Source: Deep Learning on Medium
The Six Fronts of the Generative Adversarial Networks
This research summary is just one of many that are distributed weekly on the AI scholar newsletter. To start receiving the weekly newsletter, sign up here.
Generative Adversarial Networks (GANs) is a buzzword in machine learning. Its flexibility enabled GANs to be used to solve different problems, ranging from generative tasks, such as image synthesis, image completion, style-transfer, superresolution, decision tasks including classification and segmentation, and more.
The fast evolution of GAN methods enabled GANs to surpass other methods in most of their strengths, and today it is the most studied generative model, with the number of papers growing each year rapidly.
There are some challenges when it comes to generative models. For instance, a multifaceted review of different trends of thought for guiding authors is lacking and some scholars have now done something about it.
Generative Adversarial Networks: A Review
It’s true that Generative Adversarial Networks spurred lots of interest in generative models bringing about a wave of new works that new researchers may find intimidating.
This paper aims to help the situation by splitting that incoming wave into six “fronts” including Architectural Contributions, Conditional Techniques, Normalization and Constraint Contributions, Loss Functions, Image-to-image Translations, and Validation Metrics.
The division in fronts organizes literature into approachable blocks, that ultimately shows how the area is evolving. Previous studies in the area, which this works also tabulates, focus on a few of those fronts, leaving a gap that the researchers propose to fill with a more integrated, comprehensive overview.
“We do not intend to provide an exhaustive and extensive literature review, but instead, to group relevant works that influenced the literature, highlighting their contribution to the overall scene”, researchers say.
Potential Uses and Effects
“Generative Adversarial Nets” the seminal paper from Goodfellow et al.was cited more than 12, 000 times, and the pace accelerated significantly since 2017 according to Google Scholar. As such, surveys that can make sense of the evolution of related works are necessary. This work does a straightforward review of GANs with a target of becoming an entry point to its vast literature. The work also focuses on updating experienced researchers to the newest techniques.