Formula consists of two steps:
- The first step: adjusting the weight discriminative model D, such that the two has its maximum value V
- Step Two: generative model G adjust weights V in the second term such that the minimum value acquired
First, analyze the meaning log D (x) of:
- D (x) indicates a higher discriminative model D original score for a sample, score indicates more D tend to believe that the sample is a sample of real
- D (G (z)) represents the discriminative model D to generate samples of a score, the higher the score, the more likely that the D represents a real generate a sample as a sample
Therefore, network training process is summarized as follows:
- The first step: training D, so that the above two desired maximum
- The maximum expected value of the first term, represents a true sample D will be given a higher score
- The second maximum expected value, D represents a given sample will generate a low score
Step Two: Training G, such that the minimum desired value of the second term
- The second minimum desired value, namely: to find a G, it is possible to obtain samples generated a high score in the discriminative model D
Figure: GAN training convergence
Blue indicates D
Green indicates G
Black represents the original data
August 18, 2019
In South Lake