Deepmind: How was the "$ 400 million -dollar super intelligence" created with Alphago?

By : ilikephone / On : 28/01/2023

The future artificial intelligence began with "Space Invaders".At first, the enemy aliens were overwhelmingly strong, and the player's laser turrets were destroyed in less than a few seconds.It's a game over in no time.But in 30 minutes, the weak player begins to grab the rhythm of the game, when to shoot back and when to hide.After continuing the game overnight, I finally got into the game so that one shot was not wasted.While repelling the aliens, the enemy's mother ships in the meantime are easy to destroy and hit a high score.Nobody in the world is no longer an enemy.

In fact, this player is not a human.It's a computer algorithm developed by a company called Deepmind.This algorithm called Deep Q-Network (DQN) is programmed on a graphic processor, giving a simple instruction to "Maximize the score" and a 30,000-pixel data stream per frame.

The next challenge of DQN is a breakout, which breaks the walls made of seven -color blocks with a ball like table tennis.


"It's terrible when you play 100 games in 30 minutes for the first time, but DQN is learning how to hit the paddle on the ball in the meantime."The 39 -year -old Demis Hasabis, 39 -year -old, co -founder and CEO of Deep Mind (CEO), and artificial intelligence researchers.

"The scores will go up in one hour, but it's not so clever yet, but after two hours, you will be able to master the game almost perfectly and cope with the high -speed balls. 4 hours later. , I will come up with a new technique myself. For example, a hole is made on both sides of the wall and the ball hits back to pass through the hole with a much accuracy that exceeds humans. Algorithm developers are such techniques. I did not know"

In February 2015, a paper written by Hasabis and his colleagues was published in the Nature magazine.The artificial intelligence programs that they developed were able to play 49 kinds of TV games for Atari2600 without any teaching.DQN mastered all games, from fighting games to 3D racing games, and overwhelmed (human) professional gamers many times.

"It's just a game now, but the data in the stock market is fine," says Hasabis."Deep neural networks and reinforced learning algorithms. Deep Mind has integrated these two promising research areas with basic Revels. We are interested in the knowledge learned in a certain field.It is an algorithm that can be applied to the field of

Deep Mind has not announced any products so far.I have not even found the way to connect artificial intelligence that masters the game to revenue.But that trivial thing was not a problem for Google.In January 2014, Google acquired an investor, such as Elon Musk, Peter Teil, and Li Kashin, and acquired a Deep Mind -based Deep Mind in London.The $ 400 million acquisition was Google's largest investment in the European region.