DeepMind has released a new artificial intelligence that can automatically find the most efficient and well-functioning algorithm for matrix multiplication. AlphaTensor thus improves a fifty-year-old mathematical algorithm, according to the Google company.

DeepMind, which comes under Google, writes that AlphaTensor builds on AlphaZero. That is the AI that has become known before because it can play board games like Go. AlphaTensor is an AI that can set up its own algorithms to solve matrix multiplication issues. This algebra process is used, among other things, for small-scale algorithmic applications such as recognizing speech commands or compressing images. The more efficiently an algorithm can solve such a problem, the faster the task can be completed.

The researchers at DeepMind describe in a paper published in Nature how it improves efficiency in finding such algorithms in AlphaTensor. This specifically concerns an implementation of the Strassen algorithm from 1969. That algorithm is currently the best way to perform matrix multiplication on large matrices. Before that algorithm, the idea was that the more matrices had to be multiplied, the proportionally more difficult the calculation became. The Strassen algorithm contradicts that, but it has always been difficult to translate that algorithm into code. AlphaTensor can now do that; the AI can find the most efficient algorithm on its own, without basic knowledge.

DeepMind researchers describe in the paper how AlphaTensor discovered an algorithm that can solve two matrices of four rows of four digits in 47 multiplications. With the Strassen algorithm there would be 49. Ultimately, the researchers found several thousand functional algorithms. 14,000 of them worked on four-by-four digit grids.

According to the scientists, algorithms on certain hardware, including the Nvidia V100 GPU and Google’s own Tensor TPU v2, a speed improvement of ten to twenty percent. The researchers say it proves that AlphaZero can also be used outside of games and can also solve math problems.