Basic idea of GAN
We will control what to generate latter -> Conditional Generation
We need to prepare database which is real pictures.
This is where the term 'adversarial' comes from
You can explain the process in different ways...
Note that we fix discriminator parameters and use **gradient ascent ** to update weight.
Why Structured Learning Challenging?
Can generator learn by itself?
Answer is "The most probility is noise!", how to solve it ?
Can discriminator generate?
Discriminator training needs some negative examples.
Use generator to slove argmax
Conditional Generation by GAN
c:text, x:image pair
Unsupervised Conditional Generation
平均分数 / 5. 投票数量：