Title: Generative Adversarial Nets
Summary:
Shorthand: The author proposes a framework through an adversarial process, in which two models will be trained at the same time.
The first model is the generative model G, which is used to capture the distribution of the entire data
The second model, the discriminative model D, is used to estimate whether a sample is generated from G or not.
If G and D are an MLP, then the real system can be trained entirely through an error backpropagation
introduction:
Shorthand: There are two models in this framework, one is a generative model and the other is a discriminative model.
The generative model is like a counterfeiter, making counterfeit money
The discriminant model is similar to the police distinguishing real money from counterfeit money.
In this process, counterfeiters will continue to improve their ability to counterfeit, and the police will continue to improve their ability to distinguish.
Model:
The simplest application of this framework is when your generator and discriminator are both MLPs, the generator needs to learn a distribution called pg on the data X. Define a prior pz that maps z to x. The discriminator also has its own learning parameters, which are used to put data in and judge whether the data comes from real data or data generated by the generator.
min max equalization function:
min max: If equilibrium is reached, D cannot progress, and neither can G
The result of the final request of GAN:
The discriminator is completely indistinguishable from the distribution between the data generated by the generator and the real data.
Theoretical results:
When the generator G is fixed, the optimal solution of the discriminator D is calculated in this way.
When G and D have enough capacity, we allow each step D to reach the optimal solution, then if we iterate the following step for the optimal time zone of G, then this G has been replaced with the optimal solution, then the final Pg will converge to Pdata.
Summarize:
The writing is very clear, focusing on what GAN is doing.
In related work, I wrote a lot of my own ideas, which have been done in previous work, and I also said that really great work doesn't care whether your ideas have appeared elsewhere.
The third block is to introduce the objective function of the entire GAN and how to optimize it
The fourth block talks about some proofs, why the objective function can get the optimal solution