Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs)

1. 1960

   • Herbert Simon

2. Unsupervised Machine Learning

3. Goodfellow et al, 2014

   • Architecture ofGenerative Adversarial Nets
     • Two networks
       • Generator
         • Arch-nemesis
           • Adiversarial
           • Benefitting the overall model
         • "Regression problem?"
       • Discriminator
         • Classification problem

4. Types

   • CycleGAN
     • Exemple
       • FaceApp
     • Architecture
       • Generators
         • Two
           • First domain -> Second domain
           • Second domain -> First domain
         • Often variations of autoencoders
         • Input Image and output imagem
       • Discriminators
         • Two appraisers
         • Input image -> output only number
           • Possibility of the image being real or false
     • Unsupervised
       • Two different domains
     • Cycle consistency
       • regularization of the generator models (additional loss)
     • 2017
     • Transform of Image
       • Transform image-to-imagem in different styles
         • to perform image to image translation on unpaired sets of images
     • PatchGAN
       • NxN patch
         • Real or fake
       • 2016
   • StyleGAN
     • Generate high resolution image
     • NVIDIA
       • ProGAN
         • Progressive GAN
     • Mapping Network
       • Consists of 8 fully connected layers
     • 2018
   • Vanilla GAN
     • First GAN
     • 2014
   • TextToImagem
     • Differs from the others by data input
     • Conditional GAN
       • G(z, c) and D(x, c)
         • G to generate images conditioned on variables c
       • DC-GAN
       • GAN-CLS
       • GAN-INT
       • StackGAN
       • StackGAN++
   • DiscoGAN
   • ImageToImagem
     • BicycleGAN
       • 2017
     • Pixel regression or classification

5. Training

   • Distance
     • Wasserstein-GAN 2017
   • Finding the Nash equilibrium
     • Game theory

6. Generative models

   • Auto-Regressive models
     • Learns an explicit distribution governed by a prior imposed by model structure
     • Advantages
       • Provides a way to calculate likelihood
         • returning explicit probability densities (unlike GANs)
         • Interpretability
       • The training is more stable than GANs
       • It works for both discrete and continuous data
         • It’s hard to learn to generate discrete data for GAN, like text
     • preferably used in image completion
     • PixelRNN
     • PixelCNN
     • PixelCNN++
   • Autoencoders
     • aims at maximizing the lower bound of the data log-likelihood
   • Different types
     • Naive Bayes Classifier
     • Gaussian Mixture Model
     • General Adversarial Networks

7. learn implicit data distribution

   • probability distribution of a quantity

Images and files in the map