Back in 2014, Ian Goodfellow developed this concept of dueling neural networks – taking two neural networks and making them compete with each other “in a digital cat-and-mouse game.” This was later known as generative adversarial networks (GANs).
How it Works
Neural networks are mathematical systems that learn by analysing data sets. The two dueling neural networks train using the same data set. The generator, one of the neural networks, creates new images by modifying pictures it has already seen; for example, putting an extra arm on a picture of a person. The other neural network, the discriminator, is trained to identify if the image it sees is from the initial data set or a fake.
As with all AI algorithms, the generator overtime learns to produce photorealistic images such that the discriminator cannot spot the fakes anymore and essentially “loses”. In fact, it’s been said that generators can produce images that can even fool humans!
“The models learn to understand the structure of the world,” Goodfellow explains. “And that can help systems learn without being explicitly told as much.”
Applications
GANs are extremely promising in several applications. Most notably, MIT technology review reports that they have created realistic fake images and realistic sounding speech. For instance, researchers from Nvidia gave a GAN a set of celebrity faces. From there, it produced hundreds of realistic and credible faces of people who don’t actually exist; these machine-created faces resembled the famous faces.
In biotechnology, GANs have been used for drug discovery (i.e. developing new medications). Data of molecules is fed into the GANs to find molecules that can be developed into new medicines.
Furthermore, Wired reports that GAN can potentially deliver unsupervised learning too. Ultimately though, this technology can help with predictions of the future and constructions of simulated worlds – undoubtedly useful for many companies.
Sources
Written by Nichapatr (Petch) Lomtakul