; multi-tag set generation left for future work.
超参数搜索有限，架构未完全优化。 Hyper-parameter search limited; architecture not fully optimized.
论文章节 · Sections（共 8）
摘要Abstract
1 引言1 Introduction
2.1 图像标注的多模态学习2.1 Multi-modal Learning For Image Labelling