On the Effects of Skip Connections in Deep Generative Adversarial Models
Yulin Yang, Rize Jin, Caie Xu · 2020
Deep convolutional neural networks provide significant contribution to GANs in stabilizing the GAN training. However, the convolution operator has a local receptive field and therefore geometric or structural patterns of complex images can only be processed after passing through several convolutional layers. Recent studies suggest that the network depth is indeed of crucial importance for GANs to generate images with consistent objects/scenarios to a certain extent, but it also amplifies the mode collapse problem. As a solution, we investigate a skipping mechanism for going deeper in GANs. By passing the input of layers as a weighted residual, it alleviates the phenomenon of mode collapse. Experimental results show that the proposed method increases the capacity of modeling long range dependencies while retaining the local invariance property obtained by using a relatively small convolution kernel.