Exploring the Role of Recursive Convolutional Layer in Generative Adversarial Networks

This paper aims to study the potentialities of incorporating recursive layers into Generative Adversarial Networks (GANs). Drawing inspiration from biological systems, in which feedback connections are prevalent, different studies investigated their impact on artificial neural networks. These studies have shown that feedback connections improve performance in tasks such as image classification and segmentation. Motivated by this insight, in this work we investigate whether also image generation can benefit from recursive architectures. To support our argument, we introduce a recursive layer into a standard generative architecture, specifically a Wasserstein GAN with gradient penalty (WGAN-GP), resulting in a novel model we refer to as the Looping Generative Adversarial Network (LoGAN). The performance of the LoGAN architecture is compared with the corresponding feedforward WGAN-GP both qualitatively and quantitatively. Preliminary experiments suggest that the use of recursive layers holds significant potential to generate higher-quality samples in GANs. The code is publicly available at https://github.com/bcorrad/LoGAN.