Lee: What does AlphaGo mean if he beats the world championship?

Innovation Factory chairman Lee Kai-fu in the Alphago and Li Shishi of the man-machine war, he said that four months ago Alphago defeated Li Shishi basically impossible, but this four months alphago progress a lot, the game should be very exciting. But whatever the outcome, the machine will surely triumph over mankind within 1-2 years. After the victory of mankind? Is it possible to make a universal brain? Does that mean the machine can think? What's the problem with machines that can't go beyond humans?

The following is Kai-fu Lee "Alphago can conquer Li Shishi?" The answer is in the original:

Answer this question directly, and then analyze the future of Alphago and AI. I think Alphago this game to defeat Li Shishi, but within 1-2 years must win over the human race.

By the Elo of both (go grade), it can be calculated that the Alphago defeated Li Shishi at the end of last year is quite low. How is it calculated? Alphago the top distributed version of ELO at the end of last year is 3168 (see the first picture below), while Li Shishi's ELO is about 3532 (global go hand elo:go ratings, see the second picture below).

According to these two grades of two chess players,

Li Shishi the odds per disk is 89% (see formula: How to guide:converting Elo Differences to Winning probabilities:chess link address: https://www.reddit.com/r /chess/comments/2y6ezm/how_to_guide_converting_elo_differences_to/tips: To access this link to be copied and pasted into the browser open, the same below). If you play a game, Alphago still have a 11% chance of winning, and the whole match five wins three or more, Alphago only 1.1% of the possibility. (Of course, according to the original October Alphago, only 1.1%, but now that greatly improved, it is not the same, perhaps today has surpassed: see 3rd below).

Didn't alphago defeat the European Championship? Some people think that Alphago defeated the European champion 樊麾 at the end of last year, so there should be hope for a challenge (former) world champion. However, 樊麾 is only two segments (ELO 3000 or so), while Li Shishi is a career nine (Elo 3532). The two-bit difference is huge and cannot be confused at all. For example, a person table tennis defeated the African championship, does not mean that he can successfully challenge the Chinese championship.

Is it possible for Alphago to beat Li Shishi by leaps and bounds in the past few months? "The outside world doesn't know we've made a lot of progress in the past few months," said Alphago's head. (From: http://www.geekwire.com/2016/alphago-lee-sedol-whos-underdog-in-google-ai-million-go-match/). This is indeed possible. There are two ways to Alphago progress: (1) Add Hardware: we can see from the Nature article that the ELO from 1202 CPUs to 1920 Cpu,alphago increases by only 28, and linearly increases the CPU without seeing linear ELO growth. To achieve an increase in 364 ELO integrals, the required CPU will be astronomical (an article estimated at least 100,000 cpu:http://www.milesbrundage.com/blog-posts/alphago-and-ai-progress )。 Of course, Google has the money to have machines, but purely machine will encounter the bottleneck of parallel Computing co-ordination (that is, assuming there are 1 billion machines, their total computing power is strong, but the coordination of each other will become a bottleneck). It should not be easy to increase the number of CPUs in a few months and adjust the algorithm to reduce the bottleneck by two. (2) To increase the learning function: Alphago has two learning functions, the first is based on the master game of Learning, the second is self-chess, self-learning. The former has used 160,000 master competitions, and the latter have been trained on giant units for 8 days. There will certainly be progress, but it may not be easy to go beyond the world championship. Finally, in a different way: if the "growth process" of deep Blue Beat the world champion in the past, dark blue reached the professional master level in 1993, 4 years later in a six-game tournament to defeat the world champion (about 500Elo points of Ascension). Today's alphago should be similar to the 1993 navy, just into the professional master level. It may not take 4 years to beat the world championship, but months don't seem to be enough.

Are there any other factors that have not been considered, leading to Alphago winning? If Google deliberately did not make the effort to 樊麾 confrontation, or have other learning or parallel computing beyond the nature inside the description, that Alphago is entirely likely to win.

(with the latest news: Li Shishi expected to win, not 5﹣0 is 4﹣1, his goal is to reach 5:0, a plate will not lose. Alphago in charge of the 50% probability of defeating Li Shishi, because the last four months of progress is very big. ）

Since we have written so much, I would like to make some comments on this topic:

What is AlphaGo? In January of this year nature (http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html) had a detailed introduction of Alphago, Alphago is a well-designed deep learning engine for go optimization, using the neural network plus MCTS (Monte Carlo tree Search), and using huge Google Cloud computing resources, combined with CPU+GPU, plus the ability to learn from Master games and self-learning. This system has improved the ELO by nearly 1000 points compared to the previous go system, from the amateur 5 segment to the level that can beat the 2 segment of the profession, surpassing the previous predictions of the Weiqi field, and reaching a significant milestone in the field of AI.

Is AlphaGo a breakthrough in scientific innovation? Alphago is a well-designed engineering excellence, but also to the historic industry milestones, but Nature article does not have a new "invention", Alphago is characterized by: the integration of different machine learning technology (for example: reinforcement learning, deep Neural network, Policy+value network, MCTS integration is innovative), game learning and self-learning integration, relatively very expandable architecture (let it take full advantage of Google's computing resources), cpu+ GPU Parallel play the advantages of integration. This "project" not only has the world's top machine learning technology, but also has very efficient code, and the full use of Google's world's most magnificent computing resources (not only the game, training Alphago is also critical).

Alphago's leap-forward growth comes from several factors: 1) 15-20 of the world's top computer scientists and machine learning experts (this is the never-before-seen luxury team in the Go field: You might think it's nothing, but consider the scarcity of such experts), 2) The technology, innovation, integration, and optimization mentioned earlier. 3) The world's largest Google backend computing platform, supply team use, 4) integration of CPU+GPU computing power.

Alphago is a universal brain that can be used in any field? Alphago's deep learning, neural networks, MCTS, and Alphago's ability to compute the capacity to expand are all common technologies. Alphago's success also validates the scalability of these technologies. However, Alphago has actually done a lot of optimization in the field of Go, in addition to the above-mentioned system adjustment integration, there are even manual settings and adjustment of some parameters. Alphago team in Nature also said: Alphago is not completely self-play end-to-end learning (such as before the same team to do Atari AI, with End-to-end, without any manual intervention to learn playing video games). If Alphago is to enter a new application area today, it should be possible to develop solutions faster and more efficiently with Alphago's underlying technologies and Alphago teams. This is where Alphago is really better than deep blue. But the development also takes time, and the world's most scarce depth-computing scientist (now up to $2.5 million in annual treatment). So, Alphago is not a generic technology platform, not an engineer can be used by the transfer API, but also far away.

If this alphago did not defeat Li Shishi, how long will it take? IBM Blue took four years from entering the master level to the tournament to defeat the world championship. Alphago should be faster than deep blue, because deep blue requires new versions of hardware, and manual tuning optimizations for Kasparov, while Alphago is based on Google's hardware computing platform and a relatively common deep learning algorithm. So, a few months too short, 4 years too long, it is expected between 1-2 years.

Is it a huge breakthrough from chess to Weiqi? Certainly yes, in this article (in the Chess field, the computer can already beat the human brain, then go field computer is still how far?) Link: https://www.zhihu.com/question/21714457), the first respondents analyzed the complexity of go and chess only. When Deep Blue defeated the World Championship in 1997, it was thought that dark blue was using a manually adjusted evaluation function, and that the complexity of the chess level was conquered by specially designed hardware and "brute force" search (Brute-force), but that Weiqi was not exhaustive because its search was too broad ( The choice of each step is hundreds of rather than dozens of) too deep (chess has hundreds of steps instead of dozens of steps). and the development of Alphago let us see, over the past 20 years of development, machine learning + parallel Computing + massive data can overcome these digital challenges, at least enough to surpass the top human.

AlphaGo beat the world champion, that means the computer beyond the human brain? Or can you think about it?

My answer is:

On the question that can be calculated by logical analysis, the machine is going to leave the human being far behind. Machine speed will be faster, learning ability will become stronger, data will be more and more. That year, we discussed "chess lost to the machine is not what, go is the real wisdom" is only our human dignity to maintain their own but not practical fantasy! Today, we should face the reality!

In the era of big Data + machine learning + massively parallel computing, we will see countless opportunities and products that can generate significant business and user value in the areas of forecasting, analysis, and recommendation. However, these solutions do not make much sense compared to humans because they are too far away (for example, the recommendation engine will be able to recommend the products you are most likely to buy, the ones you want to eat, the people you want to know, the higher return on investment and the ratio of risks to automated trading ...).

In terms of perception, humans will also be overtaken by machines. Today's speech recognition, face recognition, the future of automatic driving, are examples.

But for those fans of sci-fi movies: The above is still cold technology, whether the robot will be human? This is still unknown. After all, in the emotion, emotions, emotions, humanities art, Beauty and love, values and so on, the machine is far from the people, and even the foundation is not. For AI researchers, this is the next challenge. For us humans, before the next breakthrough, we still have to develop the right brain!

P.s.-perhaps some wonder why this topic I said so much, because in 1986, when I was reading, I had developed a black and white system (complexity), defeated the black and White Chess World Group Championship, and that system also has (very superficial) self-learning ability. Interested netizens can see my article in the year: a pattern classification approach to evaluation function learning (link: http://www.sciencedirect.com /science/article/pii/0004370288900768).

Lee: What does AlphaGo mean if he beats the world championship?