At A Glance – Generative Adversarial Networks – GANs

Neural networks can now train and improve themselves

It is often said that if you want to improve at a certain game, you should play against somebody who is better at it than you are. It is upon this principle that Generative Adversarial Networks, or GANs are based. GANs are actor critical, algorithmic modelsĀ used to train deep neural networks.

In a GAN, two networks compete using unsupervised machine learning. The first network, the generator, generates a data instance (for example, an image) which mimics real world data. This is fed to the second network, the discriminator, along with authentic examples from the real world. The generator’s aim is to convince the discriminator that the generated data is authentic, while the discriminator has to discern between which data is real and which is replicated. GANs are based on a zero sum game framework, which means that one party’s gain is equivalent to the other’s loss. The two algorithms are locked in a constant cat and mouse chase that enables them to improve exponentially over time.

As Artificial Intelligence expands, GANs have become an important tool to improve the capabilities of deep neural networks without requiring constant human supervision. They have been used to produce samples of images for interior and industrial design, apparel, and computer game environments. They have also been used to model patterns of motion in videos, and have even created 3D models of objects in authentic images. However, whilst they are incredibly useful, they can take days to train. Both networks need to be developed carefully, in tandem, so that they improve together.

To learn more about the expansion of AI, sign up for our free newsletter.