When it comes to creating from what has been learned, machines fail; the images they produce are often flawed and fall short of convincing realism. How to teach a computer to invent a face that does not exist in reality? During the discussion in that bar in Montreal, the suggestion was made to develop a statistical treatment of many essential details in the representation of an object. But this method would multiply the data in such a way that each new concrete application would require a monumental work. Goodfellow had a better idea: why not put two neural networks in competition with each other to learn from their mistakes?
That night, Goodfellow began writing the code that would give rise to GANs: one of the networks, the generator, learns to create images; the other, the discriminator, evaluates them to decide if they are real or not. The generating network improves its creations to try to deceive the discriminator, which in turn improves its ability to distinguish between the real and the artificial. Unlike generating networks without an opponent, GANs can be trained with only a few hundred images.
But if the concept of GANs is distantly reminiscent of the Turing test, in which a machine tries to trick a human evaluator into believing that it is a person, it is because the idea of antagonistic training has been in the making for decades. In the early 1990s, Jürgen Schmidhuber, today the scientific director of the Swiss AI laboratory IDSIA, published a system made up of two networks that fight each other.
The GANs are also providing services to science and medicine; for example, improving astronomical images or modeling the distribution of dark matter in the universe, or aiding diagnostic imaging. All these applications are based on the great virtue of the GAN.
However, even these amazing tools have their limitations. In 2021, a study by the State University of New York revealed that GANs often ignore something that even a child knows: that human pupils are round. Those seemingly perfect fictional faces created by GANs often have irregular pupils, because the machine has yet to grasp such a basic concept of human anatomy.