Overview
This posting series is a study note that records the process of learning the book "Understanding Deep Learning".
This time, I will cover Chapter 15, Generative adversarial networks.
1. Generator vs Discriminator
As you can see from the word "Adversarial", in GAN, learning proceeds in a competitive structure between two models.
GAN is divided into a Generator model that tries to generate data as similar as possible to the sample data, and a Discriminator model that tries to distinguish the data generated by the Generator from the sample data as much as possible.
I was very surprised by the concept that a generative model is learned through competition between two models like this.
2. Conditional generation
I was able to confirm several GAN-derived models such as Conditional GAN, Auxiliary classifier GAN, and InfoGAN.
While using an actual image generation model, I input a more detailed prompt as well as a latent variable, and I could see that such a prompt enters as a parameter of conditional generation.
The actual generation model is a generative model based on a Diffusion model, but I was able to recall this by seeing it enter as a conditional parameter in GAN.
Reference
[1] Prince, S. J. D. (2023). Understanding Deep Learning. The MIT Press. Retrieved from http://udlbook.com