One thing is for sure, the ultimate goal of Artificial Intelligence is to do one day all what the human brain is able to do. The thing is this is not science fiction anymore, we're entering the age of artificial intelligence.
Machine Learning
Ever since we started making computers, we tried to make them intelligent but we had no success doing this. We had no idea what intelligence meant, we had no data to do that and we had no computer power which was very expensive at that time.
Now we do have all this and in addition to that, we have a rough idea of what intelligence is, so we can make computers smarter now.
A very easy way to understand machine learning is that this is the total reverse of traditional computing. In traditional programming, every software, every computer programs were programmed with rules. A developer creates a set of rules that tell the machine what to do in a specific situation. It’s a lot like a series of complex behavioral guidelines that preempt various possibilities. The problem with this is that the program relies on the developer’s ability to accurately predict a series of future situations. While this is useful, it’s no longer immersive enough. And this limits the potential outcomes. as it's following the rules.
But we all know that intelligence is not about executing rules, this is about discovering the rules, changing the rules, understanding the world, abstracting concepts. Well this is exactly how we create machine learning today. Instead of programming rules, we feed computers algorithms with data and we let them discovering the rules.
In fact, this is pretty much like a baby who doesn't know how to speak, he doesn't know the rules of grammar when he’s born. He's just surrounded by people talking to him and then implicitly understands the rules of grammar. And with machine learning, we feed algorithms with the data and we let them discover concepts, rules, to understand the world. With this new ability given to the machine, Artificial Intelligence was born!
Deep Learning
What happened a few years ago is that we had a revolution in artificial intelligence because we used algorithm that can use a lot of data. They start from the simplest concepts in the pictures, which is the pixel, then they link those little concepts, pixels, to abstract more and more complex concepts in the pictures so that we can recognize what's in pictures today. These algorithms are called deep learning and they use what we call neural networks. Neural networks are pretty much like what we have in our brains that are capable of grasping pictures and make sense of it.
This is actually what you're using on Facebook. When you tag someone in Facebook, you're teaching the algorithm who's on the picture. And the more picture of you Facebook has with tags, the more it's able to recognize you on other pictures. It works with pictures on Facebook but it also works with medical images. We can teach computers today to recognize cancers on medical images. First, we feed cancer images to the algorithms, then we tell the computer what type of cancer is in the pictures and then the algorithms can recognize what kind of cancer is in new medical images. It works so well that it reaches human level at some task and then goes beyond human level.
Also, we could developed an algorithm feed with pictures of brains with and without Alzheimer. Now the computer can tell with your brain picture whether or not you’ll develop Alzheimer in six years from now.
So you understand the whole logic, we don't program computers anymore. We teach them, we train them like kids to understand the world, understand concepts and reach goals. With this new machine ability to see the world, Artificial Intelligence can see!
Reinforcement learning
Now that computers can see, we’ll give them another feature by providing them a visual field. Everything we do in a screen, your smartphone for instance, learning algorithms can see it, so we’ll let them understand what they see and try to reach goals in this environment.
And to make sure we could do that, at the beginning, we used video games. So we gave computers simple video games like Super Mario or racing games, and we gave them the ability to play, to go left and right, to jump, to accelerate. And then we let the computer, just by trials and errors, by millions of trials and errors, try and play the games.
At the first trials, the computer was really horrible at playing Mario. It couldn’t move Mario, it would fall into traps etc. But the more the computer played all by itself, the more the algorithm calibrated itself, until Mario became excellent! The logic behind it is that when the algorithm is closer to reaching its goal, we reward it, we strengthen it, we reinforce it. And when it moves away from the goal, we punish it, we prevent it from continuing this way and this is how the computer can now play video games in a screen. The funny thing, if you look at the results, beyond the fact Mario became excellent, is that Mario plays in a very weird way. And when you analyze the results more deeply and see how weird Super Mario plays, this is because the computer found out the optimal way to play Super Mario.
And this is the same that was happening with Google’s artificial intelligence AlphaGo in 2016 with the game Go. Google used the ability of computers to recognize their environments so they fed the algorithms with the games of the best Go players in the world. Then they train AlphaGo against itself, and using this technique AlphaGo managed to beat the best Go player in the world, live worldwide. And again, like with Super Mario, the best go players watching the game as well as the Google scientists working on the AlphaGo were not able to understand why the machine was playing in such a way. The machine discovered a new way to play Go. It is fascinating because Go is the game that’s been played for centuries, for thousands of years, and no one among human beings had the idea to play this way! Remember what we mention before, computer reaches human level and goes beyond it.
Same thing happened with Google cars, people were scared to ride in the Google car because it was driving in a weird way. And this is because it actually discovered a new and more optimal way to drive. So now, the Google engineers are "untraining" the car's AI from driving like a machine to train it more like a human being. We are not the best at doing things, and when computers decide to do things differently and very optimal, we’re kind of surprised.
If we’re using algorithms to recognize the environment and do things in the visual field of our screen. We could have an algorithm running in the background of our device that will be watching what we do in our screen, when we do a search, move the mouse, launching a browser, typing keywords, all this in order to teach him performing such tasks on our behalf. In fact, this is already what you and many people are doing by sending your GPS location to Google Maps so you’re teaching it the best route to take. With this new machine ability to learn to reach a goal, Artificial Intelligence can reach a goal!
AI is the umbrella
You should have now an easy-to understand definitions and explanations of machine, deep and reinforcement learning along with real world applications. Those are various cutting-edge technologies that all fall under the umbrella of artificial intelligence. And there are many others currently being developed that also fall into AI.
I wanted to conclude saying that the revolution we’re currently experiencing is actually enrolling everywhere in every industry. As Google's CEO, Sundar Pichai, said, we are switching from a Mobile-first world, that was mainly driven by social networks and mobile phones for about 10 years, to an AI-first world. Internet giants are putting artificial intelligence everywhere today and every industry is getting results with artificial intelligence.
AI Century