What Are Generative Adversarial Networks and What Can They Achieve? [ELI5]

Not long ago, the wider sentiment in the AI industry was that "AI can't be creative." Even today, some people hold to that view, though AI is being used to compose music, poems, sculptures, and draw paintings, like the one below:

You can create your own art using deep learning with tools like deepart. To listen to AI-generated music, you can follow YouTube channels like Aiva, which has almost a million video views to date according to SocialBlade. Individual AI-composed hits like Daddy's Car have reached millions of views.

You can create your own AI-generated poetry with Google's PoemPortraits project. Check out this poem written by an AI:

the sun is a beautiful thing
in silence is drawn
between the trees
only the beginning of light

Google actually runs a number of arts and culture projects, many of which are AI-related. You can check them out at Experiments with Google. Another AI-driven experiment is called "Please Feed the Lions," an interactive public sculpture driven by machine learning:

You might ask, how does AI do this? For most AI-generated creative work, whether it's art, poetry, sculptures, or music, it's using what's known as a "generative adversarial network," or GAN for short.

To give an ELI5 explanation, a GAN is based on unsupervised machine learning, which means that it works on data without labels, like images, audio, and video, and it's implemented via two competing neural networks.

For example, let's say you want to build an AI that can draw. Of the two competing neural networks, one neural network might start with a random static image, and calculate the error between this image and an inputted painting. Of course, the error is extremely high, so it makes adjustments to the algorithm in an attempt to decrease the error, until it successfully descends the gradient and makes an image with high likeness to the inputted paintings. In short, it can draw!

If you want to go very in-depth into how GANs work, check out this Towards Data Science article. Ultimately, the evolution of AI into creative fields is incredible to witness, and has grabbed the attention of even the most vocal AI naysayers.