OpenTensor: Reproducing Faster Matrix Multiplication Discovering Algorithms

Matrix multiplication (MM) is a fundamental numerical operation that is used everywhere. To search for faster MM algorithms, DeepMind proposed AlphaTensor [1] based on AlphaZero [3] and constructed a Monte Carlo Tree Search (MCTS) architecture. AlphaTensor [1] not only finds a faster algorithm for matrix multiplication but also provides a new paradigm for using machine learning to solve scientific problems. However, due to the lack of open-source codes and too many algorithmic tricks, researchers may get lost in the myriad of details and find it hard to understand the key points, let alone reproduce the performance and implement it to solve other problems. In this paper, we reproduce AlphaTensor [1] and hope that it will be helpful for others to fully understand the scientific problem-solving paradigm.