DeepMind Is The First AI Company To Reveal Quicker Matrix Multiplication Methods

Can AI develop its own algorithms to accelerate matrix multiplication, one of machine learning’s most essential tasks? DeepMind presented AlphaTensor today in Nature as the “first artificial intelligence system for discovering novel, efficient, and provably correct algorithms.” The findings, according to the Google-owned lab, “shed light” on a 50-year-old open question in mathematics concerning determining the fastest way to multiply two matrices.

Since the Strassen algorithm was released in 1969, computer scientists have been striving to outperform its speed for multiplying two matrices. While matrix multiplication is one of the most fundamental computational processes and, as it turns out, one of the essential mathematical operations of today’s neural networks, it is also one of algebra’s simplest operations, taught in high school arithmetic.

Matrix multiplication is used to process smartphone photos, understand oral instructions, generate computer visuals for computer games, compress data, and other applications. Companies currently use expensive GPU hardware to improve matrix multiplication efficiency, so any additional speed would be game-changing in terms of cost reduction and energy savings.

According to a DeepMind blog post, AlphaTensor is based on AlphaZero, an agent that has demonstrated superhuman ability in board games such as chess and Go. This new study extends the AlphaZero adventure beyond games and into the realm of unsolved mathematical issues.