A Step Beyond Generative Multi-adversarial Networks

摘要

In this paper we modify the structure and introduce new formulation to improve the performance of the Generative adversarial networks (GANs). We achieve this based on the discriminating capability of the Generative Multi-Adversarial Network (GMAN), which is a variation of GANs. GANs in general has the advantage of accelerating training at the initial phase using the minimax objectives. On the other hand, GMAN can produce reliable training using the original dataset. We explored a number of improvement possibilities, including automatic regulations, boosting using Adaboost and a new Generative Adversarial Metric (GAM). In our design, the images generated from noisy samples are reused by the generator instead of adding new samples. Experimental results show that our image generation strategy produces better resolution and higher quality samples as compared to the standard GANs. Furthermore, the number of iterations and the required time for quantitative evaluation is greatly reduced using our method.