AlphaGo challenges Ke Jie to 8 things to help you re-understand its owner
(Original title: These eight things helped us re-understand its owner DeepMind before AlphaGo challenged Ke Jie)
Today, the two shots of the "AlphaGo", which shocked the masses of the melons, are about to be returned. The opponent of this battle will be Chinese player Ke Jie. Its significance is of course very important to the Chinese audience. Coupled with the brewing of the "ultimate battle" and up to $ 1.5 million in the history of the highest go tournament bonus, this "to defend human dignity" contest must be much attention.
However, compared to the victory or defeat in the chessboard, we are more interested in talking about the creators of AlphaGo, the famous artificial intelligence company DeepMind.
For most people, common sense AlphaGo is made by Google. In fact, DeepMind was acquired by Google in 14 years. However, the headquarters and the project have continued to remain in London, and personnel have maintained a high degree of mystery. On the other hand, even people who know about DeepMind know more about its founder, Demis Hassabis, who was a young genius, entered chess at the age of 16 and entered Cambridge at the age of 16, etc. . But DeepMind's planning, product sequence, and goal implementation as a startup are always hidden behind some veils.
From the current news, Google AI and DeepMind still maintain a highly independent property. Although there is a combination of strategy and technology, Google AI's focus on the project list can be said to have completely disappeared from the sight of DeepMind.
AlphaGo, as one of the core creations of DeepMind, is not only to challenge the human chess world. As the key carrier of the core technology, it inherits the overall company strategy.
Therefore, what has been done through DeepMind to re-recognize this company that Google brain experts call "AI strength in the world" can help us better understand AlphaGo and its greater ambition behind it.
We use the key actions of 8 DeepMinds to tease out the company's strengths and goals, hoping to provide you with a little new thinking beyond the deciding game.
A blockbuster game system
DeepMind's founders are masters of chess and video games, and the company's first step is related to the game.
In 2013, a little-known paper published at DeepMind was about their own AI game system. The computing network described in the paper is not intended to serve games. Instead, it allows an AI system to play games on its own.
The miraculous thing is that DeepMind's game system can automatically play a series of primary video games through self-learning without any contact.
This system can make selective actions by whether the score above the screen and the score in the game rises or falls.
Although this behavior is not difficult for people, it is of great significance for machine learning. Because it involves the establishment of tasks, the establishment of artificial mental networks, the establishment of deep learning models and the improvement of several key parts of the learning process, and the need for a large number of graphic processing alone to assist.
This "playing out" AI program eventually helped DeepMind get onto Nature. Then began to attract the attention of major Internet giants, and succeeded in a series of tug-of-war, was acquired by Google at the price of up to 600 million US dollars. Even more frightening is that this is also known as Google's most successful acquisition ever.
From this piece of work that has taken off, you can see several features of DeepMind. The first is the integration of a variety of complex technologies, followed by the ability to solve AI applications far beyond industry standards. His genetic traits, which love games and love for humans, are also exposed at the very beginning.
Second, give AI 3D game lessons: Open Source DeepMind Lab
At the end of 2016, DeepMind has open sourced DeepMind Lab, one of its core deep learning platforms, for use by researchers and developers.
Different from the previous deep learning open source platform, DeepMind Lab is special, it is actually a set of 3D games dedicated to AI play.
Just like humans play a first-person shooter game, this open source program can design multiple complex environment architectures designed specifically to train artificial intelligence and machine learning systems. It is used to train artificial intelligence to learn to perform complex tasks in large environments, in some visual environments, and in diverse visual conditions.
DeepMind Lab is said to have evolved on the basis of Quake 3 and has relatively strong extensibility and applicability. Can let the applicable personnel design the level and the environmental effect by oneself, the specific discrimination and the processing mechanism that the AI ​​trains out AI.
Compared to machine learning systems for data samples, DeepMind's open source system can focus on AI's visual+perceptual interactions in real environments. This is a huge brain hole for the AI ​​industry, especially for research and entrepreneurs in areas such as drones, AR, map navigation, and robot memory.
From this system of inspiration that is still derived from the game, it is not difficult to see that DeepMind's feature is that it attaches great importance to the study of AI and human beings. The goal is to transplant human perception and spiritual thinking into the machine, and they also encourage others to try in this direction.
Third, the human brain model and the classic computer fit: differentiable neural computer
Another noteworthy action was that at the end of 2016, DeepMind announced the creation of a "Differentiable Neural Computer" (DNC).
The characteristics of DNC are the combination of the operation principle of the mental network and the computing power and external storage capacity of the classic computer. In simple terms, its solution is to separate the ontology of the neural computer: the mental network set up with the human brain as the blueprint of the biological network, and separate from the external memory that can be read and written, and set up a double-layer processing and operation structure.
The core feature of this type of computing system is to solve the problem of machine memory in the actual operation of neural networks. Made a machine that can think like a human being, and can calculate and remember data at high speed like a computer. In the published paper, the computer can plan the best routes between subway stations that are far apart from each other and figure out the complicated relationships—especially if there is no a priori data.
Integration of multiple capabilities, creative emancipation of the algorithm, in this computer show the most vivid. Although the principle sounds simple, the solutions that are actually used are very complex and the collaboration in multiple areas is designed.
IV. Develop artificial intelligence training environment for "StarCraft 2"
At the 2016 Blizzcon Carnival, DeepMind announced that it will collaborate with Blizzard to create an AI system dedicated to StarCraft II. The system will be like humans to think and make decisions, and hope to replicate the miracle of AlphaGo, and eventually defeat all the human masters.
Here you can't help wondering how much the company loves to play.
However, the fact remains that in the completely dynamic game environment of "StarCraft 2", the human players always handle the situation, timing, and environment more than the computer system.
The problem of Go is that the amount of computation is huge, and the problem with this type of strategy game is that the variability is too strong and it is the core problem that artificial intelligence faces.
Obviously, it is a good show for DeepMind to do PR, storytelling, and push products by challenging humans to excel in the field.
V. The most accurate speech generation system WaveNet
In addition to playing games, DeepMind also does some "genuine things" that everyone is doing.
For example, in the last two years, DeepMind has announced its achievements in the fields of image generation and speech generation. For example, WaveNet, a voice generation system announced in 16 years, claims to reduce the gap between computer output audio and natural human voice by 50%. At least according to personal testers, this system sounds much smoother than both Google and Apple's speech production systems.
(WaveNet's WaveNet Synthetic Sound)
The advantage of WaveNet is that it synthesizes the approximate human voice through the original waveform, instead of splicing the speech samples word by word.
This makes the sound of the future machine may be closer to people, think about it is also a matter of careful thinking.
Sixth, medical application plan DeepMind Health
The above is basically the trend of DeepMind at the R&D end. The core carrier is the algorithm architecture and thesis. On the application side of the product, DeepMind also has some actions. For example, DeepMind Health, which it launched, points to the smart medical system. Based on the data obtained through cooperation with the UK's National Medical Department, it will create a system based on artificial intelligence for help in diagnosis and treatment. For example, DeepMind's intelligent diagnosis and treatment system can learn millions of eye-monitoring data and build models to identify early signs and early detection of eye diseases. In addition, DeepMind also set up some non-artificial intelligence products for mobile applications.
However, the cooperation with the British medical system also caused great trouble. DeepMind obtained the data of all British patients in the absence of supervision, and caused the media to send out a lot of panic and opposition.
Seven, help Google Province
Last summer, DeepMind, who rarely tried the application domain, tried a small knife at the home of Google.
DeepMind uses a deep learning algorithm to save 40% of the data in the cooling system of Google Data Center. The cooling system is an important system to maintain the operation of the data center server. However, because of the large number of equipment and high demand, a large amount of energy is actually wasted.
DeepMind uses the AI ​​system to optimize the power efficiency of the entire cooling system. It is said to have helped Google save hundreds of millions of dollars in one-time spending. The outside world is more concerned that such technologies can be used on a variety of large-scale industrial systems. Google has taken a case from its own building, apparently also to advertise for future large-scale development.
Eight, fully enter Google TensorFlow
In the end, one thing I have to say is that although DeepMind has maintained a high degree of independence in its products and research, it still fully integrated its research into Google's TensorFlow open source architecture in 2015. And the official highly affirmed TensorFlow's high applicability, extensibility and operating experience. In 2016, DeepMind also developed a high-level framework Sonnet that can quickly create a neural network module on TensorFlow and open source it.
It can be seen that DeepMind still supports the core business and ecological foundation of Google's AI system. And I am willing to help Google improve the ecology in this area.
In fact, what Google Ontology needs is DeepMind as an elite enterprise to provide more support at the core of the ecology. It can be done by Google, but Huashan is still very good at DeepMind.
to sum up
By summarizing DeepMind's research direction, application, and several cases related to Google. We can easily find several features of DeepMind, perhaps we can make a simple and clear portrait for it:
In terms of technology exploration, DeepMind is more directed at AI systems that are highly analogous to the human brain in the core field. Whether it is Go, games, or environmental judgments, image and audio generation, all point to this huge ambition.
In terms of technical advantages, DeepMind has a strong and multi-disciplinary talent pool on the one hand, giving it the advantage of building algorithms and products across multiple disciplines. On the other hand, DeepMind's technical advantage lies in its high application level. Although he did not publish much, it did not allow us to see a very basic and simple application model.
In terms of the commercial value of products, DeepMind may focus on medical health and machine learning applications in the vertical field, which has been emerging from their 2016 trends.
Taken together, this company is a semi-academic, semi-commercial entrepreneurial team with high technical strength, but it is also not too eager to make money – there are Google backstage, and they are still busy playing games.
Of course, this company's strongest may be the artificial intelligence brand PR, this is not, he came back to use PR people.
燑br>
Incremental Encoder is commonly used, and Absolute Encoder is used if there are strict requirements on position and zero position. Servo system should be analyzed in detail, depending on the application situation. Commonly used incremental encoder for speed measurement, which can be used for infinite accumulation measurement; Absolute encoder is used for position measurement, and the position is unique (single or multiple turns). Finally, it depends on the application situation and the purpose and requirements to be realized.
Incremental Linear Encoders,Linear Optical Encoder,Linear Position Encoder,Encoder Bearing Tester
Yuheng Optics Co., Ltd.(Changchun) , https://www.yhencoder.com