Source: Smart Things (ID: zhidxcom)
Smart things reported on July 11th that since the international Go and chess defeated the top players, DeepMind played in the "Quake III Arena" last week. The double team defeated the top human players, and the group winning rate was 74%.
For people who care about the field of artificial intelligence (AI), DeepMind is by no means a strange name. Since AlphaGo defeated South Korean Go Champion Li Shishi for the first time in 2016, DeepMind AI has launched a long-term challenge to top players in the gaming arena.
However, after refreshing the surface of the human record again and again, DeepMind faces a serious loss, loss of staff and a survival crisis that may be abandoned by Google. In other words, the presence of DeepMind every time, behind the tears.
And defeated humans, the addictive AI company
Since AlphaGo defeated South Korean Go Champion Li Shishi 4:1 in 2016, DeepMind has become more and more courageous in the field of games, constantly developing new AI systems and challenging human limits. In June 2016, DeepMind trained the AI system to achieve master skills in the Atari game "Montezuma" s revenge. After half a year, the masters of China's Yucheng Go and Wild Fox Go Games played against dozens of Chinese, Japanese and Korean Go players and 60 consecutive games without a defeat. The new AlphaGo system was released in 2017. :0 Conquered Ke Jie, the world's number one professional Go player.
▲AlphaGo and Li Shishi battle
If you asked any professional chess player before December of last year, what is the most powerful commercial chess software on the market? The most likely answers you might hear are Stockfish, Houdini, and Komodo. These three softwares have better chess performance than any professional player.
But by December 6, everything has changed. After AlphaGo retired for 5 months, DeepMind launched a new version of the AlphaGo series ——AlphaGo Zero, which learned the rules of chess from scratch in just 4 hours of training and won 28 wins in 100 games. The excellent performance of 72 levels has exceeded the outstanding performance of Stockfish. Chess games are rated using Elo, professional players are rated between 1800 and 2000, masters are rated above 2,500, Stockfish is rated at around 3,300, and AlphaGo Zero may be around 4,000 after evaluation. In addition, AlphaGo Zero also trained to beat the old AlphaGo for 3 days, 40 days to defeat the Master, and 2 hours to beat the strongest Japanese player Elmo.
▲AlphaGo and Stockfish Chess Wars
In addition to the world-famous AlphaGo series, the DeepMind training agent learns games such as Super Mario by watching videos such as Youtube.
Last week, DeepMind's game journey has a new milestone — — in the "Quake III Arena" flag-winning game and teamed up with humans to defeat the human players.
▲ "Quake" game screen
In the research that DeepMind is advancing, the most concerned is StarCraft II. In August 2017, DeepMind announced that it began training AI to play Blizzard's StarCraft II game. At that time, Gu predicted that the plan would let AI defeat the StarCraft World Championship in five years.
StarCraft is an extremely complex strategy game that requires a high concentration of players, superior sensitivity and strategic decision making wisdom. This game can fully test AI's real-time strategy and human-machine confrontation. It requires AI learning to balance development with limited resources, learn how to develop high-tech, armor confrontation, and how to mobilize to ensure maximum benefits in the future. These cover the three major problems that AI needs to solve: one is to make decisions in the case of limited vision and incomplete information, the other is to balance short-, medium-, and long-term development strategies, and the third is to deal with cooperation and game between multiple agents.
David Church, a professor of computer science at Memorial University in Newfoundland, believes that StarCraft is so complex that solving StarCraft's AI will solve any other problem.
Why do researchers love to let AI play games?
Although AlphaGo has dominated chess and Go, it does not mean that AI has the ability to solve real problems. The way AI learns to play games is not the same as the way humans understand games. The rules of the game such as chess and go are normative. Although complex but the rules are very stable, AI can exert its superior computing ability under the premise of “limited”. Although AI researchers are trying to enhance the general capabilities of AI and bring AI closer to the human brain, this vision is still only making initial progress.
The game itself can simulate the real life scene of human beings. By observing human behavior, it can achieve the goal task with half the effort and help humans make intelligent decisions in the fields of personalized marketing, resource scheduling, self-driving vehicles and drones in the e-commerce and advertising industries. Let AI play games, which can bring the following advantages to the development of AI.
1. Simulated reality + simplified process
An important reason why AI researchers are passionate about games is to solve the difficult problems that the real world is difficult to learn and deal with directly. Most of the game scenes are derived from the real world and are virtual simplifications of the real world. When training AI with games, researchers don't need to worry about hardware maintenance, they don't need to disassemble equipment, and they can easily adjust the test environment, which makes the training of new AI algorithms much less difficult. If you let the robot do related tasks in real life, the financial resources and time that may be consumed are incalculable.
Games can sometimes replace complex data sources in the real world. For example, in 2016, Princeton University's Arthur Filippwicz wanted to teach cars to identify traffic signals without human assistance. To train this algorithm, he needed to collect a comprehensive picture of traffic signals, including new, old, clean, dirty. Chaos, occlusion, glare, rain, fog, darkness and other scenes. However, to get such a complete data set is very time-consuming and laborious, so Filippwicz chose to use the traffic signal depicted in the game “Grand Theft Auto V” as the source of the training set, and he got the number. A photo of thousands of traffic signals to let his AI system learn and digest.
In addition, many games require different cognitive skills. By training and learning in different games, researchers can help researchers better understand and build a better AI system.
The task of reducing energy consumption in the data center is no different from gaming. Google has used DeepMind to learn the same algorithm used to play Atari games in February 2015 to reduce the power consumption of its large data centers. Depending on the user's needs, the server's energy consumption and the amount of heat dissipated vary greatly. DeepMind's algorithm can be used to predict the air conditioning and air conditioning required by a large number of servers, helping the data center to save 40% of the cooling system and enabling the entire data center. Reduced consumption by 15%.
2, migration learning
For a person, learning a task and learning another task is a laborious task, but this is a bit difficult for AI. At present, most machine learning algorithms assume that the feature set of the training set and the test set are the same, but this is often not feasible in reality. The main ability of migration learning is to let AI apply the knowledge and experience learned from an environment to the new one. In the environmental learning task, solve the problem of catastrophic forgetting of neural networks.
The process of AI playing games can help with migration learning. DeepMind's early neural network can only play one game at a time, even if it performs well in one game, when it is used in another game, it must reshape the already established neural network architecture, & ldquo; forget it before learning Memory, re-learning new knowledge. To get AI to do the task like a human brain, you can keep a long-term memory of your expertise while training it to play a game, and use that knowledge to master other games. DeepMind has overcome this problem in a paper published last March that allows AI to master the gameplay of multiple games at the same time as the human brain.
3, never stop
Another advantage of using game training is that AI can freely perform long-term training without the limitations of various objective conditions such as hardware devices. Through the game, AI computing performance is improved, generating a lot of data, and progress in this area is also helpful for other AI research on real-world problems.
4, to ensure that AI does not "sliding head"
In the process of researching AI technology, companies such as DeepMind did not ignore the problems and warnings of AI. DeepMind and OpenAI decided to work together to find ways to prevent AI from bringing unexpectedly bad results. DeepMind allows the AI to only deal with the problem of “seeing” in its own field of view. It does not allow the AI to request coordinates and other information directly from the computer running the game, in order to prevent the AI from going to some ordinary players who will not use it when playing the game. ; shortcuts rdquo;. Taking OpenAI's experiment in the competition game CoastRunners as an example, in the process of intensive learning, AI found that he and his quick completion of the task, it is better to always get the scores of the original, which makes researchers worried. In order to avoid similar situations, the two companies provide AI with more “human advice” to verify how AI behaves. However, it takes a lot of time to examine the test, which is better than letting the AI accidentally go crazy and destroy the earth.
The shadow behind the brilliant record
DeepMind has a world-famous reputation for defeating humans with AI. It can be said to be a magical existence in the Alphabet Group. Freedom and branding are loud, and behind it are hidden long-term losses and data undisclosed.
1, free lone ranger: refused to pick up the robot, can not see the cloud service
DeepMind is a maverick company. Its work is still focused on the development of algorithms in an ideal environment, with a focus on futurist work. It currently has more than 700 employees and writes weekly academic papers describing their work progress and latest achievements.
It is said that when the father of Android, Andy Rubin, left, one of Google's founders, Lawrence Edward Page, wanted DeepMind to take over the Google Robotics department. But Demis Hassabis, founder and CEO of DeepMind, thinks that Boston Dynamics doesn't use AI technology, which distracts DeepMind, so he rejects Page's proposal.
▲DeepMind Founder and CEO Demis Hassabis
In addition, when Google co-founder and former CEO Diane Greene was invited by Google to lead the cloud computing business, she wanted to use the high reputation of DeepMind to market Google Cloud Services. However, considering that Google Cloud's unclear market goals will undermine DeepMind's brand, DeepMind also rejected the offer.
2, burning money: long-term negative profit, the flow of funds is unknown
With the freedom of research, DeepMind has to bear the corresponding price.
According to information released by the UK government in October last year, DeepMind lost £123.5 million (US$162 million) in 2016, which is still not a small amount compared to Alphabet's total annual profit of $19 billion. Of this, 40.2 million pounds (about $52.7 million) of revenue comes from work done for other parts of its parent company Alphabet, not external customers. DeepMind also has a £41.1 million “management service fee” including the operation and maintenance of real estate and computer systems. The largest capital expenditure is in terms of “employee wages and other related costs”. DeepMind spent 104.7 million pounds ($137 million) on wages, travel, office hardware and software, compared with 44.20 million pounds in the previous year. More than double.
Legal costs for DeepMind are also rising, from £144,881 in 2015 to £658,144. According to foreign media speculation, behind this high level of increase may be related to DeepMind being found illegally holding medical health information of the British people.
3, mysticism: weaker contact with Google
The relationship between DeepMind and Google is quite interesting.
▲Deep Hind founder and CEO Demis Hassabis (left), Korean Go Champion Li Shishi (middle), and Alphabet co-founder Sergey Brin (right)
In January 2014, Google acquired the under-recognized DeepMind company for £400 million. In 2015, DeepMind also belonged to Google. In the official website, it was written in large fonts. “DeepMind is very happy to be part of Google”, but it is here. In 2015, this slogan was changed to “DeepMind is happy to join Google’s team”.
In 2016, the new version of DeepMind's official online line, “Google”, has no trace, DeepMind only on the “About Us” page. DeepMind is part of Google's parent company Alphabet Group.
DeepMind wants to gain research freedom, it needs Alphabet to provide research funding, but refuses to share data with Alphabet.
When Google acquired DeepMind, it agreed to establish an ethics safety committee to ensure that its AI technology is not abused. But then, DeepMind has not disclosed board members and discussions.
According to a news release by Financial Times in June this year, Alphabet has doubts about the reasonableness of DeepMind's expensive expenses. The Alphabet AI department urged DeepMind to explain its business model and explain to the board the flow of funds. According to the review team, DeepMind must ultimately prove its value by sharing algorithms and data or by making money, although for the time being, there is no fear that Alphabet will prevent them from doing what they want to do, but there is no guarantee that the Alphabet board will come up with different opinions and in conclusion.
Next year is the fifth anniversary of Google's acquisition of DeepMind. With DeepMind coming to Google's 75 employees, including DeepMind CEO Demis Hassabis, they can decide to stay. DeepMind will continue to rely on the budget support of Alphabet in the future, or whether it will develop independently like other departments, and it has reached a critical period of choice.
Conclusion: AI investment is still in the loss period, DeepMind's game AI new path
At present, DeepMind's work is still focused on the development of algorithms in an ideal environment. It is at the industry leading level in building AI systems, defeating humans in complex games, and learning 3D space. DeepMind believes that AI can play the same role as the game in more complex problems, becoming a multiplier for technology and even human creativity.
Although the breakthrough of AI technology in the game will eventually be transplanted to the real world, it seems that this vision is still far away, and the data in the real world is not as easy to obtain as in the game. The success of DeepMind AI in games is more like academic achievement and will not have a major impact on the company in the short term.
Although DeepMind's research on games is futuristic, some of its other research has been used by Google. For example, Google announced in October last year that it used the DeepMind AI model WaveNet in Google Smart Assist to make the machine's pronunciation closer to real people (although the price of DeepMind conversion voice service is four times higher than that of Amazon's competing products); in the same year, Google relied on DeepMind. The algorithm greatly reduces the energy consumption of the data center; DeepMind's "You might also like" suggested that the application install rate in the Google Play store on Android devices increased by 20%.
Google CEO Sundar Pichai has repeatedly said that Google's future success will be based on AI. However, the huge amount of money that Alphabet bets on AI doesn't know how long it will take to return. In addition, it is also working on AI ethics, trust funds, medical aspects, and AI ethical practices, hoping to explore and understand the impact of AI in the real world, so that AI can really play a good role in the real world.
▲Alphabet's tax and operating losses in the “Other Areas” that contain AI
As you can see from the picture, Alphabet is suffering huge losses due to research in health, robotics, connectivity and AI. In 2016, these losses amounted to US$3.77 billion, accounting for 19.8% of the total amount of the Alphabet Loss business unit. On the whole, DeepMind's loss does not seem to be serious enough to affect the Alphabet's planning for it.
Although DeepMind has not really made a profit, it may help Google to occupy the AI Highlands in the future and promote the further development of its products. However, while the DeepMind team is studying the next major challenge, the road to AI R&D is long and the business realization and long-term operational issues cannot be ignored. I hope that DeepMind's role in AI and the real world is just beginning.