A year and a half ago, artificial intelligence succeeded in the name Alpha-Go Defeat Lee Sedol, one of the greatest human champions in the game Go. And not only that, but she beat him 4:1, while he (and the other experts in the field) didn't think he had a chance of losing. Now her older sister, Alpha-Go-Zero, has taught herself to play from scratch - without human involvement - and within three days of learning has already beaten the original software.
To understand the magnitude of the achievement, we must first talk about the original Alpha-Go. It is an artificial intelligence engine that acquired the basic skill of the game by feeding it 30 million different steps collected from 160,000 games. At this point the artificial intelligence had already reached an impressive level of play, but not yet one that was close to competing with the experts. But this is a new kind of artificial intelligence: one that is able to learn from observing games. That's why the researchers let the artificial intelligence play against itself, and learn and learn and learn from each game. She played a huge number of games against herself - more games than any human expert has played in a lifetime - and when it came to the decisive battle against Lee Sedol, she completely eliminated the hapless human player.
Actually, almost completely.
Alpha-Go played at a superhuman level. It revealed strategies that human players weren't even aware of. In the second game, she made such an unusual move, Shelley Sedol had to take a few minutes break just to relax from it. This step is called in the parlance of the Go players - "divine": creative, innovative and special. And she won that game, and was also sure of her victory from the middle of the game - although the watching experts were not sure who was leading all the way.
But Sedol was not a cane killer himself. He is considered a creative and unique go player, and in the fourth game he won back one victory for himself. He made a move of his own that the spectators also called "divine" - a move so strange and unusual that even Alpha-Go was not prepared for it, and he won the game in the end. One win, four losses, and two opponents - a person and a computer - each of whom demonstrated creativity and innovation.
Shortly after, Alpha-Go went on to further upgrade to Alpha-Go-Master defeating all the Asian Go Champions, one by one.
All this was a year and a half ago, and now there is a new version of Alpha-Go, which is even stronger than the previous one. How strong? The new version is called Alpha-Go-Zero, and when she played against the version that beat Lee Sedol just a year and a half ago, she crushed her in XNUMX games out of XNUMX. And not only that, but Alpha-Go-Zero programmed itself: it didn't even need the human programmers to feed it the first millions of steps. She was just playing against herself from the very beginning, according to the basic set of rules of the game, and in just three days she managed to reach the same level of Alpha-Go. Within forty days she also surpassed Alpha-Go-Master, becoming the best Go player in history. And all this, as mentioned, by self-study only, without human involvement or supervision.
These developments are consistent with Kurzweil's vision of the singularity, whereby machines will be able to learn by themselves and plan their own future generations. We are moving towards the realization of this vision at an increasing pace. The newest type of artificial intelligence can already teach itself without human involvement - at least in a well-defined environment, such as games. It is not an exaggeration to guess that in any situation that can be well modeled and artificial intelligence can be run on it, it will be able to provide options for optimization that will bypass everything that humans can think of. Already today, Google uses artificial intelligence to improve the energy consumption of its buildings, for example. Similar artificial intelligences will be able to develop more advanced chips and more complex means of calculation. And once we can model the human body at a high resolution level, for example, we can use advanced artificial intelligence to develop techniques and technologies to preserve human health.
But how will we keep the understandings we learn in our responsibilities? How can we make it clear to them that certain solutions to the games - for example, the total extinction of the human race to reduce the chance of future war - are not appropriate? These are questions that some of the greatest researchers in artificial intelligence are facing today, and are trying to understand how to give artificial intelligence "super values" that it cannot violate. What will those super values be? This is another question - but one that we all need to think about.
So go ahead - to the singularity. It's going to be exciting.
See more on the subject on the science website:
10 תגובות
rival
I agree with everything you wrote. What I am saying is that there is not necessarily a connection between all of this and the way the biological mind thinks. The fact that they both contain components called neurons means nothing.
Miracles,
I tried to understand what you wrote and answered you accordingly, if I didn't understand you correctly please explain again without quizzing me.
With all due respect to the application you wrote, I don't think it is related to the matter, I have also written nice programs that learn in all kinds of interesting ways.
Alpha Go is much closer in its principles of operation to the brain of a mouse than to your or my apps, of that you can be sure.
Regarding the worm, although we no longer have a complete understanding of this worm, we have nevertheless created neural networks that also have "self-awareness" and know how to look for food and stay away from dangers like your worm:
The https://www.youtube.com/watch?v=bIuu8URVF58
Again I'm trying to explain to you, the fact that we still don't fully understand how the neural network in our brain learns to play a computer game, doesn't mean we can't create neural networks that learn to play, the same goes for networks that know how to count or behave like a worm.
rival
Your last sentence emphasizes to me how much you don't understand what I'm trying to say. I ask that you read my comment again and refer to what I wrote.
Opponent - Build an amazing machine that plays Go at a high level. Part of this machine is a neural network. These neurons are a bit (really a bit) similar to the neurons in our brain. But, this network has nothing that even resembles self-awareness.
Think for a moment - a worm with about 300 neurons has self-consciousness. The worm knows it is hungry, knows how to look for a partner and stay away from certain dangers. Alpha-Go has no such ability.
Alpha-Go is a computer after all. Its working method is not fundamentally different from much simpler learning systems. I previously implemented software for an autopilot that learned by itself how to keep a gob and perform a coordinated maneuver. With the help of a different learning method, the software knew how to perform air combat and even defeated veteran fighter pilots True, it is orders of magnitude simpler than playing Go, but I don't think there is such a fundamental difference.
What's more - I'm sure that Alpha-Go is much, much closer to my software than to the brain of a mouse.
Miracles,
Playing Go is a task a thousand times more complicated than counting (almost every 5-year-old child knows how to count, but there are not many children of that age who know how to play Go, certainly not at a professional level) Do we already know how the mind of a professional Go player makes the moves that lead him to victory?
Of course not, but here we have an example of a neural network that taught itself how to play Go, and reached the level of a champion without seeing a single example of a game between humans! How do you explain this wonder miracles?
Note that you keep confusing the ability of a neural network to perform some task, and our ability to understand how it did it, one independent of the other.
And specifically regarding counting, a quick Google search shows that there are neural networks that know how to count.
rival
I guess you agree with me that simple counting is a basic and essential operation in the human brain. Do you think we know how the brain counts?
Miracles,
You can laugh until tomorrow but we know a lot about how the brain works and how to simulate it on a computer, the results as you can see in the article speak for themselves.
Even if they see a neural network that does everything a person does, you will continue to laugh that we still don't understand anything and that it won't happen soon, then you will continue to laugh and the field will continue to advance.
Soon, there will be two-legged robots that can run faster than Usain Bolt. Simply amazing.
But you have to remember - the robot's running action is much more similar to human running, than the "thinking" of this computer is to human thinking. The explanation is simple - we do not know how man thinks. We know little about how a neuron works, but not much beyond that.
simply incredible.
The next unofficial goal of artificial intelligence researchers is according to what is running in the AI corridors in artificial intelligence research groups is medicine with the help of artificial intelligence. It does not currently seem within quick reach. In my opinion, keeping the machines that will save the human race, will not work. We built it as our images, And we don't respect intelligences lower than ours. We kill them, eat them, certainly don't give them rights.
It is important to note that the software trains on 1.5 million scenarios before it wins a world champion, and that it receives human guidance during the training phase, otherwise it does not achieve good results. It will be fixed gradually. From 2012, a regularization of the intentions it receives is introduced. It is also important to note that the same intelligence that broke ground in 2012 was tested on at least 60
Different fields of specialization and excelled more than Adam. Before that, neural networks did not have a dizzying success even though the field started in 1969. What made the difference from 1969: the power of calculation. NVIDIA GPU card, and the regularization of the intentions that the algorithm receives in training. And a lot of faith from the first group, who broke the ground, who trained the algorithm for two months, without knowing that they would see something groundbreaking. They lived in an era where there was no software to program artificial intelligence algorithms and that group programmed manually as in assembler. Hard work on 1.4 million photos. A GPU card is designed for gamers and not for artificial intelligence. Today the same training lasts 3 hours and there are GOOGLE tools that are provided for free.