You became so GAN!

The GAN concept is often times explained through the idea of a counterfeiter, i.e. the Generator, and police, i.e. the Discriminator [1, p.1]. To introduce it through the game theory approach this is very helpful. Nevertheless, there sticks this bland aftertaste, doing some illegal stuff.

Neural networks are usefull when an explicit analytical function is hard to derive or simply still unknown. The combination of neural network and GAN elevates that idea to the next level. For example in law, there exist regulations which might prohibit certain properties or actions on a product. Besides that, creativity is allowed and even desired. Though, from the regulators, there comes no clear rule to obtain a certain product for a potential producer, or creator, or generator. By having samples previously examined by the regulators officials, a training dataset can be obtained. In that sense, a GAN, especially the generator, can fill the gap between the blurry rule for generation of samples in accordance with the law. This means, in this scenario producer-regulator, it is the other way around than to the counterfeiter-police intro. It is bascially the same, but from a obliance to the law mindset.

Applicant and examiner scheme
Fig.1 - A business owner on the left tries to learn a certain style and image configurations needed to fullfill the examiner restrictions on the right. The owner hands over his application image and the examiner knows some static rules and has experience due previous applications, which he learned from.

One could imagine somebody wants to create images from a specific style, type or which contains some specifications or regulations. To draw those images, the first player could start drawing images and hand them to an examiner which either knows all the rules and regulations or has a bunch of previous images on the side and compares them. By doing so, the examiner forms some intrinsic knowledge of the regulations and deploys them to new images. This makes then sense, when there are regulations or rules applied, but often times law is formulated in a interpretable way and not in an analytical way so that from that rule a model can be deduced. Here some bells for neural networks are already ringing.

Two adversarial model schema
Fig.2 - Two adversarial models instead of people like in figure 1.

That said, the usage of neural networks is justified due to multiple reasons. In case no analytic or fixed rule based model can be deduced, either due to lack of knowledge or lack of resources, a neural network may help out. On the other hand, the learning process defined by the iterative training process for GAN can be implemented through backpropagation is a second reason.

Two adversarial neural networks schema
Fig.3 - Two adversarial neural networks instead of the general term model like in figure 2.

To automize a process, instead of using real persons to fullfill the above task, a model has to be applied. The models can be choosen to be neural networks.

[1] A game–theoretic approach for Generative Adversarial Networks - This paper reflects on the idea of a GAN as a stochastic Nash equilibrium problem.