Alpha Go "ZERO" autonomous learning for three days, more than human millennium, Google to
chinatopwin
chinatopwin
2017-10-20 09:28:22
Beijing time on October 19th morning, Google's artificial Intelligent Company DeepMind
announced the new progress of AlphaGo in the world's top science journal Nature, can in
the absence of human intervention in the case of self learning, self learning in the new
AlphaGoZero after 3 days, with 100 than 0 results over the first generation AlphaGo.
AlphaGo zero pictures coming from the web
Self learning ability, for artificial intelligence and machine learning, is a new breakthrough. "In
AlphaGo Chinese chess player Jie Jie pictures from the network
During the game in May this year, a number of executives has been to DeepMind < > financial
(graph: 72 hours Atlas of AlphaGo training)
Human top player Ke Jie is considered to be a go genius, 6 years old began to learn chess,
AlphaGo zero pictures coming from the web
Self learning ability, for artificial intelligence and machine learning, is a new breakthrough. "In
the past, people generally think that machine learning is based on large data, but from
AlphaGoZero, we found that the algorithm is more important than data." David Silva is mainly
responsible for the AlphaGo project (David Silver) said.
Because of the use of more algorithms and less data, AlphaGoZero uses only one computing
Because of the use of more algorithms and less data, AlphaGoZero uses only one computing
machine and 4 TPU, and its beat generation AlphaGo uses multiple machines and 48 TPU.
When people are surprised at the AlphaGoZero in the field go God level, for the DeepMind
team, this is just the beginning, their purpose is to cultivate the autonomous learning ability, to
solve more in other areas it can not solve the difficult problem.
From AlphaGo, AlphaGoMaster to AlphaGo Zero
AlphaGo available in October 2015, before the well-known and chess player Li Shishi, it had
From AlphaGo, AlphaGoMaster to AlphaGo Zero
AlphaGo available in October 2015, before the well-known and chess player Li Shishi, it had
defeated the European chess championship. Said fan fan Hui Hui "accept financial reporters at
the time >, in his opinion, a program to defeat the occupation player is impossible.
He lost 0-5 to AlphaGo, but he also joined the DeepMind team, helping to train the AlphaGo.
He lost 0-5 to AlphaGo, but he also joined the DeepMind team, helping to train the AlphaGo.
2016 in March, when he helped to 4-1 beat the human top players Li Shishi.2017 at the
beginning of training under the pseudonym AlphaGo AlphaGo, "Master", the challenge 60
human players in the network, the unbeaten score.2017 in May, in Wuzhen, called Master
second generation AlphaGo 3-0 to beat the strongest human players kija.
AlphaGo Chinese chess player Jie Jie pictures from the network
During the game in May this year, a number of executives has been to DeepMind < > financial
reporters, Master has realized the self learning ability, and even have their own "intuition", "we
have found the AlphaGo does not need to rely on the human trainer." David Silva told the
financial reporter. < >
Chess and kija in AlphaGo has been under a lot of human players could not imagine the way,
Chess and kija in AlphaGo has been under a lot of human players could not imagine the way,
kija said after the game, the first generation of AlphaGo can find flaws, Master has achieved
the "God" from the leap.
AlphaGoZero in the "independent" further, in the process of training, it is self training game.
AlphaGoZero in the "independent" further, in the process of training, it is self training game.
From the figure can be seen as a start is not familiar with chess, players level is very weak,
but as time progresses, in just 3 days for each game after 4 million 900 thousand innings,
more and more strong, realize the chess level breakthrough.
(graph: 72 hours Atlas of AlphaGo training)
Human top player Ke Jie is considered to be a go genius, 6 years old began to learn chess,
at the age of 17 in the world ranked first, first human genius more than ten years of learning,
AlphaZero was 3 days to go beyond
But the DeepMind team is not here, "the meaning of AlphaGo is not to beat people, but to
But the DeepMind team is not here, "the meaning of AlphaGo is not to beat people, but to
understand knowledge, to solve more problems." David Silva said
Commercial acceleration: 1/4 energy for commercial realization
In May this year, DeepMind founder Jamis Hassabis in an interview with finance reporter
In May this year, DeepMind founder Jamis Hassabis in an interview with finance reporter
said: "if Google is a rocket, DeepMind is fuel.""
Under this positioning, although AlphaGo evolved rapidly, it will never stop on weiqi
Under this positioning, although AlphaGo evolved rapidly, it will never stop on weiqi