The news that "computer" outplays a professional Go player has spread through the Internet. A Korean Lee Se-dol, one of the best Go players in the world and owner of the ninth dan, has already lost two of five games to the AlphaGo program created by Google Deep Mind. For Lee Se-dol this was, apparently a big surprise, because he was sure that will cope with the program, watching her playing with Fanem Hui champion of Europe, held in October 2015 (at that time, incidentally, Hui also lost to AlphaGo, but Fanu Hui is far from Lee Se-dol level). Why it is yet impossible for one of the strongest players to win the game against the machine, as well as what is artificial intelligence (AI) and what other problems it can solve, the most important person in the company Yandex on Artificial Intelligence Alexander Kraynov told to reporters.
First, you need to understand on what basis artificial intelligence is running. "Artificial intelligence - it is a special case of machine learning", - said Alexander Kraynov. - But there is a certain line when a machine learning becomes artificial intelligence. "
Oddly enough, this does not come in contact with the area of logical thinking but is rather in the field of experience. When we solve any problem (decide), we rely on certain factors. These factors are inadequate. For example, when buying a garden we look at the price of the site, nature, neighbors, distance from the town, and many other "variables". But the main thing for us is, for example, the price - if the value of cottages is higher than we expected, we do not buy it.
So, when we do machine learning we ask relevant factors, and the machine selects the weight coefficients, that is, how important they are. But artificial intelligence determines on its own which factor is important - for each specific case. A so-called decision tree handle well with this task. This is an algorithm, where the machine must answer a certain question each step. Such questions specify the scope of the search.
Yandex being the browser should handle and correctly interpret user requests. For example, if you create a decision tree for the search engine the first question may be whether the searched word is in the fresh new. Then it searches if there is an appropriate material in "Wikipedia", and so on.
A neural network is a very large multi-layered structure in which each layer is made up of thousands neuronic points, analyzing the incoming information. Each 'neuron' has two states: "yes", this information is correct, and "no", this is not true. If the first layer of neurons is focused on self image, text, sound or information provided in other way, the deeper layers (by the way. Interface itself is called a deep neural network, and the process of learning is Deep learning) «looked» on the previous layer, and thereby way receive more complete information. And each "neuron" run through itself all the information in the search for matching his request.
To train a neural network to recognize some features (for example, pictures of a man among the landscape and interior images), it is necessary to provide access to a large number of pictures with a man, as well as to a no smaller pool of images without human. Sorting millions of images and receiving feedback, the program remembers the important features and starts the process anew using a new picture. So gradually it gains a certain "pool" of significant symptoms and easily can deal with exactly the same problems.
"The difference is that ealier we invented the factors ourselves but in the case of a neural network, we stopped doing that, we began to give examples - summed up Alexander Kraynov
The neural network learns about in the same way as children. We do not explain to the child that the cat is different from the form of doggie ears, whiskers, and so on. The child is shown in the book: here's a cat and here is a dog. How the child consolidates the acquired knowledge? He asks all members of the family million times wrong at first, but gets a confirmation or denial and thus reinforces his skills. As a result his neural network will develop signs on its own and using them it will be able to distinguish a cat from a dog, but we did not teach it these grounds. "
The artificial neural network produces an experience that turns into a kind of intuition, that is, the ability to take right decisions in an environment where it is very difficult to figure them out. That is why the Go game has been selected to test artificial intelligence where chess was rejected.
Go is a true product of Asian culture. Its essence is to "isolate on a gameboard more territory than the opponent using playing stones of a certain color " and when the size of the standard board is 19 × 19 cells and gaming chips of each player are equivalent there is a huge number of turns. Therefore players at every turn has to re-assess the situation. According to Alexander Kraynov, who himself is an ardent fan of the game, to figure a game out is simply impossible. And experts will explain this or that turn using such categories as "beautiful" or "creating balance", which are difficult to logically explain, but if you play a hundred games, you can tell the difference. Therefore, other programs with machine learning programs could not beat the man in Go.
AlphaGo program watched a great number of parties, moreover, it was set up so that the winners were perceived as "good", and the losers - as "bad." "Of course, not all the turns of good players are good, certainly the Kraynov said. - But on a large sample they ar on average better than losers turns ".
AlphaGo still continues to learn: watch other games and play with itself. Lee Se-dols fatal mistake was that he did not consider the program's ability to self-improvement. Therefore, in October, when the program beat the European champion, he alleged with a certainty that the computer will not be able to win the game against him. By the way. AlphaGo is a computer only conditionally: now the device playing against the Korean champion contains as many as 1202 processors. But be that as it may, in the past months the program has significantly improved the level. According to experts, the gap between strong and weak turns of is AlphaGo very small, and this is a sign of good player.