They find that AI can learn by imitating you: a big step towards AGI? – LEBIGDATA.FR

Google DeepMind researchers have just made a major discovery: just like humans, AI can learn in real time through imitation! The implications could be huge, as the entire AI training process could be redefined… find out everything you need to know about this new breakthrough!

Surely you know: human intelligence relies heavily on acquiring knowledge from other people. THE knowledge is passed down from generation to generationsince the dawn of our cultural evolution.

This the ability of social learning allows us to imitate actions and behaviors in real time. However, at the time of arrival Artificial Intelligencea fundamental question arises: machines can also learn Like this?

AI researchers were interested learning by imitation. The idea is to instruct algorithms to observe humans as they perform a task and then try to imitate them.

However, AI tools usually need more examples and exposure to massive amounts of data to successfully copy actions. At least it was until now…

The artificial intelligence managed to explore the virtual world by imitating an expert

Groundbreaking study published by researchers by DeepMind reveals that AI agents can also demonstrate social learning skills in real time, imitating people in new contexts’ without using any pre-collected human data “.

More specifically, the team focused on a specific form of cultural transmission: observational learning or imitation (multiple shots). In short, it is about copying body movements.

This experiment was carried out in a simulated environment called GoalCycle3D : a virtual world where the terrain is uneven, strewn with obstacles and where agents have to navigate.

The ability to quickly acquire new cultural knowledge is one of humanity’s distinctive strengths. For the first time, this ability appears among incarnates #HAVE in complex 3D worlds @edwardfhughes, @reverettai @GoogleDeepMind

— Nature Communications (@NatureComms) November 29, 2023

To help AI researchers learn teaching technique of strengthening or Reinforcement Learning : A machine learning method that consists of offering AI rewards for each behavior leading to a desired outcome.

In this case, this was the result find the right path to complete the route. In the next step, the team added expert agents, coded or controlled by humans who already knew how to navigate the simulation.

Agents quickly learned that the best way to achieve a goal is to learn from the experts. AND DeepMind researchers were very surprised by two observations.

DeepMind believes it has found the key to AGI

They realized that AI not only learns faster by imitating experts, but that it is also capable apply the acquired knowledge to other virtual journeys.

In addition, DeepMind found that AI agents can continue use your new skills even in the absence of experts. In the eyes of the authors, the conclusion is clear: this is indeed an example of social learning.

So while the authors admit that more research is needed, they believe that cultural evolution could play a major role in the development the first AGI or Artificial General Intelligence : a An AI capable of learning to perform any task with the same versatility as a human…

In addition, they hope their work will lead to greater interdisciplinary collaboration between the fields of AI and cultural evolutionary psychologyespecially to better understand the human brain.

A revolution for the AI ​​industry?

This discovery is not that the beginning of the journeyat the end of which there could be a real revolution for the artificial intelligence industry.

It might allow it reduce massive resource requirements for training algorithms and at the same time significantly increase their problem-solving abilities.

It can also be a major twist humanoid robots like Tesla Optimus. They could learn to play manual tasks by directly observing human workers !

Finally, AI could learn directly adopting social and cultural elements of human thinking… if you want to discover this research paper in more detail, you can view the full study in Nature Communications by clicking this link !

This item They find that AI can learn by imitating you: a big step towards AGI? was published on LEBIGDATA.FR.

Leave a Comment