Meta’s prestigious AI director, Yann LeCun, wants to create autonomous artificial intelligence. In a study, he unveils an innovative architecture enabling a machine to learn like a human being. Find out all you need to know about this crazy project!
In recent years, artificial intelligence and machine learning have experienced unprecedented growth. For the time being, however, there is no an AI with human consciousness..
Recently, a former Google engineer was fired for claiming that a chatbot had become conscious. In reality, we are far from having reached this stage.
Compared to humans or even animals, current AI systems are not endowed with reason. Yet this is an essential feature for the development of a truly autonomous machine.
Such an artificial intelligence would be capable of learn directly from its observations of the real world, without the need for training sessions to accomplish a specific task.
However, for better or for worse, the first general-purpose autonomous AI could soon see the light of day. The director of AI at Meta, Yann LeCuna leading figure in Machine Learning, has just unveiled his vision for a system capable of learning in the same way as human beings.
Current artificial intelligence lacks common sense
In a study published by the Open Review.net websiteYann LeCun describes an autonomous AI capable of learning and experience the world in a way similar to a human.
His proposal is to train learning algorithms to learn more efficiently. Current AIs are not capable of predict and plan for changes in the real worldwhile humans and animals can understand how the world works by observing it, without even needing physical interaction.
For example, teenagers can learn to drive a car in just a few hours of lessons. They don’t need to experience a road accident themselves to do so.
On the contrary, today’s Machine Learning systems need to be trained on astronomical volumes of data to learn how to perform any task.
As LeCun explained when presenting his work at UC Berkeley, ” an AI-powered car needs to fall off a cliff many times to understand that this is a bad idea. Then a few thousand more times to figure out how not to fall off the cliff. “.
The difference lies in common sense”: a notion found in humans and animals, but not in today’s AI. In his study, LeCun describes common sense as a collection of patterns that can help a living being distinguish between what is probable, what is possible and what is impossible.
It is this skill that enables a person to explore their environment, fill in missing information, and imagine new solutions for unknown problems. Although we take common sense to be an innate skill, no scientist has ever been able to prove it. instill it in an AI or Machine Learning algorithm.
Always during his speech, LeCun criticized learning techniques such as Reinforcement Learning. This training method involves rewarding good behavior and punishing bad.
However, according to the expert, this approach will never match human reliability on real-world tasks. According to him, ” it’s a real-world problem because we really want machines with common sense. We want autonomous cars, domestic robots, intelligent virtual assistants… “.
AI architecture similar to human memory
To kick-start the next decade of AI research, LeCun proposes an architecture for minimizing the number of actions required by a system to successfully learn and perform an unassigned task.
The way different sections of the brain are responsible for the various functions of the body, LeCun suggests a model of artificial intelligence composed of five separate, configurable modules.
One of the most complex parts of this architect would be the “ world model module “ allowing the state of the world to be estimated, and imagined actions and sequences of other worlds to be predicted in the manner of a simulator.
Using this unique world-modeling engine, knowledge about how the world works could be shared very easily between different tasks. This architecture is therefore similar to human memory.
However, major efforts will be needed to create autonomous systems capable of dealing with uncertain situations. This ability is nonetheless indispensable for evolving in a world as chaotic and unpredictable as ours…
Yann LeCun: one of the founding fathers of modern AI
It’s far from the first time someone has raised the possibility of autonomous AI, but Yann LeCun isn’t just anyone. Also professor at New York UniversityThis specialist has spent most of his career developing the learning systems on which many modern artificial intelligence applications are based.
In 2013, he founded the Facebook AI Research (FAIR) group Meta’s first foray into the world of AI research. A few years later, he endorsed the role of chief AI scientist within the company.
Since then, Meta has enjoyed several major successes in the field. In 2018, its researchers trained an AI to replicate eyeballs in the hope of simplifying digital photo editing.
In early 2022, Meta’s BlenderBot3 chatbot opened the debate on AI ethics and biased data. And more recently, the firm unveiled its Make-a-Video tool capable of creating a video from text or an image.
By trying to give machines a better understanding of how our world works, Yann LeCun could open the door to a new generation of artificial intelligence…