Exploration of the evolution-based footprint on the Generative Adversarial Networks

In the last few years, the Generative Adversarial Networks (GAN) have proved to be an excellent tool for image generation and found a variety of practical applications. In this paper, the specific evolutional concept of mutation for the generative model training improvement is proposed and examined. The idea behind the research is that the generative model learning can be optionally tackled as a complex optimization task. Such an approach is associated with a two-fold problem raised in the paper, meaning that comparing different generators to each other is hard to define, so is the fitness function of the weight vectors. It can be assumed that the fitness function, if existed, might have an unknown but potentially high number of local minima, which could restrict the gradient descent to move in some directions. The vast number of degrees of freedom usually seen in the image processing layers exceeds the standard problems for intelligent optimization algorithms. Therefore the hybrid combination of the gradient descent with the smart usage of mutations is considered and examined, in particular with the proposed novel approach for the mutation masking strategy.