A small step for man, but a great leap for artificial intelligence.
DeepMind, the company that Google acquired in 2014, has released the latest version of its artificial intelligence, AlphaGo. The software, which they have called AlphaGo Zero, is capable of learning on its own without human help and without previous knowledge of Go, the traditional Chinese game that the machines first beat last year. The latest version of Google's artificial intelligence, AlphaGo Zero, has beaten its predecessor by a hundred games to zero. The previous AlphaGo beat Lee Se-dol, world champion of Go, in March 2016, so the new software could be considered as almost unbeatable.
According to Demis Hassabis' team, CEO of DeepMind, the latest version of Google's artificial intelligence uses a single neural network and four specialized chips for neural network training, as explained in the work published in Nature. In contrast, the previous software used multiple machines and 48 specialized chips (TPU). The program is designed to learn from scratch in a self-taught way with the aim of predicting the selection of movements and the winner of their games, improving with each repetition of the game.
After a few days of training, AlphaGo Zero was able to complete more than five million games with itself and beat the human capacity and previous known versions. The self-learning of artificial intelligence allowed the DeepMind program to discover for itself the principles of the same game that took thousands of years to discover human beings, as well as conceptualizing and developing novel strategies to be unbeatable in this ancestral Asian pastime.
It's amazing to see how far AlphaGo has come in just two years. AlphaGo Zero is now the most robust version of our program and demonstrates the progress we can make with even less computing power and avoiding the use of human data,"notes Demis?? Hassabis, co-founder and CEO of the company.
The researcher, who came to be described by The Guardian as "the superhero of artificial intelligence,"points out that breakthroughs like his algorithm could help solve "all kinds of pressing real-world problems,"such as solving protein folding or helping in the design of new materials.
Satinder Singh, a computer specialist at the University of Michigan, views the work as "a breakthrough. The scientist points out in a podium published in Nature News & Views that the method of learning software is "novel", although it combines "basic and familiar aspects" of learning by autonomous reinforcement. DeepMind's conclusions, in the opinion of the researcher, who has not participated in the study, show that the results of artificial intelligence are probably better if they use this technique than if they rely on human experience.
However, Singh warns of the "media obsession" with the confrontation between artificial intelligence and humans. "Yes, another beautiful and popular game has fallen into the hands of the machines, and yes, the self-reinforcing learning method can be applied to other tasks. However, this is not the beginning of the end because AlphaGo Zero, like other successful programs, is extremely limited in what it knows and what it can do compared to humans and even other animals,"the expert warns. DeepMind's CEO also highlights the positive side of his research, noting that "if we can make the same progress with the current problems as with AlphaGo, we will have the potential to advance human understanding and make a positive impact on our lives.