2 min readAlpha Go Zero shows how close we are to generalization of artificial intelligence

It was 1997 when the IBM Supercomputer called Deep Blue showed the world that the brainpower of machines could surpass that of humans.  Though several theories downplaying the intellectual superiority of the machine over human has been put forth, there is no denying that this moment was symbolic because it was seen as a sign that artificial was catching up to human intelligence.

From this initial victory, several chess programs such as Houdini, Rybka, Deep Fritz, and Deep Junior have been made to help train chess enthusiasts. Machines have since won against humans in a variety of games – Othello, Scrabble, poker, backgammon, even winning $1 million in Jeopardy. But what machines have failed in for years was in the classic game of Go, an eastern version of chess. When DeepBlue won against Gary Kasparov, computers could not even challenge an amateur in Go. In fact, in 2014, after years of programming and experimenting, Remi Coloum, a pioneer developer of a computer Go program, said that it will take at least another 10 years before a computer can beat humans at Go.

But all that is history. In just two years after Coloum’s pronouncements, Google DeepMind’s AlphaGo has managed to win over Lee Sedol, 18-time world champion in Go. AlphaGo won 4 out of 5 games, and in all these four victories, Sedol was forced to resign. It was a match reminiscent of the the DeepBlue-Kasparov. But what if I tell you that the odds against humans has just become greater?

In October of this year, AlphaGo Lee was knocked of its top spot, and the amazing thing was that its opponent had no previous knowledge of the rules of Go. This is AlphaZero. Unlike the original AlphaGo, AlphaZero was untrained. It learned the rules of the game through self-play. In just 40 days, surpassed the knowledge of all AlphaGo versions that beat human world Go champions.

Why it matters

The computing power of artificial intelligence is accelerating faster than was previously expected. If a narrow AI like AlphaGo can be generalized into AlphaZero and could learn a complicated game like go and emerge as champions despite without historic data,  aren’t we closer to artificial general intelligence than we think? AlphaZero could surpass all its previous versions without training in just 40 days, how fast will an AI designed for warfare learn?

AlphaGo Zero: Learning from scratch | DeepMind

AlphaGo Zero: Learning from scratch | DeepMind

We introduce AlphaGo Zero, the latest evolution of AlphaGo, the first computer program to defeat a world champion at the ancient Chinese game of Go. Zero is even more powerful and is arguably the strongest Go player in history. Previous versions of AlphaGo initially trained on thousands of human amateur and professional games to learn how to play Go. AlphaGo Zero skips this step and learns to play simply by playing games against itself, starting from completely random play. In doing so, it quickly surpassed human level of play and defeated the previously published champion-defeating version of AlphaGo by 100 games to 0.

https://deepmind.com/blog/alphago-zero-learning-scratch/

Leave a Reply

Your email address will not be published. Required fields are marked *