(Original title: AlphaGo with a public relations mission, face and back of Mianjiang Lake)
Director Xu Haofeng’s “Master†described this story: The only passer of Wing Chun Chuan came to Tianjin, a northern martial arts center in Tianjin, hoping to open a martial arts house in Tianjin to carry forward the outstanding knowledge of Hunchun. Tianjin Wulin did not tolerate him. While laying down vertical and horizontal rules, he also wanted Chen to deliver his real skills.
The unbearable boxer finally stated that I simply put you down.
Although the logic of the story is completely different, but in the past few days, defeating Ke Jie and AlphaGo, the masters of Go, always reminded me of the story of Master. The difference is that the boxer in the movie stays face-to-face and can't stand it down until he defeats all the masters. AlphaGo is destined to beat the masters from the start, but at the same time they must stay face to face... Perhaps that's all.
The same master challenge, the same in accordance with the rules, the same built-in splendour and tacit consent - in the extremely harmonious and commercialization of the contest + exhibition match + speech + forum + media reports, AlphaGo completed in a pleasant atmosphere A major public relation to the Chinese market and Chinese netizens, and announced that they basically completed the basic exploration of the AI+ chess field.
Even more cleverly, DeepMind and the Google team also played a "content e-commerce" homeopathic market and made several homeopathic products that have just been deployed at Google's developer conference. The strength is true, and it is true that three actions are one.
According to the “Gender of a Generationâ€, someone became a face, some people lived into a living being, they had a respectable face and smoked one, and Li Zi might want to get rid of one person. The battle of the AlphaGo's official building this time, I am afraid there is also a difference between face and inside.
Of course, there is no intention of opposing the feast of the human machine race. On the contrary, the more such games and routines, the better. But too many routines do make it difficult for us to see the most precious value that AlphaGo brings. After all, the pursuit of chess is a matter of a few, and AlphaGo's core may be more relevant.
AI+ Go's Public Relations Mission: Accurate, kind, and routine
Let’s talk about the essence of this trip to AlphaGo in China. Before the start of the game, we were swept by the news about restricting the coverage of the game. But after a few days on the spot, we will find that Google and DeepMind team leaders do not understand the official requirements and positions?
Even more than expected, the AlphaGo team seems to have received some kind of training in both words and actions, not only taking care of official emotions, but also knowing the preferences of Chinese netizens.
For instance, Demis Hassabis, the founder of DeepMind, repeatedly mentioned in the interviews and keynotes that Go is a sport with greatness and artistry, and talks about the same value and charm as exploring the universe. And also do not forget to use Go to praise China's greatness and long history. Of course, it is also necessary to indicate that it is willing to carry out more cooperation in China in the future.
In addition to the EQ online, the lineup of the AlphaGo team is good, and sincerity and fisting are also impressive. You know, DeepMind is a technology company known for its high cold. The three founders basically did not show up at the same time. This time not only the three people arrived at the same time, but they also took Google executives together and wanted to talk about the future and talk about the future. They wanted to say technology, say technology, and said that the market also said that marketing was accompanied by you. In short, strength rejects all violence.
In addition, the attitude toward the battle between Ke Jie and AlphaGo is also intriguing. In the first game, Ke Jie's half defeated the goal, the AlphaGo team immediately clarified that there was no letting. In the second game Demis Hassabis directly stated that Ke Jie's performance was almost perfect. If you want to come to the third game, even if you are not Ke Jiesheng, you should have a more “kind†winning game. .
This story can not help but think of Tianjin Competition in "Master": If you can't win, you can't open martial arts, but you won't win face, I'm sorry, it still can't open.
Google's Face: TPU, Google Cloud, and TensorFlow
In the very premise of giving China, even to human face, Google also met its own face in one fell swoop.
We can take a look at the key words of the man-machine warfare reported by the tech media in the past few days. You will find some high-frequency words that may not appear logically.
For example: TPU
As Google's processor specially tailored for machine learning, the Tensor Processing Unit (TPU) is Google’s favorite pearl since its birth. At the just-concluded Google Developers Conference, it also focused on the disclosure of information and values ​​for the second-generation TPU. Although there are only photos, it is still used as the core of a large number of products and projects.
It should be noted that TPU has strategically demonstrated that Google is a company led by products and computing services and is moving toward hardware + computing + product. Google clarified the AI ​​first strategy several times, and the chip processor provided for AI is self-evident.
It stands to reason that AlphaGo has played TPU driven calculations with Li Shiying, and in this version of the game with Ke Jie, the use of TPU has only decreased in number, and has not changed in terms of hardware computing capabilities. Should not be counted as a major change to the AlphaGo upgrade.
But during the speech shared by the founder of DeepMind and the relevant person in charge of Google, as well as answering questions from reporters. The TPU is being mentioned as the core of AlphaGo computing power. Arguably, this AlphaGo feature is a huge reduction in the amount of computing required, and it seems that there is something wrong with the processor that used it.... but that's what they did. This is the face. Google's face.
Under the roof, people naturally bow their heads. Whether you are a technical genius or an idealist, taking Google’s acquisition money will naturally serve Google.
In fact, the current market application value of TPU has yet to be considered.
First of all, this is a no-sale product. You can only obtain the computing power of TPU by purchasing Google Cloud's services. On the other hand, as a chip tailored specifically for its machine learning platform, TensorFlow, TPU's efficiency in processing other platform algorithms and other machine-learning computing networks is still questioned.
At the Wuzhen summit, David Silver, co-founder of DeepMind, deliberately emphasized in his speech that “the TPU is programmable like a CPU or GPU. It is not designed for a neural network model and can be used on multiple networks. Execute CISC directives such as convolutional neural networks, LSTM models, large fully-connected models, etc."
But in fact, according to the test value of a generation of TPU, the operating efficiency of the LSTM model is still quite low, which is far less than the operating speed of the convolutional neural network. This is why Nvidia, with machine learning hardware as its core, always seems to have some disdain for TPU: After all, the AI ​​road is ten million, and it's impossible for everyone to walk on TensorFlow.
However, it is clear that Google is hoping to package TensorFlow's platform, Google Cloud's cloud services, and TPU's computing power to more companies. This is almost the lifeblood of the entire AI application industry in the Valley singer.
This kind of thinking is unrealistic at present, after all, TPU's adaptability is too narrow, the startup company chooses GPU enough. How can big companies give their lives to Google?
Therefore, in the human-computer war, Google’s "face" was actually supported by various "little children".
AlphaGo's Lison: A low-cost and high-figuration algorithm
So, what exactly is AlphaGo's real "living child" that supports Google's face?
Logically speaking, this is the deeper reason why the new version of AlphaGo has mercilessly spiked the old version, making it easy to freehand to overcome humanity.
In the past few days, there was a concept suddenly fired, which is the "letter" in Go. What is very strange is that everyone seems to think that letting a few words correspond to several levels of strength. In fact, the "let's win" in Go is by no means counting numbers.
Every time you let a son, the winner will have a huge advantage. In general, up to four children are allowed, and letting four children basically is a teaching game for professional chess players and introductory players. Whether or not AlphaGo can play against a human player won't be known, but if it can make three sons for the old version of AlphaGo, it means that the game has evolved hierarchically.
What is the core strength of this evolution? Obviously it is not from the acceleration of computing speed and the increase of data processing volume. According to Demis Hassabis, “The new version of AlphaGo has a tenfold smaller computation, more self-playing ability, and is simpler, better, and less power consuming.â€
This mystical effect may come from several abilities. The core is that AlphaGo has improved the Monte Carlo tree search algorithm. The essence of this algorithm is to continuously exhaust the maximum and minimum values ​​in the case of complete information game, and match the search results that can achieve the player's goal based on the data results. Follow the search tree and eventually reach the optimal result.
This algorithm is the machine learning algorithm that AlphaGo began to use in its early days. However, the disadvantage of this algorithm is that the amount of computation required for the game goal will be very large. If it is not limited, it is very likely that there will be a brute force and exhaustive calculation method. This is obviously unwise for Go.
The way to solve the computational problem is to use the convolutional neural network to use the strategy network and the value network to determine the direction of motion. This greatly reduces the amount of computation and computation time of the search tree. It is said that the mental network formed by this version of AlphaGo's strategic network/value network has been increased from 12 to 40 layers. This implies not only an increase in the number, but more complex algorithmic logic.
AlphaGo's extensive learning of the human chess game has also provided a powerful foundation for the evolution of the version. This version of AlphaGo is more self-engaging to complete reinforcement learning and assist with a small amount of human chess. Obviously, it will improve on high-quality data search, and it will be even less predictable by humans.
At the same time, although AlphaGo is still a training method that uses supervised learning and enhanced learning, it has already reached some unsupervised learning. This is probably the source of many strange moves and senses of layout when the master wins 60 games.
In simple terms, AlphaGo's “little child†is that it uses a completely improved algorithm and high-quality data sources. Although there is very little information at present, we can't glimpse what it is. But AlphaGo's combination of algorithmic logic, training logic, and learning resources allows the machine to acquire something that is almost intuitive and creative, and it is probably the most important message it conveys.
Although this information is wrapped behind layers of commercial purposes and news gimmicks, it is still necessary to find out... Because I don’t know what, but here it seems that I can quote another “Guru of a generationâ€: only two kung fu Words, a horizontal one, standing right, lying down in the wrong.
The same is true of AI.
燑br>
WENZHOU TENGCAI ELECTRIC CO.,LTD , https://www.tengcaielectric.com