0%

(NIPS 2016) Coupled generative adversarial networks

Liu M Y, Tuzel O. Coupled generative adversarial networks[C]//Advances in neural information processing systems. 2016: 469-477.



1. Overview


In this paper, it proposed coupled GAN

  • based on existence of shared high-level representation in the domains
  • learn a joint distribution of multi-domain images
  • unsupervised
  • used in image transformation, domain adaption

1.1. Model



Generator

  • share the same high-level concept
    Discriminator
  • sharing constraint can reduce parameters

1.2. Loss Function




  • VAE
  • Attention Model
  • Moment Matching
  • Diffusion Process
  • Cross-domian Image Generation
  • GAN
    • Laplacian Pyramid
    • Conditional GAN



2. Experiments


2.1. Metric

  • ratios of agreed pixels

2.2. Digit



  • [digit-edge], [digit-negative]
  • without weight-sharing constraint, GAN generate unrelated image


  • correlated to the weight sharing of G
  • uncorrelated to D

2.3. Face

2.4. Color and Depth Image



3. Application


3.1. Unsupervised Domain Adaption (UDA)



  • [MNIST(labeled)-UDA(unlabeled)]
  • attached a softmax layer c to last hidden layer of D, train on MNIST, predict on UDA


3.2. Cross-Domain Image Transformation

Given x1 in domain 1, find corresponding image x2 in domain 2. As for CoGAN



  • get the most suitable z* for x1
  • use z_* generate x2