DevOps Articles

How DeepMind’s AlphaTensor AI Devised a Faster Matrix Multiplication

After developing an artificial intelligence that can achieve https://thenewstack.io/new-google-ai-achieves-alien-superhuman-mastery-chess-shogi-go-mere-hours/, in addition to another AI that can predict https://thenewstack.io/deepmind-ai-makes-breakthrough-with-protein-folding-problem/, the researchers over at https://www.deepmind.com/ have done it again — this time using a deep learning AI model to efficiently solve a fundamental mathematics problem, while beating a 50-year-old record to boot. In a https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor from earlier this month, the DeepMind team introduces https://www.nature.com/articles/s41586-022-05172-4, an AI system that is designed for discovering new and more efficient algorithms for solving crucial mathematical operations — in this case, matrix multiplication.

In fact, the team approached the matrix multiplication problem like a game, with AlphaTensor being built upon the lessons learned from its game-playing predecessor, https://thenewstack.io/deepminds-new-milestones-on-the-road-to-artificial-general-intelligence/.

In the case of AlphaTensor, the team reformulated the problem of finding efficient algorithms matrix multiplication as a single-player game, where the “board” is translated as a three-dimensional array of numbers.

“We trained an AlphaTensor agent using reinforcement learning to play the game, starting without any knowledge about existing matrix multiplication algorithms,” explained the team.