Generative Adversarial Networks are transforming what we’re able to do with neural networks, and it’s unfortunate that almost all the press goes to those wildly accurate facial constructions like that of This Person Does Not Exist. GANs have some incredible potential so let’s take a look at some really unique use cases that display the power of this advancement.
What Are GANs?
For those of you non-data scientists, GANs are a type of neural network that relies on two different components, one to generate content (adaptive network) and the other to test it (discriminator)– thus “generative adversarial.’ The discriminative network attempt to distinguish between real and generated content, helping the generator learn through each iteration.
It’s a type of unsupervised training in which the adaptive network attempts to fool the discriminator, improving accuracy and helping the machine learn what constitutes an acceptable degree of accuracy.
So what does this mean for you? GANs is a new way for neural networks to generate accurate, high-quality content through unsupervised training, making it cheaper and faster to produce content required for deep learning training.
Unique Uses for GANs
Here are some ways we’re taking advantage of this new learning power. Each has the potential to revolutionize a field or at the very least, making research and production a whole lot easier.
Text To Image Synthesis
What if you could type a description in and a machine could generate an image based on the text? GANs makes this possible. This has historically been very difficult for machines to translate, but recent research has shown promising results.
Examples from existing architecture yielded only generalized images or poor quality ones that failed to capture the nuance of the descriptions adequately. While the ability to produce good results isn’t quite ready for production yet, GANs has shown us that it offers results far superior to existing text to image models.
Standard mapping inputs work on the entire function, but what if you only want to manipulate part of the photo? Let’s say a subject is wearing sunglasses in a photo and you want to remove them. GANs features could allow us to manipulate model distribution, removing the sunglasses for better recognition.
This could also have implications for things like aging photos, improving resolution, and helping with facial recognition in poor conditions. Freedom to manipulate only a small portion of the photo without distorting the whole image could help us in a wide range of uses.
It also helps in training data because it’s time-consuming to create GANs with different facial features. Once the system learns this joint distribution, it can generate its own manipulations for learning.
Drug discovery is a slow process. Traditionally, a single hypothesis is tested over a range of years if not decades with massive human input and substantial resource investment. GANs, on the other hand, can rapidly generate novel biological components to test hypothesis simultaneously.
For example, Insilico is using GANs to research root causes and drivers of disease by generating novel proteins that could account for a particular disease. They’ve shown really good promise now that experimental validation of these novel biological components generated by GANs. These models can run generate these hypotheses and run them simultaneously, vastly speeding up the time needed for experimentation without also increasing human investment.
GANs can manipulate images based on training data to create new works of art that mimic the training rules and add unique details that would please buyers. This isn’t fantasy. First Christie’s and now Sotheby’s have auctioned off art created exclusively through artificial intelligence.
In some cases, the artist uses the network to manipulate the outcomes, and in others, the network makes its own rules to generate the images, but we could be seeing the next generation of art in which a machine is a tool rather than traditional mediums.
GANs is already being used to create alternative examples to share in instances where we need to share sensitive data with a third party. The generation can approximate the data without revealing anything potentially compromising.
The networks can also create and test code for bugs, helping developers quickly and accurately. Two networks compete against each other to create code and then cracking that code, allowing the networks to learn through each generation. Google currently has a considerable research avenue open for precisely this kind of privacy training.
It’s expensive and time-consuming to score scenes, and expensive and time-consuming to license already created music for that same distribution. Enter GANs. It uses a discriminator to learn the distributions of melodies and exploiting prior knowledge to generate new melodies based on our parameters.
A recent model named MidiNet came up against Google’s MelodyRNN model and was found to be just as pleasing to listen too, but with the added bonus of being more interesting. The compositions based on algorithms are very realistic and could soon find their way into scores we hear as we watch our movies or tv shows.
The Future of AI Learning
We’re trying to help machines learn more quickly and more efficiently without so much human intervention. By pitting two networks against each other, we may be able to produce faster results in a shorter period without increasing our human input. It’s an exciting prospect and may yield the next big wave of advanced learning and capability.