DeepMind has unveiled a new machine learning and AI-based model, AlphaTensor, the first AI to solve matrix multiplication algorithms. It was released in a paper published in Nature and is an extension of AlphaZero, the model which searched for new optimized algorithms for matrix multiplication.
Matrix multiplication is a mathematical branch that is trivial to perform but traditionally challenging to perfect. The majority of computing devices, in particular GPUs and those designed for machine learning, place a high priority on effective matrix multiplication.
The traditional algorithms for multiplying 2×2 matrices consisted of 8 multiplication steps believed to be optimal. However, more efficient algorithms with 7 multiplication steps were discovered later on. The novel method innovated by AlphaTensor uses lesser multiplication steps. Despite the new method having many more addition and subtraction steps, for computers, those are insignificant in comparison to the multiplication step that was saved.
AlphaTensor considers the search for new multiplication algorithms like a game. And to make things more interesting, depending on the size of the matrices, AlphaTensor’s complicated game of matrix multiplication algorithm optimization has larger game state magnitudes.
AlphaTensor aims to identify efficient algorithms, and it accomplishes this by learning how matrix multiplication functions. Over time, it improves until it begins to recognize the technological advancements, at which point it moves beyond.