The A-Z of AI

GANs

Two AI systems that learn by competing with each other.

Generative adversarial networks (GANs) are pairs of AI systems trained to create content and accomplish tasks faster than one system working alone.

Imagine the task is to generate an image inspired by the artistic style of Picasso. AI design teams could gather together all of Picasso’s paintings and train a GAN to spot the colors, characteristics and individual brushstrokes that make his artworks unique.

One AI system would attempt to copy Picasso’s work, while the other judges its attempts.

The copycat applies its knowledge to producing thousands of new images in the style of Picasso — drawing on features from existing artworks — while the other AI system judges how similar the creations are to Picasso’s style and rates them. Any that aren’t convincing are given back to the copycat for improvement.

Eventually, after bouncing ideas back and forth millions of times, the copycat gets better and better at creating paintings in the style of Picasso.

The ability for GANs to go beyond simply memorizing what’s been made before to create new content was believed to be an important milestone within the AI research community. Designers and architects are already exploring their potential to help generate 3-D models of cars and buildings, based on the study of 2-D photos.