Neil Zhu，ID Not_GOD，University AI 创始人 & Chief Scientist，致力于推进世界人工智能化进程。制定并实施 UAI 中长期增长战略和目标，带领团队快速成长为人工智能领域最专业的力量。
作为行业领导者，他和UAI一起在2014年创建了TASA（中国最早的人工智能社团）, DL Center（深度学习知识中心全球价值网络），AI growth（行业智库培训）等，为中国的人工智能人才建设输送了大量的血液和养分。此外，他还参与或者举办过各类国际性的人工智能峰会和活动，产生了巨大的影响力，书写了60万字的人工智能精品技术内容，生产翻译了全球第一本深度学习入门书《神经网络与深度学习》，生产的内容被大量的专业垂直公众号和媒体转载与连载。曾经受邀为国内顶尖大学制定人工智能学习规划和教授人工智能前沿课程，均受学生和老师好评。
Tommorrow, DeepMind AlphaGo, an AI Go player program from Google, and Lee Se-dol, South Korean professional Go player of 9-dan rank, will give the world an important game. -- This is the first time, AI player challenges the top human player in Go game which has been recognised as the most complex game in human history ever. Almosthuman got the chance to interview Hui Fan, during the interview, He recollected the game with AlphaGo, commented on AI technology from his perspective, and then gave the comparison between Go and life. Hui Fan will play the role of judge for the "AlphaGo VS Lee Se-dol" game. In this interview, we hope the deep and extensive thinking of Hui Fan can give a nice introduction of this "Century Fight", and more importantly, we believe this long piece of knowledge and experience sharing could inspire us beyond just a single game. It's about the relationships between machine and human, between technology and culture, between Go and philosophy, connecting our history and future.
Play with AlphaGo
Almosthuman: What is the opportunity of u engaging in playing with AlphaGo? Any interesting story to share with us?
Hui Fan: Last August, I went to a small town in Czech for European Go Championship or Congress (EGC). After EGC, I travelled to Bordeaux. And about the end of August or the beginning of September, I received Google DeepMind's email, in which they invited me to take a tour of DeepMind. At that time, I didn't even understand what they were doing. But I am a person of open characteristic, so I agreed. During the communication in Skype, I knew they belonged to Google, conducting AI research. They mentioned that they were researching an exciting project related to me, but then didn't tell me the true goal, skipping all the concrete things, including AlphaGo. I signed a confidentiality agreement with them, and the game with AlphaGo naturally took place. Due to this agreement, I can't tell some details, but I guess here the reason they contacted me is that I take the European Go champion in recent 3 years. If they want to have a test player, they need a professional one, with a title, so I was the "natural" selection.
The first play with AlphaGo was in October, 2015. I knew the result, knew the power of AlphaGo. But then other people still have no idea about this, I can't talk to anybody this result. During this period, there were two interesting things. The first is that at that time I went back to China, watched a Go program competition, I clearly knew that Go system was not in the same level with AlphaGo, but I can't say anything; second thing is in November 2015, there was an amateur Go player of 7-dan rank from South Korea, who was also top player in Europe. He said that he already played with Crazystone, he winned easily in the case of 让 3 子. Therefore he concluded that AI Go player was still long path to win human player. Only I can bear the funny situation. Not only this player, but others considered that Go programs need another 10 years to achieve the state of winning top human players. But actually, at that time, AlphaGo already beated me.
Almosthuman: At first, you just took a tour to DeepMind, then you knew they want you have a game with AlphaGo. Did you worry about anything? Did you consider, for example, what happened if I lose this game? And even would like to refuse the fight?
Hui Fan: No body can guarantee beat others everytime. Absolutely I consider the situation that I lose the game. However at that time the reason I received the Gaming invitation is that at first, I thought I could win the game with high probability; second, what if I lose the game? there were no decisive influence on myself. I am not afraid of losing a game because Go playing is full of losing or winning, no matter who your opponent is. But I will do my best to play the game, it is better if I can win, but losing also was receivable.
When I am young, my father told me many similar ideas, for example the report about Lee Chang-ho. He also worried about losing, but his idea is simple -- peacefully go each step. So my view on the game with AlphaGo was that I didn't care about the result of it. Actually all these years, I held this view during my games in Europe. Therefore it was a normal thing to me. The real situation was that DeepMind was very cautious then, they worrried that I may refuse the game. Since I am a professional Go player, if I lose, it seems that my honor will be gone. I told them to forget about this kind of worrying stuff. To win or to lose is normal for Go players.
Almosthuman: But this is not a simple Go game. What is your inner state during the playing of 5 games? Are there strong pressure? or more relaxed, an normal inner state for this Machine-Human game?
Hui Fan: As a professional Go player, when I engage in a match with win/loss, I absolutely try my best to win it. Recently, I heard many views, such as the view I lose game with intent, etc. I don't want to comment these things, but one thing it is valuable to point out is that in first game, I underestimated AlphaGo. My consideration was that my opponent was an AI player and we need to play 5 games, therefore I tried to start the first game with a simple strategy, then search for new things, if I can win the first game easily, then I knew how to suppress AlphaGo; if I get in trouble, then I could switch back. Moreover, I think the global view in Go game is important, evaluation of the plate state is the hardest task, so I start with a easy-to-sense strategy, AlphaGo should not identify with little effort. This was my consideration at that time. But the fact is when game was going, many moves were surprising to me. At least no computer-wise move occurred and the evaluation of situation was excellent. After I take a wrong move, I had no chance. Even before I made this wrong move, I thought I should win the game. But after the mistake, all done. And the controlling of the endgame phase was amazing.
The mistake in the first game is that there was "手筋", I missed it, that caused the big loss, but still have chance to win. That "勺子" is not hard to notice, but at that time I thought I was in a good situation, considering I will win the game, therefore neglect the "勺子". But after it, I knew that I would lose the game and I had no chance to win from then on. This was the first game, I payed little attention, but it caused DeepMind's big interest.
After the first game, I was in a bad mood. Why was that? The reason was -- This was the first time that computer beated professional Go player in history, this was a historic moment, never happened before. And I lost completely. It is not about the losing of game, but the feeling of AlphaGo's impossible to be beaten. As the first mistake occurred, I got no chance to win back. After that, my inner mood state was bad, second game, I switched the strategy to design some complex change to game with AlphaGo. The first 定式 is "大雪崩", one of the most complex transformation in Go. That time, I caught some good chance, but the problem was later on I made a "勺子" again, and was caught by AlphaGo. The catching is very accurate, once I had been caught, there was no chance to escape. And as I made a mistake, the going of the game was back to AlphaGo's way. Then I won't win. Each of the last games all went under the same way described above. It make no obvious mistake.
Almosthuman: When you are in the best stage, based on your estimation of the global game, if you do the right thing, will it be a hard game for AlphaGo in that game?
Hui Fan: Very tough. But the biggest problem is that human all make mistake, machine are more accurate. Therefore afterwards this is a burden for me. No matter the situation is good or bad for me, I was worrying about makeing a mistake. But AlphaGo didn't have the kind of conception. This case made me lose the inner balance. When I was in good situation, I didn't believe I can control the going of the game; when I was in bad situation, I thought I must lose. So all these things influenced my playing. Why do people say that my playing is not good? Because for a professsional player experienced lots of games, playing during game is not coincident to your real ability, there is a mentality problem. Take Lee Chang-ho for example, who beated many best players before, also was called "Stone Buddha" due to his good mantality, no facial expression, therefore no psychological fluctuation. But what I want to point out here is that AlphaGo is more "Stone Buddha"-tive than human, when they meet, the true Stone Buddha should be the machine. As a result, machine is hard to play with because of the nonlive feeling. It is like a void, which can control your characteristics, totally imbalanced.
Almosthuman: in the game is there such case that actually you made a hard-to-find small mistake, AlphaGo missed it.
Hui Fan: Yes. But I don't want to say that it didn't find the mistake, actually it didn't want to catch me in the mistake. Because it is not necessary to catch me there for a win. In the third game, there is a situtation I am dead. It was easy to eat me. But it didn't, and let me live. If AlphaGo chose to eat me, there would be a possiblity that the game would go complicated. And if it chose not to eat me, then it can easily win at last. Therefore AlphaGo took a more safer strategy to gain a easy win. There were also some Go report reviewed our game, one professional Go player also considered that at that time, AlphaGo had the idea about how to kill me, but it didn't. Because in real games, it left me alive then several steps later I admitted defeat. If it chose to kill me, the situation would be more complicated. Like Jie Ke said, he can not distinguish which is machine, which is human. Nobody can.
Almosthuman: from your point of view, are these small mistakes seriours? After 5 games, which mistake are more serious? How many times they occurred?
Hui Fan: The confidentiality agreement makes me not to give comments these contents. But you can ask professional Go players to analyze the question.