Video

Ian Goodfellow: Generative Adversarial Networks (NIPS 2016 tutorial)

Slide

Why study generative modeling?

• Generate possible features for reinforcement learning
• normal generative modeling sometimes intractable
• Monte Carlo's failure
• Produce better samples...

Structure of GAN

• generator
• discriminator: judge whether it is real

How to choose divergence?

• The idea of initially pop up at the bar, like the Reverse EL
• then he realize there might be many other choices later on