Breaking News

Friday, 26 January 2018

How Facebook is considering the future of artificial intelligence

Far from the fantasy robot companion android, the researchers of the American giant improve the methods that will allow machines to learn faster tomorrow. The humanoid robot Sophia made the front page of the US media during CES. With some notable exceptions, it was highlighted in interviews where his hesitations and his silences were cut in the editing. 

how-facebook-is-considering-future-of-artificial-intelligence

This has led to believe in a fluid conversation between man and machine. What exasperate Yann LeCun, director of FAIR, the artificial intelligence research section at Facebook. The eminent specialist has denounced demonstrations completely "bullshit" on social networks.


"This contributes to a dangerous buzz around AI, and creates unrealistic expectations in the public and in the media about the state of the art in artificial intelligence. It's dishonest and destructive. "

Look at cognitive science :-

how-facebook-is-considering-future-of-artificial-intelligence

Visiting Paris for several conferences open to the public, he drew up the state of the art in terms of artificial intelligence. And according to him, we are still very far from being able to equal the man. "The best artificial intelligence systems have less common sense than a rat," he said abruptly to immediately cut the fantasy of the movie Her, in which the hero falls in love with an attractive AI.

The machines would still be far below our biological capabilities. "Humans and animals build models of prediction that allow them to learn without doing too much nonsense, through observation and interaction and from very few examples." The man does not need to get into a car in a plane tree or fall off a cliff to understand that it is dangerous: he knows in advance what will happen.
On the contrary, artificial intelligence systems need to be enormously trained. This is why Facebook looks beyond mathematics and collaborates with researchers such as Emmanuel Dupoux, psycholinguistics specialist at ENS.

The latter studies human learning skills to draw cognitive models. A baby learns very quickly concepts of direction, position or permanence of objects. "Common sense is a quantity of gigantic knowledge acquired very quickly during the first moments of life. If machines could be allowed to learn these methods so quickly, we would make very big advances in artificial intelligence, "says Yann LeCun.

Communicating with AIs other than through dialogue :-

Even pragmatism on the part of Antoine Bordes, who heads the FAIR laboratory in Paris and one of the Facebook M artisans, an experiment that aimed to integrate a personal assistant to Facebook Messenger.

This bot had to be able to answer the questions of the surfers formulated in natural language. "We started from what people would ideally want. It was thought to be able to automate more tasks but it turned out to be more complicated than expected because the needs expressed were infinite and very personalized, "explains Antoine Bordes.
Facebook researchers have also observed many conversational agents. Their report is without appeal. It's impossible to hold an interesting conversation that lasts more than two minutes with a robot. The models manage to initiate the conversation with formulas like "Hello! How are you ? And can even artificially prolong it by reviving you on your favorite sport but remain unable to react as soon as the course of the dialogue becomes unique. "The man is able to jump from cock to ass, to make connections. The model, he will be lost because he will endlessly find himself in problems that he has never met ", we detail Antoine Bordes.

If the challenge of dialogue is not ready to be solved, it will not necessarily put a brake on the development of voice assistants. For the FAIR teams, Internet users are already used to interacting with search engines and know how to select the right keywords to launch queries. It will not be necessary to exchange with a robot as with a similar one but to develop an effective mode of communication to obtain the accomplishment of a useful task. Like to say "light salon! To your smart speaker rather than "please could you turn on the light in the living room?" ". With in the future, the probable development of uses still today unsuspected.

Combine different methods :-

A pioneer in deep learning, which consists of teaching artificial neural networks to perform certain tasks, Yann Le Cun seems to have already moved on to something else.

Aware of the limits of this deep learning, he is now seeking to further advance artificial intelligence. There are many possibilities to experiment with new neural network structures or to optimize mathematical methods.
But for him, the key seems to be in what he calls "differential programming" and others call it also a neuro-symbolic hybrid learning system. It involves combining different methods: both deep learning and memory, or unsupervised reinforcement learning and programming, dynamics and static. "Maybe some tasks can not be learned. Hence the interest of mixing parts that could be learned with other coded or stored in a memory, "illuminates Antoine Bordes.

The hope of Yann Le Cun is that artificial intelligence progresses in this way consistently in the years to come. If this is not the case, he fears a new winter for this discipline that has had many ups and downs since the fifties. "The challenge is to know if people will continue to believe in it, while the results still do not match what they dreamed," he wondered.

No comments:

Post a Comment

Technology Tech news