DeepMind AI Breaks 50-Year-Old Matrix Multiplication Record
DeepMind's AlphaZero AI discovers a faster method for matrix multiplication, beating a record that has stood for over fifty years. This breakthrough promises to speed up thousands of everyday computing tasks while saving significant energy.
DeepMind uses its game-playing AI, AlphaZero, to discover a faster way to solve matrix multiplication, breaking a 50-year-old record in computer science. Matrix multiplication is a fundamental calculation that powers thousands of everyday computer tasks, from displaying images on screens to simulating complex physics and running machine learning models. Speeding up this core process holds the potential to significantly cut computing costs and save energy across the technology industry.
Despite being a basic concept taught in high school, finding the absolute fastest algorithm for matrix multiplication remains one of the biggest open problems in computer science. The difficulty arises because there are more possible ways to multiply two matrices together than there are atoms in the universe, making it impossible for humans to simply search through all the options. DeepMind tackles this vast complexity by converting the mathematical challenge into a custom three-dimensional board game called TensorGame.
In TensorGame, the game board represents the multiplication problem, and each move corresponds to the next step in solving it, meaning a complete game essentially writes a new algorithm. By training a new version of AlphaZero on this game, the AI successfully navigates the near-infinite possibilities to find highly efficient solutions that outperform methods developed by human mathematicians over the last five decades.