每日论文InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial
2017-03-13 16:09
531 查看
转载地址:http://hacker.duanshishi.com/?p=1766
InfoGAN
InfoGAN是一种能够学习disentangled representation的GAN(https://arxiv.org/pdf/1606.03657v1.pdf),何为disentangled representation?比如人脸数据集中有各种不同的属性特点,如脸部表情、是否带眼睛、头发的风格眼珠的颜色等等,这些很明显的相关表示,InfoGAN能够在完全无监督信息(是否带眼睛等等)下能够学习出这些disentangled representation,而相对于传统的GAN,只需修改loss来最大化GAN的input的noise(部分fixed的子集)和最终输出之间的互信息。
原理
为了达到上面提到的效果,InfoGAN必须在input的noise来做一些文章,将noise vector划分为两部分:z: 和原始的GAN input作用一致;
c: latent code,能够在之后表示数据分布中的disentangled representation
那么如何从latent code中学到相应的disentangled representation呢? 在原始的GAN中,忽略了c这部分的影响,即GAN产生的数据分布满足$P_{G}(x|C)=P(x)$,为了保证能够利用c这部分信息, 作者提出这样一个假设:c与generator的输出相关程度应该很大,而在信息论中,两个数据分布的相关程度即互信息, 即generator的输出和input的c的$I(c;G(z,c))$应该会大。 所以,InfoGAN就变成如下的优化问题:
因为互信息的计算需要后验概率的分布(下图红线部分),在实际中很难直接使用,因此,在实际训练中一般不会直接最大化$I(c;G(z,c))$
这里作者采用和VAE类似的方法,增加一个辅助的数据分布为后验概率的low
bound: 所以,这里互信息的计算如下:
这里相关的证明就不深入了,有兴趣的可以去看看paper。
相关文章推荐
- 阅读小结:InfoGAN:Interpretable Representation Learning by Information Maximising Generative Adversarial
- 论文笔记:unsupervised representation learning with deep convolutional generative adversarial networks
- 生成对抗网络学习笔记3----论文unsupervised representation learning with deep convolutional generative adversarial
- 论文笔记之:UNSUPERVISED REPRESENTATION LEARNING WITH DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL NETWORKS
- 每日论文Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
- 【论文阅读】Author2Vec: Learning Author Representations by Combining Content and Link Information
- 论文笔记之:Semi-Supervised Learning with Generative Adversarial Networks
- 每日论文 Learning from Simulated and Unsupervised Images through Adversarial Training
- Re-ID: Person Re-identification by Local Maximal Occurrence Representation and Metric Learning 论文解析
- 每日论文Least Squares Generative Adversarial Networks
- 【Paper Note】Generative Adversarial Nets论文翻译(未完待续)
- DeepID2:Deep Learning Face Representation by Joint Identification-Verification
- Deep Learning Face Representation by Joint Identification-Verification
- Deep Learning Face Representation by Joint Identification-Verification
- 论文笔记:Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks
- 论文笔记(一)Deep Ranking for Person Re-Identification via Joint Representation Learning
- 论文笔记(一)Deep Ranking for Person Re-Identification via Joint Representation Learning (续)
- 文章1-Deep Learning Face Representation by Joint Identification-Verification
- 论文笔记之:Conditional Generative Adversarial Nets
- Deep Learning Face Representation by JOint Identification- Verification(DeepID2)