A few days ago, researchers from DeepMind, an artificial intelligence company owned by Google, through its multinational corporation Alphabet AD, sent a report from the boundaries of the universe of chess.
One year earlier, on December 5, 2017, that team shocked the world by announcing the discovery of AlphaZero, an automated learning algorithm that overcame not only chess, but also steps, or Japanese chess, and Go. The algorithm began without the knowledge of those games beyond its basic rules, and then played against itself millions of times to learn from its mistakes. Within a few hours, the algorithm became the best player, human or computerized, the world has seen.
Now the details of AlfaZero's achievements and internal activities were officially evaluated by colleagues and published in the December Science Magazine. The new article deals with numerous serious criticisms of the initial request (among other things, it is difficult to decide whether AlphaZero played against its chosen opponent, a computer monster called Stockfish, with complete impartiality). Let us consider that these problems have been wasted. In the last twelve months, AlphaZero did not become stronger, but the evidence increased for its superiority. The algorithm has a kind of intellect that people have never seen and will continue to think for a long time.
In the last twenty years, computer chess is a long way. In 1997, Deep Deep's chess program, IBM, managed to defeat the then world champion in humanity, Garry Kasparov, in a game of six matches, not so surprising result. Deep Blue can estimate 200 million positions per second. He never got tired, never made mistakes in the calculation, and never forgot what he meant a few thoughts before.
He calculated more than Kasparov, but he could not overcome his thinking. In the first game, "Deep Blue" eagerly accepted Kasparov's victim on a bishop chain, but lost the game 16 strokes later. The most sophisticated chess programs of the current generation, such as Stockhish and Komodo, still play with that inhuman style. They want to catch the opponent's pieces, but they do not have a real understanding of the game.
It changed with the emergence of machine learning. By playing against himself and updating his neural network by learning from experience, AlfaZero himself revealed the principles of chess and quickly became the best player ever. Not only could he defeat all the leading human masters – he did not even try to try – he also destroyed Stockhish, a world chess chess computer. In a game of one hundred games against that dangerous machine, AlfaZero won 28 wins and 72 boards. He has not lost any game.
The most disturbing thing is that AlfaZero seemed to express a certain understanding. He played in an intuitive and nice way, as no other computer did, and a romantic style in the attack. He made gambits and took risks. In some games he paralyzed Stockhish and played with him. In the tenth, while carrying out his attack, AlfaZero withdrew his queen into a corner of his field, far from the king of Stockfish, the view that the Queen does not usually attack. But this strange retreat proved destructive: regardless of his answer, Stockfish was doomed.
This experience reveals that AlphaZero has won because it thinks more intelligently, not because of its speed: AlphaZero only estimates 60 thousand positions per second, while Stockfish analyzes 60 million. After discovering the principles of chess, AlphaZero developed a style that "reflects the truth of the game" instead of "the priorities and prejudices of the developers," Kasparov wrote in a comment in the science.
Now, the question is whether machine learning can help people discover similar truths about major unsolved problems in science and medicine, such as cancer, immune system enigmas, and the mysteries of the human genome.
The first signs are encouraging. In August this year, Nature Medicine published two articles on the use of machine learning in the medical diagnosis. In one of them, DeepMind researchers worked with clinicians at the Moorfields eye Hospital in London to develop an algorithm capable of classifying a wide spectrum of retinal pathologies with precision similar to that of human experts. Ophthalmology suffers from a worrying lack of experts in the interpretation of millions of retinal scanning, and artificial intelligence assistants can provide tremendous help.
The other article was about an automated learning algorithm that decides whether the patient's computerized tomography tomography in a emergency room shows signs of stroke, intracranial haemorrhage, or other critical neurological event. For the victims of the blow, every minute counts. The new algorithm recognizes these and other critical episodes with an accuracy similar to human experts, but 150 times faster. This would make it possible to prioritize the most urgent cases.
However, the frustrating thing about machine learning is that algorithms can not articulate what they are thinking. AlphaZero seems to have discovered some important chess principles, but we can not share that information with us. At least until now. As people, more than answers, we want to understand what will be a source of tension in our interactions with computers.
In fact, this has happened in mathematics for years. Think about a long-standing mathematical problem called the Four-Theorem Theorem. It suggests that, under certain reasonable restrictions, each map of neighboring countries can always be colored in only four colors, so that the two neighboring countries do not have the same color.
Although the Theory of Four Colors was tested in 1977 with the help of a computer, no human being could confirm all the stages of thinking. There are still parts that require the calculation of a brute force like the one used by the ancestors of the AlphaZero chess computer. This fact surprised many mathematicians. They did not want to know that the theorem of the four colors is true, they already believed in it. They wanted to understand why it was true, and those demonstrations did not bring anything new.
But imagine that one day, in a not so distant future, AlphaZero will become a more general problem-solving algorithm. We call it AlphaInfinity. Like his predecessor, this will have a perfect understanding: he can present beautiful demonstrations, as elegant as the chess games that AlphaZero has challenged against Stockfish. And everyone will discover why one theorem is true. AlphaInfinity should not force us to accept forcefully hard and ugly judgment.
For mathematicians and humanistic scholars, that day will mark the beginning of a new era of knowledge. But maybe it will not take too long. Because machines get faster and people continue with their neurons running with a slow millisecond scale, the day will come when we can not stick with them. The dawn of human knowledge would soon give way to his twilight.
The New York Times
Translation: Jaime Arrambide