Man-machine game vs the second inning, AlphaGo falls to the blank area of 37 stunned world -- see left after lee se-dol, remote viewing and explanation of Redmond also shocked. The west only nine masters said, "I really don't know this is a good or a bad move." And at the same time, the weiqi association vice President of communications English commentator, said Chris and Locke, "this is a mistake."
This step lee se-dol test for about 20 minutes long, but more than four when I was a child, he still lost. Then later on, the man-machine war, link to the Google data centers around the world hundreds of distribution of the AI programs, defeat may be the human design of the most complex of the best players in the game.
Of 37 hands equally confused is probably fan, he not only had lost 5-0 to AlphaGo go champion of Europe, and from then on, he became AlphaGo practice partner. Lee se-dol war with more than five months, in front of the fan people with AlphaGo hundreds of sets. He looked at AlphaGo grew up day to day. Fan people more and more failure, but he is the most understand AlphaGo. Looking at AlphaGo 37 hands, he knew that there must be a ordinary people difficult to understand, calculate after 10 seconds, he said "was so wonderful hand".
Yes, most people think that AlphaGo victory is to calculate the strength of the rolling force, however, hand prove AlphaGo 37 not only the calculation, it shows a certain degree of understanding the ability to go, like a human player, has the sense of "chess". So, 37 hands of historic significance, it shows that the machine and human are finally starting to have a real fusion of the day.
AlphaGo founder, sabe was born in London in 1976, 4 years old he began to contact the international chess, 13 in the "master" (a chess master), ranked the world's second in the players under the age of 14. AlphaGo team leader silva said, "I have seen him in our town, to win the match, and then leave." The two of them officially met in Cambridge read undergraduate course. In order to understand the human mind, whether the machine can also become smart, two major is computational neuroscience. , sabe si at university college London (UCL) to study cognitive neuroscience PhD program, the main hippocampus of the brain involved in navigation, memory and imagination (hippocampus), to make thinking more approximate human computer basis, he put forward a new theory for science magazine 2007 top ten technology breakthrough.
In 1997 when IBM's deep blue computer beat chess champion, is precisely the sabe, studying computer science at the university of Cambridge. He was in contact for the first time in Cambridge to go - the one thousand - year - old board movement, and just getting into the game, sabe si couldn't help thinking: why the machine never break this intellectual games? Also because of this, sabe si decided, I want to make a play weiqi is better than a human computer system. In terms of game theory, the game like chess and checkers, is a game of perfect information game - no luck, full disclosure of information. Generally speaking, the computer should it can be easily mastered, but it is conquer not go.
Ha sabe, says that, in the game, whether human or machine can't work out every step of the final result, top players by intuition, rather than hard - that is, it is move feeling. "Go exquisite aesthetic layout, good layout look full of aesthetic feeling."
In 1998, two people goes into a computer games company after graduation. The game is a good way to test of artificial intelligence. But in 2005, their game companies went bankrupt. Silva, went to the university of alberta research primary form of artificial intelligence, enhance learning. Let the machine learning techniques, it can repeat the same task, find out the best decision, so as to realize autonomous learning. Hasa, go to the university college London, received a doctor's degree in the neuroscience. Two major is computational neuroscience, in order to study whether the machine can also become smart. In 2010, they meet again, sabe si in London has set up a company called DeepMind of artificial intelligence, silva joined him.
When Google CEO brin met, sabe si, ha sabe, said: "a few years, DeepMind may be able to defeat world go champion." Even with foresight outstanding brin also feel incredible, but they did it.
Man-machine war after the second inning, silva into AlphaGo control room, to monitor its operation is normal, and track it have any change on every game result of forecast. Silva, bring up the AlphaGo record decisions during the game, see AlphaGo next out of the hands of 37 what happened a moment before.
Before there were DeepMind and AlphaGo machine chess rely on brute force method, namely exhaustion, IBM's deep blue with this. At that time, the dark blue also out of the step of human unexpected, however, the violence cannot solve the go. Go there are too many changes, the computer can't hard to calculate.
So, DeepMind only path - machine learning.
DeepMind team will step 30 million humans go out under step input to a depth of neural network. The network simulation in the human brain neural network, the team also hope that it is able to think like a human brain, autonomous learning. Computer vision technology such as Facebook, Google voice recognition. Watch enough cats, it can recognize the cat; Enough language input data, it can understand natural language; Also, input enough, move it can also learn how to play chess. However, creative rules of lenovo and use are two different things, such as 37 hand does not step in the 30 million chip, then how to do it? AlphaGo also calculates that, in fact, a human expert player under such a probability is only about one over ten thousand of step, but it still chose this step.
"It know professional guy so the chance is very low, but when it after his own calculation, it can be the overthrow of the original input chip reference," silva explains, in a sense, AlphaGo began to independent thinking. It's decision not to its creator in its digital DNA into rules, on the basis of but its self-learning algorithm based.
Let itself after learned to play chess, silva AlphaGo and his game - a version with the different neural networks. In the process of self game training, AlphaGo records on those -- this is the optimal way to silva studied the enhancement of learning technology.
Give yourself j - this is an effective method for QiLi increase, but this is part of the skill. Analyse the situation, has the logic calculation is not enough, also found in the vast checkerboard hand by instinct, is the perceptual forecast according to move form. After enhanced learning technology, silva's team will these non-human go step input to the neural network, the second teaches it like dark blue chess chess as prediction go chess. With his game after more than all of the information collected after losing in AlphaGo start can predict a go game the way possible. This is intuition. Such as AlphaGo 37 hands. Even back to the background check process silva, cannot know how AlphaGo the results - which is the formation of move feeling.
AlphaGo DeepMind entered the field of AI is an important step, but for "AI replace human theory", ha sabe said don't have to worry about, in his view, the AI is a tool, a structure of wisdom, let human better tools. Although AlphaGo currently has the ability to do this, but it may not really know "own" in what to do. So, the use of a tool, ha sabe si imagine how the AI world in the next five years? Google paid $650 million to buy a company, not only play a chess game.
Have a deep learning and independent thinking ability, AlphaGo today can play chess, can learn the design tomorrow. More than a dozen deep learning and neural network to support the service provided by Google, including its omnipotent search engine. AlphaGo a less secret weapon - enhancing learning in teaching has the lab machine people picked up and move all kinds of items.
But the business problem is not the most important. When asked, sabe, see lee se-dol: how did you feel he lost the game, he said, pointing to his heart: "I'm sorry." See their own creation achievements he felt proud, but for human instinct, he felt sad and he hopes lee se-dol can win a game.
In the fourth, however, AlphaGo dominates on the board with the same mysterious technique control - a particular area of chess board when lee se-dol stormed, AlphaGo response reflects startling vision, can not only resist lee se-dol local strangulation, but also for all the stage -- in hand, 37 AlphaGo puts a black piece next to a only a white place, away from the Internet.
And then, under the AlphaGo 77th hand, lee se-dol into long test again. Lee se-dol walk 78 before hand, AlphaGo calculate their winning percentage is 70%. But at the same time, a AlphaGo became the next Garry kasparov. It can't believe that a human will bottom out the "one hand" god 78 - nearly one over ten thousand of its possibility. But in another Angle to think, this thinking more frightening: it is challenging the human fear, it has a fluky psychology. It has a bit of human thinking, there is also a weakness.
After 78 hand reversal, lee se-dol attack on the right, in AlphaGo designed for defense. In the process of it to keep for attack, AlphaGo once seem to want to be in the right "disaster", but did not succeed, but ran out of the "death". Then, after the dish every move later, AlphaGo are thinking more than 3 minutes on average. By "packaged" stage, lee se-dol still cautious, no mistakes. Finally, AlphaGo listing to throw in the towel, lee se-dol first win. This is the first time since AlphaGo beat people fan "surrender" to people.
Finally the fifth, though still ended in defeat for lee se-dol, but both sides making already put it out, it's a contest of principles. Before the start of man-machine war, sabe, referred to as "the AI is a tool for promoting human progress, then such groundless, but as the AI into real life and production application, the more the more let people believe that: the machine can also be done humanization, and can with human mutual coordination, promote each other. The most simple, no AlphaGo 37 hand, and to "one of god hand" 78?
