Hassabis: What does it mean how AlphaGo was developed

Hassabis: What does it mean how AlphaGo was developed

Media Intelligence News May 24, the Go Summit entered its second day. At the Artificial Intelligence Summit Forum held today, DeepMind founder Demis Hassabis gave a speech and introduced the development process of AlphaGo , And what AlphaGo means.

AlphaGo is an artificial intelligence system created by DeepMind. DeepMind was founded in London in 2010 and joined Google in 2014. According to Hassabis, DeepMind participated in Google s Apollo Project of artificial intelligence, and also created a new way of researching and developing science.

Founded in 2013, Giiso Information is a leading domestic technical service provider in the field of "artificial intelligence + information". It has domestic top technologies in the fields of big data mining, intelligent semantics, and knowledge graphs. At the same time, its research and development products include information robots, editing robots , writing robots and other artificial intelligence products! With strong technical strength, the company received angel round investment at the beginning of its establishment, and in August 2015 it received a pre-A round of investment of USD 5 million from Jinshajiang Ventures.

"DeepMind's vision is to study what AI is, and then use intelligence to solve all problems, that is, how we can make effective suggestions to solve problems, and we ultimately hope to establish general artificial intelligence." Hassabis said.

According to Hassabis, the universal learning machine built by AlphaGo has two characteristics. One is "learning", that is, it is not pre-programmed and can learn original materials independently. The other is versatility, that is, the same system can perform multiple tasks. "A series of algorithms and systems can do a series of tasks. These tasks may be unprecedented." Hassabis said that general strong artificial intelligence is different from weak artificial intelligence. At present, weak artificial intelligence is all The preset, in fact, the chess program designed by IBM in the 1990s is also the preset artificial intelligence. "It uses brute force search. The machine passively accepts this program and cannot learn by itself." Hassabis Speaking of.

DeepMind hopes to build a reinforcement learning framework. Hassabis said that the reinforcement learning framework is to let the "agent" deep in a real/virtual environment to accomplish a specific goal. In order to accomplish this task, the agent can observe the environment, which includes visual, auditory, and tactile observation methods. After this, set an idea in your mind, and then take action to implement the plan. You can act in real-time observation. "If this problem can be solved, strong artificial intelligence has actually been realized, which is the same as the way the human agent observes and learns." Hassabis said.

Hassabis mentioned the Atari agent, which is an AI program developed by DeepMind before AlphaGo (three years ago). It passed the 2600 test bench and tested more than 100 eight-bit Atari games from the 1980s. At that time, the Atari agent only inputs raw pixels (~30), and the goal is to use deep reinforcement learning to play Atari games well.

AlphaGo is currently DeepMind's latest artificial intelligence system. Hassabis said that although AlphaGo plays Go, it can also do other things.

Why is it so difficult to play Go on a computer? This is because its complexity makes it difficult to solve exhaustive search. This problem includes two aspects. One is that it is "impossible" to write an evaluation program to decide whether to win or lose, and the other is that the search space is too large.

For Hassabis, what is more difficult is that Go does not rely on calculations like chess and other games, but on intuition. "There is no level concept in Go. All chess pieces are the same. Go is a game of defense, so you need to calculate the future. When you play chess, you have the board in your heart and you must predict the future. A small piece can shake the whole situation and lead one. Start the whole body. Go's wonderful hand is like being inspired by the apocalypse." Hassabis explained the intuition in Go.

In terms of technology, AlphaGo uses two types of networks: strategy networks and valuation networks. These technologies were published in "Nature" last year and inspired many researchers to design their own artificial intelligence systems.

Hassabis recalled that we conducted tests since then. In 2016, we let AlphaGo play against Li Shishi. In the end, AlphaGo defeated Li Shishi 4:1. "At this moment, we waited for ten years. It was really ten years to grind the sword. "Li Shishi said with emotion.

AlphaGo has attracted 280 million viewers, 35,000 reports, and board sales have increased tenfold in the West. Hassabis recalled the wonderful part of it (AlphaGo vs. Lee Sedol). The 37th move in the second game was amazing. It made us think: For thousands of years, you humans have underestimated the 5th line. the value of. In addition, there was an extraordinary 78th move in the fourth game, and he won a set. Lee Sedol said, "I think this introduces new ideas to Go, and I feel that I have found a new reason to continue playing Go."

Hassabis said that a lot of art is subjective, and AlphaGo regards Go as an objective art, and every step will analyze the impact. "Therefore, my definition of intuition is to get a preliminary perception through experience, which cannot be expressed, and its existence and right or wrong can be confirmed through behavior." Hassabis said that AlphaGo can already imitate human intuition and is creative. The ability to combine existing knowledge or unique ideas. So AlphaGo already has intuition and creativity, but these abilities are currently only limited to Go.

After that, DeepMind hoped to fill the gap in AlphaGo's knowledge, and then released a new version of "Master", played games on the Internet, and also won a lot of victory. After playing with the Master, Ke Jie lamented that after thousands of years of actual combat exercises have evolved, computers have told us that all humans are wrong. Hassabis said, "In the 1930s and 1940s, Wu Qingyuan brought revolutionary power to Go. I believe that AlphaGo can also open a new era of Go. Chess program talks about tactics, and AlphaGo talks about strategy."

"How far are we from the best? What is the perfect game? 3000 years of games are not enough to find the best game. AlphaGo allows us to explore these mysteries." Hassabis said.

Founded in 2013, Giiso Information is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots . At the beginning of its establishment, the company received an angel round of investment, and in August 2015, it received a pre-A round of investment of USD 5 million from Jinshajiang Ventures.

In addition to Go, Hassabis hopes to apply artificial intelligence to various fields. Hassabis said, Human-machine cooperation can achieve the effect of 1+1>2, and human intelligence will be amplified by artificial intelligence. Both artificial intelligence and AlphaGo are tools, just like the Hubble Telescope, which can promote the progress of human civilization. "Hassabis said that countless other fields will also be bombarded by combination. Strong artificial intelligence is also the best tool for us to explore. For example, AI is used in material design, new drug development, and real-life applications, such as Medical care, smartphones, education, etc.

At present, DeepMind has successfully used AlphaGo in the optimization of data centers, and the results show that it can save 50% of power.

Finally, Hassabis concluded that information overload and system redundancy are huge challenges, and we hope to use AI to find meta-solutions. "Our goal is to achieve artificial intelligence science, or artificial intelligence-assisted science. Of course, AI must be bound by ethics and responsibility. In short, artificial intelligence technology can help us better explore the mysteries of the human brain.