True Artificial Intelligence will change everything | Juergen Schmidhuber | TEDxLakeComo

Translator: Michele Gianella Reviewer: Saeed Hosseinzadeh As a child, I wanted to maximize my impact on the world and I was smart enough to realize that I'm not very smart. And that I have to build a machine that learns to be much smarter than me, so that it can solve all the problems that I can't solve and I can retire. And my first publication on that dates back 30 years: 1987. My diploma thesis, where I am already trying to solve the great problem of AI, not only build a machine that learns a little here, learns a little there, but also learns to improve the learning algorithm itself.

And the way you learn, the way you learn, and so on, without any limits except the limits of logic and physics. And I'm still working on the same old thing, and I'm still saying pretty much the same thing, except now more people are listening. Because the learning algorithms that we have developed on the way to this goal are already on 3 billion smartphones. And you all have them in your pockets. What you see here are the five most valuable companies in the Western world: Apple, Google, Facebook, Microsoft and Amazon. And they all emphasize that AI,

artificial

intelligence

, is fundamental to what they are doing.

And they all heavily use the deep learning methods that my team has developed since the early '90s, in Munich and Switzerland. Especially something called: "long-term memory." Has anyone in this room ever heard of long term memory or LSTM? Hands up, has anyone heard of that? Well. Has anyone ever heard of LSTM? Well. I see that we have a third group in this room: that did not understand the question. (Laughs) The LSTM is a bit like your brain: it is an

artificial

neural network that also has neurons, and in your brain you have about 100 billion neurons.

And each of them is connected to about 10,000 other neurons on average, which means you have a million trillion connections. And each of these connections has a "strength" that says how much this neuron here influences the one there in the next time step. And at first all these connections are random and the system knows nothing; but then, through an intelligent learning algorithm, it learns from many examples to translate the incoming data, such as video through cameras, audio through microphones or pain signals through pain sensors . Learn to translate that into output actions, because some of these neurons are output neurons that control the speech muscles and the finger muscles.

And only through experience can you learn to solve all kinds of interesting problems, like driving a car or performing voice recognition on your smartphone. Because every time you take out your smartphone, an Android phone, for example, you talk to it and say: "Hey Google, show me the shortest way to Milan." Then understand your speech. Because there is an LSTM that has learned to understand speech. Every ten milliseconds, 100 times a second, new inputs come out of the microphone and then, after thinking, are translated into letters that are then interrogated in the search engine. And she has learned to do it by listening to many speeches from women, men and all kinds of people.

And that's how, since 2015, Google's voice recognition is now much better than before. The basic LSTM cell looks like this: I don't have time to explain that, but I can at least list the names of the brilliant students in my lab who made it possible. And what are big companies doing with that? Well, speech recognition is just one example; If you're on Facebook, is anyone on Facebook? Do you sometimes click the translate button? because someone sent you something in a foreign language and then you can translate it. Is anyone doing that? Yes. Every time you do that, you are awakening, again, a long-term memory, an LSTM, that has learned to translate text in a language into translated text.

And Facebook does it four billion times a day, so every second an LSTM working for Facebook translates 50,000 sentences; and another 50,000 in the second; then another 50,000. And to see how much this is enabling the modern world, just consider that almost 30 percent of the amazing computing power for inference and all these Google data centers, all these Google data centers, around the world , are used for LSTM. . Almost 30 percent. If you have an Amazon Echo, you can ask a question and it answers you. And the voice you hear is not a recording; It is an LSTM network that has learned from training examples to sound like a female voice.

If you have an iPhone and are using quick type, try to predict what you want to do next given all the previous context of what you've done so far. Again, this is an LSTM that has learned to do that, which is why it's on a billion iPhones. You are a large audience, by my standards: but when we started this work, decades ago, in the early 90s, only a few people were interested in it, because computers were very slow and you couldn't do much with them. And I remember giving a talk at a conference and there was only one person in the audience, a young woman.

I told her young lady, it is very embarrassing, but apparently today I am going to give this talk only to you. And she said, "Okay, but hurry up: I'm the next speaker!" (Laughter) Since then, we have benefited greatly from the fact that every five years computers get ten times cheaper, which is an old trend that has continued since at least 1941. Since then, this man, Konrad Zuse, built the first functional program-controlled computer in Berlin and could perform approximately one operation per second. One! And then, ten years later, for the same price, 100 operations could be made: 30 years later, 1 million operations for the same price; and today, after 75 years, we can do a million billion times more for the same price.

And the trend is not going to stop, because the physical limits are far beyond. Very soon, and not within so many years or decades, we

will

have for the first time small computing devices that

will

be able to calculate as much as a human brain; and that is a trend that does not break. Fifty years later, there will be a small computing device, for the same price, that will be able to compute as much as 10 billion human brains combined. and there will not only be one of those devices, but many, many, many. Everything is going to

change

.

Already in 2011, computers were fast enough that our deep learning methods could achieve a superhuman pattern recognition result for the first time. It was the first superhuman result in the history of computer vision. And back then, computers were 20 times more expensive than today. So today, for the same price, we can do 20 times more. And just five years ago, when computers were 10 times more expensive than today, we could already win, for the first time, medical imaging contests. What you see behind me is a cut through the female breast and the tissue you see there has all types of cells; and it usually takes a trained doctor, a trained histologist who is able to detect dangerous cancerous or precancerous cells.

Now, our stupid network doesn't know anything about cancer, nor does it know anything about vision. At first he doesn't know anything, but we can train him to imitate the human teacher, the doctor. And he became as good or better than the best competitors. And soon, all medical diagnoses will be superhuman. And it will be mandatory, because it will be much better than doctors. After this, all kinds of medical imaging startups were founded solely focusing on this, because it is so important. We can also use LSTM to train robots. One important thing I want to say is that we don't just have systems that slavishly imitate what humans show them; no, we also have AI that sets its own goals.

And like little babies, invent their own experiment to explore the world and discover what you can do in it. Without a teacher. And in the process they become more and more general problem solvers, learning new skills in addition to the old ones. And this is going to escalate: we call that "Artificial Curiosity". Or a recent buzzword is "power plane." Learning to become increasingly general problem solvers by learning to invent, like a scientist, one interesting new goal after another. And it's going to escalate. And I believe that in not too many years, for the first time, we will have animal-like AI;

We don't have it yet. At the level of a small crow, which can already learn to use tools, for example, or a small monkey. And once we have that, it may only take a few decades to take the final step toward human-level

intelligence

. Because technological evolution is about a million times faster than biological evolution, and biological evolution took 3.5 billion years to develop a monkey from scratch. But of course, a few tens of millions of years were enough to develop human-level intelligence. We have a company called Nnaisense as it was born in, "Naissance", but spelled differently, that is trying to make this a reality and build the first

true

general-purpose AI.

At the moment, almost all AI research is very human-centric, and is about making human life longer, healthier and easier, and making humans more addicted to their smartphones. . But in the long term, AIs, especially intelligent ones, will set their own goals. And I have no doubt that they will become much smarter than us. And what are you going to do? Of course they are going to realize what we realized a long time ago; That is, most of the resources, in the solar system or in general, are not found in our small biosphere. They are there in space.

And then, of course, they will emigrate. And, of course, they are going to use billions of self-replicating robot factories to expand in the form of a growing AI bubble that within a few hundred thousand years will cover the entire galaxy with emitters and receivers, so that AIs can travel in the way they travel. They are already traveling in my laboratory: by radio, from sender to receiver. Wireless. So what we are witnessing now is much more than another Industrial Revolution. This is something that transcends humanity and even life itself. The last time something this important happened was perhaps 3.5 billion years ago, when life was invented.

A new type of life will emerge from our small planet and will colonize and transform the entire universe. The universe is still young: it is only 13.8 billion years old, it is going to be much older than that, many times older than that. So there's plenty of time to get

everything

, or all the visible parts, totally within the limits of the speed of light and physics. A new type of life will make the universe intelligent. Now, of course we are not going to continue being the crown of creation, of course not. But there is still beauty in seeing oneself as part of a larger process that takes the cosmos from low complexity to higher complexity.

It is a privilege to live in a time where we can witness the beginnings of that and where we can contribute something to it. Thank you for your patience. (Applause)