It will be time for AI to learn by itself without human supervision

To design this deep self-study mode, researchers have relied on human and animal intelligence.

Nowadays, everyone talks a lot about them because they are more sophisticated and popular. However, they are still machines anyway, basically not comparable to humans, so it is normal for them to occasionally make mistakes. Therefore, perfecting AI is always an urgent issue in the scientific world.

Recently, at the end of June, 2014, at the RFIAP Science Congress (Reconnaissance de Formes, Image, Apprentissage et Perception - Identity, Image, Learning and Awareness) were combined with two other seminars in Marne -la-Vallée near Paris, Yann LeCun, Facebook's AI research director, revealed to the press about future plans for AI.

Yann LeCun - Facebook's AI research director.

Yann LeCun has been one of the pioneers of Deep Learning since the 1980s. He said that in order to develop AI to achieve perfection techniques, we need to respond to a challenge. new consciousness, it is self-study technique without supervision.

According to him, Deep Learning is a technique that 15 years ago was not trusted by people, even many magazines refused to publish works on this topic. For a long time, Yann LeCun had a hard time convincing people to believe in Deep Learning. And it can be said that this technique is like a complete "auditorium" already packaged.

In addition, due to the appropriate convenient conditions, only a few more years, Deep Learning will make extraordinary steps. For example, the error rate of the image recognition algorithm on the ImageNet database has dropped from about 28% in 2010 to less than 3% in 2016.

The fact that Facebook has cloned the artificial neural layers is a miracle, and this result has made Deep Learning a step forward. Yann LeCun also said that for Facebook, setting up 100 artificial neural layers is only normal. Therefore, it is possible to meet up to 2 billion photos posted by users online every day, each image is taken through 10 artificial neural networks and displayed online in less than 3 seconds.

As a result, we can recognize our friends in photos or provide personalized content online, which is more than a social network. He added: "Soon we will see an electronic circuit dedicated to Deep Learning on each device, from mobile phones, to cameras and other connected devices" , this electronic circuit. will calculate itself in real time without having to go through the clouds.

However, despite some success, the AI ​​must reinvent itself to overcome certain weaknesses. Just like the problem of self-driving cars is being concerned because they still cause accidents and lead to death. Don't forget that in real life, any mistake can cause an accident.

In order for self-study not to be effectively monitored, we must combine learning with deep reasoning.

A complimented AI when properly identified is not as good as identifying itself in the surroundings, just like the way AlphaGo is trained to approach the game well.

To design this deep self-study mode, researchers have relied on human and animal intelligence. Particularly as humans, when making thousands of images of mammals, immediately know what is a dog or a cat, thanks to self-study without supervision. Besides being a simulation of the human world to equip the world's model machines, only that takes a longer time.

If a person learns to drive within 20 hours, an AI must take thousands of hours. In order for self-study not to be effectively monitored, we must combine learning with deep reasoning. This is also the research topic of both Facebook and DeepMind of Google.