Theoretical Issues

Generative Adversarial Networks

  • Source code
  • Contribution Points
    • Induce the concept of Generative Adversarial Networks for replacing traditional Probabilistic Graphical Models
    • Prove the existence and the convergence of the solution of it's main objective function

Unsupervised Representation Learning With Deep Convolutional Generative Adversarial Networks

  • Source code
  • Contribution Points
    • Make GAN stable to train in most settings
      • Replace any pooling layers with "Strided convolutions(discriminator) and fractional strided convolutions (generator).
      • Use batchnorm in both the generator and the discriminator except for the generator output layer and the discriminator input layer.
      • Remove fully onnected hidden layers for deeper architecture.
      • Use ReLU activation in generator for all layers except for the output, which uses "tanh".
      • Use LeakyReLU activation in discriminator for all layers.
      • Preprocess scaling training images to the range of [-1, 1] for "tanh"
    • Show the Generators have interesting vector arithmetic properties allowing for easy manipulation of many semantic qualities of generated samples

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Energy-based Generative Adversarial Networks

  • Source code
    • Not support the author's official code
  • Contribution Points
    • An Energy-based Formulation for generative adversarial training
    • A proof
    • A EBGAN framework with the discriminator using an auto-encoder architecture in which the energy is the reconstruction error.
    • A set of systematic experiments to explore the set of hyper-parameters and architectural choices that produce good results for EBGANs and conventional GANs.
      • EBGAN is more robust
      • EBGAN can generate reasonable-looking high-resolution images from 256x256 pixel resolution


Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

  • $\hat{x} = G(z)$
    • Z is Low Resolution Image, not noise vector Z
    • Output X is Super Resolution Image from G
  • $D(x)$
    • X is Ground Truth High Resolution Images or Generated Super Resolution Images
  • Upscaling using Sub pixel Convlutional Neural Network

    We increase the resolution of the input image with two trained sub-pixel convolution [ D. Rueckert, and Z. Wang. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network. ]

Semantic Segmentation using Adversarial Network

  • Set Segmentation Networks as Typical G(z)
  • invent new network D(x) for matching function, If ground truth class map and segmentation map from G(z) is similar -> 1 otherwise 0
  • New D(x)
    • Input: Target image to segment and Class map
      • ground truth or Segmented map from G(z) Here z is the target image, There is no noise Z

Generative Adversarial Text to Image Synthesis