Generative adversarial networks with adaptive learning strategy for noise-to-image synthesis.

Generative adversarial networks (GANs) directly learn from an unknown real distribution through adversarial training. However, training the generator only by the feedback of the discriminator cannot make GANs learn adaptively from the unknown complex real distribution, and for this reason the quality of generated images is unsatisfactory sometimes. To address this problem, we propose a framework for training GANs with an adaptive learning strategy from simpleness to complexity. First, we employ a pre-trained encoder and a generator to construct a simple task that looks like a real image. Second, an adaptive learning strategy is designed based on the mathematical expectation of the discriminating results of the real image and the simple task. The designed adaptive learning strategy is well compatible with various GANs architectures. Experimental results demonstrate the proposed method can improve the performance of existing GANs. [ABSTRACT FROM AUTHOR]