We hear a lot about artificial intelligence, but concretely how does it work? Is the machine capable of thinking? Are we in the future all become sex slaves machines become crazy? Yes, today we will answer all these burning questions. By cons you leave me barely 10 minutes to make you understand this topic, it's not much. I will talk about machine learning and deep learning, no time for the rest. So if you are a fan of artificial intelligence that does not accept any simplification of concepts: getaway straight away it's better.
Artificial intelligence is already there. It is everywhere. When you upload an image, when you listen to music, when you do a Google search or when you take your shower then it will be there too. It analyzes, it studies. You send him a maximum of data permanently in the face to accelerate the pace. It already knows you very well. It's been years that it decides what to see on your Facebook thread. The worst is that it does it well since you do not realize it anymore. The problem is that you do not know much about artificial intelligence. Would not it be cool to know?
How it works before being dominated?
To understand, we will start by talking about machine learning. Concretely it is about complex algorithms, but complex kind of headache instant, who will learn to the machine by experience. We send them a maximum of data so that they recognize a pattern. The Pattern that they will use to predict a result is without the intervention of the man. The more real-world examples they will have to solve - the famous data - the more they will be effective because they will be able to refine the pattern. It's not clear my case? Let's look at an example of an alcoholic to better understand.
The most advanced of you have recognized at first sight the “Ballmer Peak " of our friend. For the team first degree yes the data is bogus and corresponds to nothing. It's a joke actually. By cons it allows me to illustrate how the machine will extrapolate a pattern (the curve) with a set of data (the table) to finally make a prediction (3.73). A learning phase is to create the curve and a prediction phase to use it. And now, roughly machine learning is that. Of course, here it is simplified to the extreme. I make you a curve with two variables like a moron and I'm happy. In real life, you want to do more complicated things. For example, to recognize faces in real-time on a video. To do that, you'll need something more powerful than my stupid curve.
One of the solutions to create artificial intelligence capable of facial recognition is deep learning. Deep learning uses the concept of artificial neurons. As the name suggests it is an algorithmic imitation of a real neuron of your brain. The high-level operation of a real neuron is the following: he receives a nerve impulse (electrical signal), he analyzes it, the module and sends it to the following neurons in his network. In an imprecise and incomplete way, it is this behavior that we try to reproduce with an artificial neuron.
We mimic the steps of this behavior via mathematical functions. Specifically, at the input of an artificial neuron, there are several variables to which coefficients (or weights) are assigned. Via a combination function, we obtain a sum of these variables taking into account the coefficients. This sum finally goes into an activation function that will decide the output value of the neuron. This activation function will use a threshold to do its calculation. If the sum of the variables is below the threshold, the neuron is said to be non-active and its output is 0 or -1. If on the contrary, the sum is above then the neuron is said active and its output is 1. It is complicated; look at the serious scheme pumped on Wikipedia to understand what I just said.
In summary, an artificial neuron will transform several input variables into an output variable through several coefficients and a threshold using mathematical functions. So we will be able to feed it with a maximum of input data and each neuron will trigger other neurons attached to it. The goal of the network will be to reach a certain value at the end of the race. If this value is not good then the neural network will adapt.
What is interesting to understand is that the values of the weights and thresholds of each neuron are adjustable. This is precisely what artificial intelligence will do in its network of neurons each time it has data. It will move the sliders until it finds a combination that works! And here we come back to my example of debilitated from the beginning with the two variables and the curve. The adjustment of weights and thresholds corresponds to the learning phase to create a pattern. And the more data we have, the more the neural network can learn by perfecting its pattern. It's infinitely more complex in a network of neurons, but the principle remains the same. The only problem, especially in the deep learning category of machine learning, you need a huge amount of data for the learning phase. However, the data is not what is missing.
Every day, 300 million images are uploaded to Facebook's virtual neural networks. Do you use Google Photo? Artificial intelligence analyzes each photo you take and trains to recognize the face of all your contacts. It trains the same with the contacts of these 500 million users. You do not have to be an Internet giant to access a phenomenal amount of data. Several free sites provide thousands of gigabytes of usable data for machine learning. Whenever you listen to music on Spotify you send data on their neural networks. I could go on like this for a long time. What we have to understand is that we feed these networks like big pigs and we do it at a crazy speed.
And this system delivers spectacular results. DeepMind, an artificial intelligence company bought by Google, has managed to teach the machine to walk. There's a very funny video that proves it and it's funny as well as impressive. Also in China, a television channel has outright replaced the presenter of the JT by artificial intelligence because why not? This is no longer science fiction, your voice can be imitated by artificial intelligence. We are already beaten to all the possible games of chess at GO through video games like DOTA or Starcraft artificial intelligence dominates us. In all areas you can think of, artificial intelligence is already used. So, we are entitled to ask the question of the limits of all that.
Must be afraid of how?
Are we all going to become mere batteries for machines that have enslaved humanity? Is it hard to say? One thing is certain: it's not for now. The reason is simple: the machine does not think. The machine solves a problem. It solves a problem quickly and well, much better than us. But it only does that. It executes code and learns to exploit it better with data. It is a tool. It does not have an awareness of herself. According to the experts, the next milestone to reach is intuition. But a machine with real human behavior seems far away. Awareness, for example, we have no idea how it works. How could we convey a concept that we do not understand?
The only thing that worried me while doing my research on the subject is Elon Musk. He explained several times that artificial intelligence is much more advanced than we think and that it would evolve today at an exponential speed, out of sight. These claims remain to be proven, but it would make sense that the state of progress of the last artificial intelligence remains discreet. It really represents a lot of money.
For the future of artificial intelligence, we must turn to the giants of the internet. They have the best developers in the world who have access to the greatest computing capacity humanly possible and all that supported by maximum money. Revolutions, good or bad, will come from home. However today there is no reason to worry about artificial intelligence, on the contrary, it helps us enormously and will continue to do it better and better. On the other hand, the future remains a real mystery.
Nov 11, 2019