Making machines learn on their own: this is, today, the great challenge of artificial intelligence, AI. With this objective and equipped with techniques that imitate the learning process of humans, the area of ”machine learning” or machine learning was born, whose focus is to enable computers to perform tasks alone through observations or interactions with the environment.
The simple idea — just feed a machine with data and define a purpose — is inspired by the way humans learn, that is, they observe patterns from the moment they are born. A child comes to recognize a dog after observing it several times. After releasing a toy in the air and repeating this action, she internalizes that such an object will fall. Observing patterns and interacting with the world is how we experience reality, an experience that is associated with how we interpret the environment.
Based on this idea, the area of AI has had several success stories, such as defeating the human champion in the game go, to solving a problem in biology that had been ingrained for fifty years. Here, however, we are dealing with situations in which he was taught both the game and the relationship between proteins. It is not surprising that we ask ourselves if, in addition to these accomplishments, it would be able to create or generate things by itself, which for human beings is natural. Our intelligence is, in general, closely associated with our ability to build, from works of art to city projects and cooking recipes – in short, imagination is the limit.
The question motivated computer scientists to investigate how an AI would be able to understand and extract information from data, which are fed to the model, and based on it, build new data. It was then that generative networks gained strength. And what are generative networks? These are networks that are taught to remake versions of data after understanding their distribution. How to produce images based on drawings or assemble a house after looking at different photos of houses. The combination of generative networks with the potential that neural networks offer was an important milestone for the advancement of this proposal.
One of the first tasks of these generative networks was to acquire the ability to reproduce human faces. Over the years, the success of this acquisition has been such that today it is difficult to discern whether an image is in fact a photo taken by human eyes or the creation of an AI. It didn’t take long and new challenges emerged, such as creating poems and remaking a city.
Recently, one of these networks managed to 3D remake the city of San Francisco, California, after studying photos taken by cars. The feat was achieved by Waymo, a Google company that aims to use this AI to train autonomous cars without the risk of accidents. Allied to a possible revolution in the video game industry, this would be one of the first direct applications of an AI production.
The big revolution in the field of AI in recent years came from the GPT-3, a monstrous model with 175 billion parameters that uses generative networks that can generate conversations, books and even code. A version with 12 billion parameters, DALL-E, was built to form images based on others or from text. But the breakthrough came with the second version of DALL-E, which both creates text-based images like “teddy bears working on a computer on the moon” and paints paintings in the style of great artists, with results strikingly similar to the painter who was emulated.
Little by little, machines are understanding how to produce the world we live in, be it the physical part, as an entire city in 3D, or the way we interpret reality through art, literature, photography. Perhaps the time has come for us to start asking ourselves if in some near future they would not be able to create worlds based on ours.
*
Roberta Duarte is a physicist and doctoral student, and works with applications of artificial intelligence in astrophysics.
Sign up for the Instituto Serrapilheira newsletter to keep up with more news from the institute and from the Ciência Fundamental blog.
