Artificial Intelligence (AI) is, currently, one of the most relevant topics in technology, with a huge potential in the business area. It must be urgently addressed within companies, regardless of size or economic branch of the business, as it can be considered an important part of innovation strategies, process improvement and, finally, delivery and creation of value for customers. By itself, AI is an immense field of study that seeks to develop intelligence through software and other mechanisms. In addition, it has several ways to be applied. In this article, I will explore one of the ways in which AI is applied to the business area: Generative Adversarial Networks (GANs). I will discuss the power of GANs in the development of creativity and innovation, as well as ways in which other companies have already been using them to grow their businesses.
What are GANs?
GAN is a class of Artificial Intelligence that works on generating content that, when analyzed, can easily deceive, leaving consumers uncertain as to whether it is real (produced by humans) or fake (produced by the network). This technique was invented by Ian Goodfellow, with his colleagues, in 2014, and, in summary, it consists of 2 main components: the Discriminator and the Generator . Both Discriminator and Generator are networks that have different responsibilities. The Discriminator is a network responsible for assessing whether a particular content is real or fake. The Generator is responsible for producing the content itself.
The connection between these two components is carried out in an adversarial way, that is, while the Discriminator is trained to differentiate the real from the fake, the Generator is trained to deceive the Discriminator through the content it is producing. Consequently, through the training process of both components, the networks develop jointly in order to improve their responsibilities over time.
In the GAN class of AI, there are several different subtypes of GANs, which are, in turn, applied to different purposes:
Deep Convolutional GANs (DCGANs): are generating and discriminating network architectures for working specifically with images. In general, they are used in the context of generating new images, 3D models, noise removal, fraud identification, among other things;
Conditional GANs (cGANs): are an extension of traditional GANs, with the addition that the data generation must take into account certain conditions or defined attributes. That is, it is possible to lead results based on pre-established conditions passed through the network. Thus, in the example of 3D model generation, it can be led to the generation of models that have conditions determined by the user;
InfoGANs: just like cGANs, an InfoGAN aims to produce content similar to those evaluated by the Discriminator, however, in addition to this evaluation, it also provides a vector of information about the evaluated data, which, in turn, is used in the Generating network in order to produce more consistent and meaningful content. This technique is commonly used in the task of generating more concise content and produces an information vector, which condenses the most relevant characteristics of the learning process;
Super Resolution GANs (SRGANs): networks that combine the idea of using GANs to produce high resolution in images that were previously in low quality. In summary, DCGANs receive low-resolution images as input and, in the end, produce the same image with Resolution Enhancement at each iteration.
These are just a sample of what GANs, among many other approaches, can do . Now that we know a little more about GANs and the scope in which they can act, let’s move on and check out how companies have been using this technology.
GANs have become a tool increasingly used within companies for the exploration of creativity, innovation and new trends and all of this comes from their results in different fields. Since the concept is relatively simple, companies have been exploring it in many contexts, such as text, audio and image generation. In addition, there has been great effort on the part of companies in the development of techniques that help in the evaluation of content that can be classified as fraudulent. Some examples are highlighted below, in order to illustrate the work of these networks in different economic sectors:
Finance: a very interesting use case, presented by an ODSC article, was the use of GANs to simulate the real distribution of the stock exchange prices, rather than the Monte Carlo method. The Monte Carlo method is commonly used to generate these distribution curves, however, by using GANs as the generating source, the approximation was more accurate and obtained a better estimate of the temporal series of the stock exchange. Consequently, the financial return from this data was maximized. In summary, a very simple network configuration was used, following the concept presented in GANs. That is, with real distributions used as input to the Discriminator, the Generator can try to get closer and closer to them, so that it can deceive the Discriminator and produce an increasingly accurate distribution.
Marketing: when talking about marketing, there is a huge range of AI application possibilities to further enhance and improve customer relationship. In regards to GANs, they can be used as strategies to predict and improve how users react to new products and services. In addition, companies in the industry have been exploring purchase profiles of users, simulating them through GANs in order to further improve product recommendations in web services. The strategy used is to present multiple purchase profiles to the Discriminator and then compare the profiles produced by the Generator, so as to classify them as plausible or not. Thus, with the development of GAN learning, the Generating network begins to produce purchasing profiles similar to those presented to the Discriminator, thus allowing a more strategic analysis, based on the synthesized data, to predict future real data .
Healthcare: in the field of Biomedicine, some pharmaceutical companies are exploring the benefits offered by GANs in the task of identifying molecules that are more likely to have desired properties, aiming to develop new drugs and medications. There are millions of components that can be chosen during the stage of the discovery of new drugs. In addition, the synthesis and testing of a new molecule costs millions of dollars. Thus, when there is an improvement in the way of finding a component that is a potential candidate to be part of a drug, this cost is reduced and the process of discovery and preclinical testing are streamlined. The process of generating these components is carried out from conditions and properties that are desired for the medication. Therefore, based on the Discriminator’s data, the Generating network is guided to produce the elements obeying the conditions established by the user [12, 13].
Based on the content presented in this article, several ideas begin to emerge regarding how to apply AI in a few economic sectors and, in a very creative way, one can imagine the use of GANs to aid the development of innovation and process optimization. As mentioned, each type of GAN has a specific characteristic and purpose, however, all of them follow the same goal, that is, to improve information through a dispute between the Discriminator and the Generator networks. Thus, by rephrasing the problem, based on this fundamental objective, several interesting applications can arise, such as those presented in the areas of Healthcare, Marketing and Finance, as well as others that have not been addressed here.