Google DeepMind AlphaGo Zero AI Can Now Self-Train Without Human Input

Anyone out there who shares Elon Musk's fear of a Skynet apocalypse may find Google's latest AI dubbed AlphaGo Zero to be frightening. The new AI is the followup to the original AlphaZero AI that dominated all human players in an ancient Chinese game called "Go". Google chose Go because it is said to be a game of intuition and the original AlphaGo AI was able to beat the top human player in four out of five games.

DeepMind AlphaGo

AlphaGo Zero completed three days of self-learning and then challenged AlphaGo for a match. Zero decimated its predecessor winning 100 games out of 100. "AlphaGo Zero not only rediscovered the common patterns and openings that humans tend to play ... it ultimately discarded them in preference for its own variants which humans don’t even know about or play at the moment," said AlphaGo lead researcher David Silver.

AlphaGo Zero so easily beat the original AlphaGo AI, which itself was able to beat Lee Se-Dol, considered to be one of the game's all-time masters. Before AlphaGo Zero was unleashed on the Go world, another version called AlphaGo Master was able to beat the top-ranked global Go player Ke Jie in three out of three matches.

The difference in AlphaGo Zero and other versions of the AI is that Zero wasn't given schooling on how humans play the game. "All previous versions of AlphaGo ... were told: ‘Well, in this position the human expert played this particular move, and in this other position the human expert played here’," Silver said in a video about the AI.

The result was an AI that learns how to best play Go without being constrained by human knowledge. Eliminating that human knowledge clearly improved the AI's ability to play the game. Zero reportedly was able to come up with novel moves that humans haven't tried all on its own. Zero also used a single machine neural network compared to previous versions of the AI using multiple machines. AlphaGo Zero had four data processing units while the original AlphaGo used 48. Zero also learned much faster, playing only 4.9 million training games over three days compared to the original spending months and playing 30 million games.

“People tend to assume that machine learning is all about big data and massive amounts of computation but actually what we saw with AlphaGo Zero is that algorithms matter much more,” said Silver.

If you worry about a Skynet-like apocalypse, Silver points out that the Zero AI is extremely limited in what it knows and can do compared to humans. He said, “However, this is not the beginning of any end because AlphaGo Zero, like all other successful AI so far, is extremely limited in what it knows and in what it can do compared with humans and even other animals." It might also make you feel better to know that AlphaGo doesn't program itself.

Tags:  Google, AI, deepmind