Today, Artificial Intelligence (AI) is already used in many fields, for the most varied purposes. In fact, this capability has proven to be very useful, although also endowed with a certain dangerousness. However, it does not yet possess a capacity inherent to humans: imagination.

So that's the flaw that scientists are addressing.

A IA is, nowadays, already capable of performing the most varied tasks, making the everyday life of many people around the world easier. However, although humans are born with a keen capacity for imagination, machines can only recognize what they have seen before.

According to the portal "Pplware", as exemplified by Science AlertOnce humans know what a cat is, they can easily imagine it in a different color, in different positions or even in different environments. However, while they can be trained to recognize a cat when they see one, for AI systems, this task becomes difficult, since it is typically trained to detect specific patterns.

Still, because a group of scientists is working on that field of AI - that of imagination.

AI Imagination may hold the cards in the future

To try to unlock the imagination capacity of AI, researchers have discovered a new method. This will allow AI systems to figure out what an object is supposed to look like, even if they have never seen exactly what it looks like.

We were inspired by the human capacity for visual generalization to try to simulate the human imagination in machines.

Explained Yunhao Ge, computer scientist, in University of Southern Californiaadding that, just as humans are able to separate acquired knowledge by attributes - such as shape, position, color - and recombine them, forming a new object, so AI systems will be able to do it. To do this, the process will be simulated using neural networks.

What the team has presented is called controllable disentangled representation learning, and it uses an approach similar to those used to create deep falsifications. That is, it triggers different parts of a sample - if the AI system sees a red car and a blue bike, it will be able to imagine a red bike, even if it has never seen one before.

In a novel way, the scientists put the information together in a framework they called Supervised Group Learning. In this way, the processing of samples is done in groups, rather than individually. This allows a semantic construction between them, throughout the study.

So the AI is able to recognize similarities and differences in the samples it sees, using this knowledge to produce and imagine something new.

In the future, the AI system could help build neural networks that are not racist or sexist. In addition, the researchers say it could also be an asset associated with autonomous driving or in a wide variety of medical fields, by imagining scenarios for which it has not been trained.