Effective shortcut technique for generative adversarial networks

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 53; no. 2; pp. 2055 - 2067
• Main Authors: Park, Seung, Yoo, Cheol-Hwan, Shin, Yong-Goo
• Format: Journal Article
• Language: English
• Published: New York Springer US 2023; Springer Nature B.V
• Subjects: Architecture; Artificial Intelligence; Computer Science; Datasets; Deep learning; Generative adversarial networks; Generators; Image processing; Machines; Manufacturing; Mechanical Engineering; Processes; Regularization methods; Gated shortcut; Residual block; Generative adversarial Networks; Image generation
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-022-03666-2

In recent years, image generation techniques based on generative adversarial network (GAN) have been used to design their generators by stacking multiple residual blocks. A residual block generally contains a shortcut, that is skip connection, which effectively supports information propagation in the network. In this paper, we propose a novel shortcut method, called the gated shortcut, which not only embraces the strength point of the residual block but also further boosts the GAN performance. Specifically, based on the gating mechanism, the proposed method allows the residual block to maintain (or remove) information that is relevant (or irrelevant) to the image being generated. To demonstrate that the proposed method significantly improves the GAN performance, this paper includes extensive experimental results on various standard datasets such as CIFAR-10, CIFAR-100, LSUN, and tiny-ImageNet. Quantitative evaluations show that the gated shortcut achieves the impressive GAN performance in terms of the Frechet inception distance (FID) and inception score (IS). For instance, the proposed method improves the FID and IS scores on the tiny-ImageNet dataset from 35.13 to 27.90 and 20.23 to 23.42, respectively.