(작성중) [Generative Model] Generative Adversarial Networks(GAN)

이 글은 GAN에 대한 필자의 이해를 높이고자 작성된 글입니다. 참고자료는 자료1 입니다.

Concept of Generative Adversarial Networks(GAN)

field of(Generative model-latent variable model), A neural net that maps noise vectors to observations

• Training: use the learning signal from a classifier trained to discriminate between samples from the model and the training data
• Pros
  • Can generative very realistic images
  • Conceptually simple implementation
  • Fast generation
• Cons
  • Cannot be used to compute probability of observations
  • "Model collapse": Models ignore regions of the data distribution
  • Training can be unstable and requires many tricks to work well