Whats next in GANs

Now that you have been deeply exposed to deep learning and Generative Adversarial Networks (GANs), in this chapter, you will learn about the possible future avenues for GANs! We start with a summary of this book, the topics that we covered, and the knowledge that we have gained so far.

Next, we address important open questions related to GANs that are essential for interacting with GAN models. We briefly pose questions related to how important architectures are, whether GANs really learn the target distribution, whether GANs are dependent on the inductive bias of architectures, and how to identify GAN samples.

Following this, we consider the artistic use of GANs in the visual and sonic arts. In the visual arts, we provide examples of painting and video generation; while in the sonic arts, we provide examples of instrument synthesis and music generation.

Finally...

What we've GANed so far

The GAN framework has revolutionized the fields of machine learning and deep learning. Over the course of the chapters in this book, we implemented many models in many domains that were part of this revolution, including images, text, and audio.

Generative models

We learned about deep learning and generative models in general, and their applications in AI. We covered many topics including GANs, autoregressive models, variational autoencoders, and reversible flow models.

We described in detail the building blocks of GANs, including their strengths and limitations. We learned how to visualize their results and how to evaluate them qualitatively and quantitatively.

...

Unanswered questions in GANs

GAN-based research is fertile and new architectures, loss functions, and tricks are being released on a daily basis. In this context of constant change, we enumerate a few questions that still need to be answered.

Are some losses better than others?

As we addressed earlier, in the paper, Are GANs Created Equal? A Large-Scale Study, the authors state that in their experiments, they did not find evidence that any of the tested algorithms consistently outperformed the non-saturating GAN. This leads us to wonder whether some losses are, in fact, better than others. We should bear this in mind when choosing a GAN framework.

...

Artistic GANs

In this section, we are going to explore the uses of GANs in the visual and sonic arts.

Visual arts

There are many GANs that produce impressive visual artifacts that can be used in the arts. Examples include the generation of paintings, anime characters, and fashionable clothes. Here, we provide a very small snippet of the visual work that is done using GANs.

GANGogh

GANGogh is the product of semester-long research performed by Kenny Jones and Derrick Bonafilia at Williams College in 2017.

In their project, the authors scoured the WikiArt database, which...

Recent and yet-to-be-explored GAN topics

In this section, we will cover a few recent and yet-to-be-explored topics of GANs that are challenging, interesting, and valuable.

In my opinion, one of the most interesting topics in GANs and deep learning is verified AI. This topic was described in Sanjit Seshia's Towards Verified AI paper in 2016 and is later addressed in a blog post by Google's DeepMind team. There are many challenges involved in achieving verified AI. Some of these challenges include testing, training, and formally proving that the models are specification-consistent.

Other fields that have recently received attention from GAN researchers include biology and its related subfields. There are GAN models that address the problem of drug discovery (3D Molecular Representations Based on the Wave Transform for Convolutional Neural Networks) and real-valued time...

Summary

In this chapter, we learned about the recent advances of GANs. We started with a summary of this book and expanded our knowledge from the simplest GAN framework to state-of-the-art GANs. We then addressed a few important open questions related to GANs. We then considered the artistic uses of GANs in the visual and sonic arts. Finally, we briefly explored new and yet-to-be-explored domains within GANs.

Closing remarks

By now, you should have acquired a broad understanding of deep learning and a deep understanding of the GAN framework. We are confident, that by now, you are able to use the GAN framework to train your own state-of-the-art models for several tasks and domains. We look forward to seeing your models being shared on GitHub and deployed in real life.

Join the revolution, seek adversarial relationships, and collaborate with the future of GANs because: yes, we GAN!

Further reading

An extensive list of GAN models, also known as GAN Zoo, and further reading can be found at the following URLs:

• Minchul Shin has a deep list of GANs and their applications on his GitHub page at https://github.com/nashory/gans-awesome-applications.
• Avinash Hindupuer has a deep alphabetic list of GANs on his GitHub page at https://github.com/hindupuravinash/the-gan-zoo.