Researchers are already confronted with the limitations of the techniques that have made artificial intelligence so successful in the recent past. New paths are being explored to get a little closer to human intelligence.
Autonomous cars, strategy games, facial recognition, language translations... Artificial intelligence has made enormous progress over the last ten years.
"Today, we can distinguish two branches: supervised learning, where the machine learns from many examples, and reinforcement learning, where the machine will be rewarded if it gives a good answer", Yann Le Cun, Chief Scientist at Facebook, explains in French in an interview.
The first one has particularly sparked off in image recognition, the second one has established itself in the world of games. But these techniques are already beginning to show their limitations...
Thus, deep learning is based on the availability of a large amount of labeled data. This is a good thing : the multiplication of sensors makes it possible to generate a lot of data for certain problems. But there are cases where data are scarce or difficult to generate. Conversely, if the data are really very large, labeling becomes complicated.
The path of learning by reinforcement also has its limits. To achieve good results, this technique requires a very large number of trial/error cycles.
The AlphaStar team explains on DeepMind that Google DeepMind's work on the StarCraft game shows that it takes the equivalent of 200 years to reach the level of a human player.
Obviously, these calculations can be accelerated in the case of a video game. But for real-world applications, such as robotics, this is not possible.
This is why the French researcher, Yann Le Cun, has been promoting the principle of Self Supervised Learning (SSL) for about five years. According to him, it could lead the IT sector - and Facebook in particular - towards a future "artificial intelligence revolution".
The idea is to take a mass of data and hide part of it. The SSL algorithm is responsible for analyzing the visible part and deducing the hidden part. As a result, he will create the labels himself that will allow him to complete his learning.
A translator who has never seen translations
This technique has already achieved some success in the field of languages. By analyzing the context of words in texts, it makes it possible to transform a linguistic corpus into a vector field.
However, these vector fields can then be easily correlated with each other, regardless of the underlying languages. This allows you to create a translation service between two languages without having to rely on translated texts, which is particularly useful for rare or regional languages.
These vector representations can also be used to predict the words that will arrive in a conversation, a very useful feature for text editors in messaging, for example.
Facebook researchers are now trying to apply SSL to other areas, such as speech recognition or image recognition.
But it is not that simple because of the complexity. The words of a language constitute a finite set of elements, which is not the case for sound or image. Predicting a sound or image in a sequence remains a challenge.
"For example, you take a pen, put it upright on a table, drop it, it will fall. You can't predict in which direction though. And it's very difficult to train a learning system when there's not a single correct answer. We have to tell him "here are all the correct answers", and we don't know how to do it very well. This is the focus of the research: predicting high-dimensional data - images, videos - in the presence of uncertainty", Yann Le Cun explains in an interview in French.
Of course, Facebook is not the only one exploring the path of SSL. Google and other IT giants are also looking at this technique. It can only be a step towards truly advanced artificial intelligence systems.
SSL is a little closer to the human learning mode, which is very much based on observation. But there is still a long way to go. Based on statistical calculations, supervised, self-supervised or reinforced systems do not have a sense of causality that would allow them to deduce the facts of an action (example: a vase falls to the ground, so it breaks).
This is a reasoning that any animal is capable of doing, but that is inaccessible to a machine.
Similarly, learning algorithms cannot conceptualize things. Fuelled by traffic images, a computer will quickly identify cars on a road. But it will have trouble recognizing a car immersed in a pool.
"We can train a machine to recognize images, it will only do that. If we train her to play the go game, she will only play go... If we change the size of the board, we must train the machine again. There is no true abstraction from one skill to another. We can work on developing machines that detect tumours in medical images, that drive cars... But it's very narrow. What we do not know how to do is machines that learn by themselves by observing the world and acquire a certain common sense in the same way as animals and humans", Yann Le Cun explains in French.
It is also a bit for this reason that it is possible to trick an automatic learning system with processes that a human would have no problem detecting.
Sources : Le Monde, Sciences&Avenir, Konbini and DeepMind
I've made a lot of articles with tools, explanations and advises to show you how to protect your privacy and to secure your computer, GO check them out!
This is my guide To Secure your PC after a fresh installation of Windows
If you think that your Phone or your PC has been hacked, you have to check it right now!
That's how you can be more Anonymous on the internet!
The Future of Cyber-Security, what to expect?
The best Crypto debit card – Wirex!
These are the best VPN to protect your numeric life: NordVPN, ExpressVPN and CyberGhost!
Your PC is slow? That's why!
Why is it important to Be Discreet on the Internet
Feel hot? Your Computer also!
How an Adware works?
That's how you should guard against Trojan!
What are the different Types of hackers?
