Since the victory of the Deep Blue computer against chess champion Garry Kasparov in 1997, the machine has continued to demonstrate its ability to surpass human beings in increasingly complex tasks… without still second to none in terms of adaptability and versatility.
Deep Blue’s victory was a “culturally incredible” event, but “technologically, it was only a great success”, explains to AFP Philippe Rolet, doctor in artificial intelligence.
Basically, the success of Deep Blue is only a victory of “brute force”, explains the co-founder of Artefact, a consulting and technological development company in this field which has 800 employees.
At this moment, the machine wins thanks to its formidable computing power, which allows it to determine all the possibilities of evolution of the game very far in the game, and to deduce the best piece to play.
In reality, the real revolution is at this time barely emerging from the laboratories: it is that of machine (or automatic) learning, and artificial neural networks, which have made immense progress in the artificial intelligence over the past decade.
While Deep Blue knew how to play chess thanks to a whole architecture of logical rules instilled by human beings, the new machines forge their own rules themselves, in a learning period where they ingest mountains of data.
“We are moving from imperative programming to programming by learning”, explains Philippe Rolet.
In games, the effect is startling. In 2017, DeepMind company’s AlphaGo algorithm, based on machine learning, beat the world’s best go player, Ke Jie.
Before AlphaGo, “there was speculation about the handicap that should be given to God for man to be on an equal footing with him”, recently said Cédric Villani, the mathematician deputy, author in 2018 of a founding report on the artificial intelligence (AI).
“And then seeing the level of crushing of the human” by AlphaGo, “we understood that the human was much worse than he thought. starting rookie mistakes turned out to be great moves,” he continued.
– “A general self-learning method” –
Today, machines are even capable of beating humans in games with uncertain universes, such as poker or bridge, as the French start-up NukkAI recently demonstrated.
These giant steps completely go beyond the world of games, where the machine no longer has much to prove.
In recent years, artificial intelligence has made “absolutely incredible progress which surprised me myself”, explains to AFP Yann LeCun, the head of AI research at Meta / Facebook, and one of the founding fathers of modern AI.
“Today we are able” to allow a machine “to translate any language into any language in a set of 200 languages” or “to have a single neural network which includes a hundred languages “, he explains.
But the machines still come up against obstacles.
“Just because we can have a fun dialogue with GPT3”, the amazing text generator from Elon Musk’s OpenAI startup, doesn’t mean it “will be able to help us in everyday life “, he nuances.
What is missing to design this virtual assistant – and probably also to manufacture the real autonomous car – is to arrive “at a general method of self-learning”, underlines Yann LeCun.
“We would put the computer in front of 200 hours of video, and from that, it would arrive” at a form of “understanding of the world”, of “common sense”.
This is what would then allow him to achieve learning capacities “closer to what we observe in animals and humans”, he explains.
The researcher is convinced that machines will one day have “a universal capacity to learn, capable of learning everything that humans learn, in most cases with superior capacities”.
“But when will it happen? The answer is not clear,” he concludes.
We would love to give thanks to the author of this post for this outstanding material
Since Deep Blue/Kasparov, the dazzling progress of the machine against…
You can find our social media profiles , as well as the other related pageshttps://www.ai-magazine.com/related-pages/