Beijing time on October 19th early morning, Google's artificial intelligence company DeepMind in the world's top scientific journal Nature announced a new AlphaGo progress, can in the absence of human intervention in the case of self-learning , new AlphaGoZero in the self learning for three days before beating the first generation of AlphaGo 100-0.

AlphaGo zero image from the web

The emergence of the ability to self-learn is a new breakthrough for artificial intelligence and machine learning. "In the past, it was widely believed that machine learning was based on massive amounts of big data, but from AlphaGoZero, we found that algorithms are more important than data." David Silver, the principal of the AlphaGo project, said.

Also because it uses more algorithms and less data, AlphaGoZero uses only one machine and four TPUs, whereas the generation of AlphaGo that it beat used multiple machines and 48 TPUs.

While people are amazed at how godlike AlphaGoZero is at Go, for the DeepMind team, this is just the beginning, and their aim is to solve more tricky problems that are currently unsolvable in other fields by fostering the ability to learn on its own.

From AlphaGo, to AlphaGoMaster, to AlphaGo Zero

AlphaGo came out in October 2015, and before its widely publicized game against chess player Lee Sedol, it had already defeated European Go champion Fan Hui. Fan Hui said in an interview that at the time it seemed to him impossible for a computing program to beat a professional player.

As a result, he lost to AlphaGo 0-5, but as a result, he joined the DeepMind team to help train AlphaGo.In March 2016, AlphaGo, which he helped train, beat top human player Lee Sedol 4-1.In early 2017, AlphaGo, which goes by the name " Master", challenged 60 human chess players on the Internet, maintaining a winning record.In May 2017, in Wuzhen's, the second-generation AlphaGo, named Master, defeated Ke Jie, currently the strongest human chess player, 3-0.

AlphaGo versus Ke Jie Image from the web

During the tournament in May this year, a number of DeepMind executives had already told reporters that Master has realized the ability to self-learn, and even has its own "intuition", "Master has realized the ability to self-learn, and even has its own "intuition". strong>, "We found that AlphaGo no longer needs to rely on human trainers." David Silva told reporters.

The game with Ke Jie, AlphaGo has been able to play a lot of human players can not imagine the road, said Ke Jie after the game, the first generation of AlphaGo can still find the cracks, Master has realized the "from man to God" leap.

AlphaGoZero is a step further in terms of "independence", and in the course of training, it is a self-playing game. As you can see from the training charts, both players were weak because they were unfamiliar with Go at the beginning, but as time progressed, after playing 4.9 million games against each other in just 3 days, they got stronger and stronger, and realized a breakthrough in Go level.

(Photo: 72-hour chart of AlphaGo's training)

Top human player Ke Jie is considered a Go genius, who started learning the game at age 6 and was ranked No. 1 in the world at age 17.A human genius's decade of learning was surpassed by AlphaZero in 3 days.

But the DeepMind team aspires to do more than that, "AlphaGo's significance does not lie in defeating human beings, but in comprehending knowledge and solving more problems. " said David Silva.