aiacademy: 生成對抗網路 GAN - Generative Adversarial Network (GAN) Introduction

GAN

• GAN Zoo

Outline

• Generation by GAN
  • Image Generation as Example
  • Theory behind GAN
  • Issues and Possible Solutions
• Conditional Generation
• Unsupervised Conditional Generation
• Relation to Reinforcement Learning

Basic Idea of GAN

step 1. Fix generator G, and update discrimniator D

step 2. Fix discriminator D, and update generator G

Algorithm

(Variational) Auto-encoder

Auto-encoder v.s. GAN

GAN in Depth

Generator

• A generator G is a network. The network defines a probaility distribution PG.

Discriminator

Can we use other divergence?

Sebastian Nowozin, NIPS, 2016

Issues and Possible Solutions

How to Tain a GAN?

JS divergence is not suitable

What is the problem of JS divergence?

Wassertein distance

WGAN

Tip: Improve Quality during Testing

Mode Collapse & Mode Dropping

• Mode collapse
  • generator 開始產生一樣的東西

  

• Mode Dropping
  • generator 同一類

  

Tip: Ensemble

Objective Evalution