How machines can learn to think efficiently

Head of Artificial Intelligence at Meta researches a new method of teaching robots to make better decisions

From science fiction to the social imaginary, a question has always permeated discussions about technology and artificial intelligence: Would it be possible for robots to have autonomy of thought and sensitivity as or more acute than that of human beings?

It may seem like the premise of a futuristic film, but it's not. In fact, scientists around the world are seeking to discover ways to make this human capacity a reality for artificial intelligence, reports the Deputy.

This is the case of the vice president and head of AI at Meta (the group that contains Facebook), Yann LeCun, who researches ways to train algorithms to learn more efficiently. This type of study is important because, unlike humans and animals, artificial intelligence does not yet have an essential characteristic for autonomous learning: reason.

This is why AIs have so much difficulty in learn with one's own mistakes or with observation of situations. “An autonomous car would have to fall off cliffs multiple times before it understands that this is a bad idea,” explains Yann. “And you will have to fall thousands more times before you understand how to avoid falling off a cliff” — this is because machines are distinguished from humans and animals by not having the capacity for common sense.

While, for us, having common sense in situations like the car on the cliff seems commonplace, scientists still face great difficulty in enabling machines with this ability. For LeCun, this is a practical problem: “we really want cars that drive themselves, home robots, virtual assistants intelligent”, exemplifies the scientist.

Will it be possible to give autonomy to artificial intelligence?

Although these technological advances are still far from reality, studies like LeCun's move towards this future. His research proposes an architecture that can minimize the number of actions a system needs to learn and perform a task assigned to it.

This architecture works in a similar way to the different sections of the brain, which are responsible for different functions of the body. Thus, the scientist suggests an autonomous intelligence model that would be composed of five separate modules, which can be configured separately.

The study still faces difficulties, but the potential of the proposal lies in the possibility of allowing the knowledge acquired by one of the modules to be shared among themselves, functioning, in a way, like our memory.

For now, robots that think and have feelings are just fiction — but, if it depends on the efforts of scientists like Yann LeCun, it won't be long before these autonomous artificial intelligences become reality.

Was this content useful to you?

Related posts