The actual state of "deep mind" which developed Google self-learning artificial intelligence DQN, what is the purpose?

"Deep Mind" is a venture company that develops artificial intelligence that Google acquired over 400 million pounds (about 75 billion yen). This deep mind utilizes machine learning technology and neuroscience to learn so that high score can be arbitrarily given while playing games such as space invaders and block breaking "DQN(Deep Q-Network) "and it became a big topic, but what is Google plotting using artificial intelligence on a solid body?

DeepMind: inside Google's super-brain (Wired UK)
http://www.wired.co.uk/magazine/archive/2015/07/features/deepmind/viewall

DQN released by Deep Mind is a general-purpose learning algorithm created from the application of machine learning and neuroscience. What is given to DQN is that "rules" that you need to remember to play games can be played by yourself, with screen output signals from game machines and simple instructions "maximize score" I will learn by playing many times.

In an experiment that tested how much learning ability DQN has, 49 kinds of Atari 2600 games are played. As a result of the test, DQN has mastered all kinds of games in a matter of hours from fight fights such as boxing to car racing, and among the 49 games, 43 games record higher scores than existing AI In addition, it is said that the score of 29 is higher than the professional gamer.

ByChris L

When I actually let Space Invaders play, the enemies are going to suffer immediately at first, but I learned how to attack and defeat the enemy in the third play, and after another 30 minutes, I grabbed the rhythm of the game It seems that he grabbed "when to attack and when to hide". In addition, when I let DQN play the space invaders overnight, I finally became able to knock out the high score without wasting shots. When it comes to this level, it seems that DQN can play Space Invaders skillfully better than anyone in the world.

You can see how DQN plays space invaders from the following movies. It has a movement that is too high to be thought of as AI and you can see how to defeat the enemy while using the Torchka (defense position) in the position well.

DeepMind: Space Invaders - YouTube


Besides that, I am also playing DQN "Breakout" which is the origin of the block breaking game. "I learned that it is necessary to hit the balls that bounce off the bar by moving the bar and playing 100 games in 30 minutes" about the progress of DQN at that time, "said a researcher co-founder of Deep Mind Mr. Demis Hasabis.

According to Mr. Hasabis, "When one hour has elapsed from the play, the grade will improve quantitatively, but it is a level that can not be said to have remained excellent grades, but if two hours have elapsed since the beginning of the game, After four hours, DQN has created a "strategy to open a hole in the cabbage and bounce the ball across the fence" as an optimal strategy for hitting high scores in the game And, surprisingly, the engineer who designed this AI is that I did not know anything about "how to break in the breakdown created by DQN", "DQN is so unlikely to be an algorithm It is appealing to have excellent learning ability.

The way the DQN plays the breakout can be seen in the following movie. Every time the number of plays increases, you gradually learn the rules, and finally you can see how to scoop high scores by sending the ball behind a rainbow-colored cabbage.

DeepMind: Breakout - YouTube


About DQN Hasabisu commented, "This is just a game, but we can change the data we handle to stock market data." Furthermore, regarding the efforts at Deep Mind "Deep learning"When"Reinforcement learningI just multiplied the algorithms of the two research fields, in fact, in a basic way. We are interested in applying algorithms learned from one field to new areas. "

In addition, Mr. Demis Hasabis, the founder of Deep Mind, is simply an excellent computer scientist, and at the age of 4, chess was first called a priest for the first time soon after he was referred to as "the most excellent chess player in history" person. You can tell what kind of person Hasabis is such by reading the following article.

What kind of person is Deep Mind's genius Demis Hasabis who leads Google's artificial intelligence development? - GIGAZINE


Deep Mind did not release a product using DQN, much less a way to change DQN into a long-term income source. Also, the move toward such commercialization was not stopped after being acquired by Google,Deep Mind Official PageOn top of it is written only briefly as "Solve Intelligence" (unraveling intelligence) about the purpose of the project. Nonetheless, Google has paid 400 million pounds (about 75 billion yen) to buy Deep Mind.

Mr. Hasabis explains Deep Mind as "project to understand artificial general purpose intelligence (AGI)", and develops advanced image recognition system or construction of speech recognition assistant that can talk like Siri and Google Now Unlike, etc, it is the purpose of the project to create a system that can machine judgments as close as possible to humans by using machine learning and neuroscience.

"The goal in AI development is to make the machine smarter, and most existing AI program the machine, but our method is slightly different from them, so that AI can learn by yourself It's in programming, which is far more powerful than regular AI, "Hassabis said. For that goal, Deep Mind has gathered excellent computer scientists from all over the world, and more than 150 staff are working in offices in the King's Cross district of London.

ByPhOtOnQuAnTiQuE

So why do we need generic AI? In response to this question, Mr. Hasabis said, "I think that human beings will need AI ascent to progress while progressing.To handle climate, economics, epidemics etc. probably a complex interactive type It is obvious that if you try to analyze all of these data only by humans, it is very difficult and it is obvious that it hits resource-related problems.In such a case, AI will notice some sort of awareness I will state my own opinion as to whether it will help to bring it.

However, general-purpose artificial intelligence is not completed at once. Mr. Hasabis is currently developing a single algorithm that behaves like a human brain, "We are currently challenging the development of something like the" mind "that is commonly referred to, The paper published in Nature is just the beginning part of the project.In order to derive a suitable solution for a specific task it is necessary to have a long-term memory and work area, and a clear vision, , Our system is currently capable of playing the game well, but IBMdeep blueI do not have the kind of wisdom like. But even a smart program like Deep Blue needs to teach all knowledge to the engineer, otherwise it will not be useful just like supercomputers. "

ByJan Hoffmann

Mr. Hasabis envisions the roadmap for the next 20 years on deep mind and cites the appearance of AI Assistant that makes daily life smarter within 5 years in a short period of time. The speech recognition assistant function installed on existing smartphones can only be used within a fairly limited range, but if you have learning abilities like DQN, the possibilities will expand to infinity. Mr. Hassabis is imagining, for example, "I want to move, I have children but please derive a good area somewhere from the report of Ofsted (British educational evaluation organization)" It seems that the AI ​​Assistant appears to answer the question. Furthermore, after 10 years, I predicted "AI is a scientist", and some of the papers published in Nature are AI's author's, he said.

In addition, Mustafa Schleimman, co-founder of Mr. Mustafa Schleiman, describes what kind of position Deep Mind is standing in Google. According to him, "I am working with five teams, five of which are YouTube, search, health, natural language understanding, Google X team." Deep mind technology seems to be applied to these fields, and specifically it seems that personalization work etc. to make YouTube's recommended function more suitable for individuals are done. In addition, Deep Mind seems not to be asked about efforts on advertising revenue, which is the main source of Google's revenue, and is not exposed to pressure and "I am immersed in long-term research," Schleiman said.

ByThomas van de Weerd

Earlon Mask, who used to fund Deep Mind in the past, said: "The evolution rate of AI is exponential, incredibly fast, within 5 years depending on the technology developed in this field There is a possibility that a serious crisis will come in. I think that it will occur within 10 years even if it is long.When you come down to reddit"We need to worry about the evolution of AI," he says. Dr. Stephen Hawking also said "Evolution of artificial intelligence means the end of humanity"I am speaking a warning against artificial intelligence development.