DeepMind’s game-playing AI has beaten a 50-year-old record in computer science

Patreon:

Tools:

Video text: DeepMind’s game-playing AI has beaten a 50-year-old record in computer science. The problem, matrix multiplication, is a crucial type of calculation at the heart of many different applications, from displaying images on a screen to simulating complex physics. Speeding up this calculation could have a big impact on thousands of everyday computer tasks, cutting costs and saving energy. The trick was to turn the problem into a kind of three-dimensional board game, called TensorGame. The board represents the multiplication problem to be solved, and each move represents the next step in solving that problem. The series of moves made in a game therefore represents an algorithm. The researchers trained a new version of AlphaZero, called AlphaTensor, to play this game. Instead of learning the best series of moves to make in Go or chess, AlphaTensor learned the best series of steps to make when multiplying matrices. "We transformed this into a game, our favorite kind of framework," says Hubert, who was one of the lead researchers on AlphaZero. The researchers describe their work in a paper published in Nature today. The headline result is that AlphaTensor discovered a way to multiply together two four-by-four matrices that is faster than a method devised in 1969 by the German mathematician Volker Strassen, which nobody had been able to improve on since. Overall, AlphaTensor beat the best existing algorithms for more than 70 different sizes of matrix. Virginia Williams, a computer scientist at MIT’s Computer Science and Artificial Intelligence Laboratory, is excited by the results. She notes that people have used computational approaches to find new algorithms for matrix multiplication for some time—and many of the existing fastest algorithms were devised in this way. But none were able to improve on long-standing results like Strassen’s. "This new method does something completely different from what the others did," says Williams. "It would be nice to figure out whether this new method actually subsumes all the previous ones, or whether you can combine them and get something even better." DeepMind now plans to use AlphaTensor to look for other types of algorithms. "It’s a new way of doing computer science," says Kohli.