Discovering faster matrix multiplication algorithms with reinforcement learning
Matrix multiplication is… Well, it is matrix multiplication, it is everywhere. When computers do things, more often than not – they multiply matrices. Originally thought to have complexity n^3 for nxn matrices, it turned out to have other formulations than the “straightforward” mathematical definition, leading to algorithms of lower numerical complexity. In this talk I will tell a story about how an attempt was made to make AI come up with new algorithms for matrix multiplication, and what was the outcome. Spoiler alert – it worked.
Bio
Grzegorz Swirszcz obtained his Masters and Ph.D. from the Department of Mathematics of Warsaw University (MIMUW). After his Ph.D. he continued his research on Dynamical Systems and ODEs as a postdoc in Centre de Recerca Matematica in Barcelona. This and subsequent research in that area later became his habilitation, obtained from Institute of Mathematics of Polish Academy of Sciences (IMPAN). In 2004 Grzegorz Swirszcz joined IBM T.J. Watson Research where his interests and work shifted towards applications of mathematics in modeling and Machine Learning. Since 2016 he is working on Artificial Intelligence at DeepMind.