Ian Goodfellow designed the GAN “Generative Adversarial Network” – that effectively allows Artificial intelligence (AI) to teach itself. This represents a significant step toward creating machines with a human-like consciousness.
While current AI applications mostly include ‘deep learning’, the actual learning is limited by the provided data set. For example, feed in a million images of a pedestrian crossing a road, tell the computer what is (and is not) a pedestrian – and sure enough, the computer will start to accurately identify pedestrians. This becomes critical for neural networks deployed in the likes of self-driving cars.
But asking a machine to draw images of a human face is an entirely different challenge. The classic approach has been to use complex statistical analysis of the elements that make up a face to help machines come up with images by themselves. But they tend to be blurry or have errors like missing ears.
Goodfellow hit on an idea to pit two neural networks against each other to create a ‘back-and-forth’ rivalry between a ‘picture forger’ and an ‘art detective’ who repeatedly try to outwit one another. The first one, known as the generator, is charged with producing artificial outputs, such as photos or handwriting, that are as realistic as possible. The second, known as the discriminator, compares these with genuine images from the original data set and tries to determine which are real and which are fake.
On the basis of those results, the generator adjusts its parameters for creating new images. And so it goes, until the discriminator can no longer tell what’s genuine and what’s bogus. Genius!
Credit: Portrait of Edmond Belamy, 2018