Google Deepmind has said that its new artificial intelligence (AI) system has made a ‘major breakthrough’ by solving geometry problems with the same skill as top-level high school students.
Geometry can be challenging for AI systems due to a lack of available data to help with geometric problem-solving.
Some 30 problems were presented to the system at the International Mathematical Olympiad, a competition in which high-performing school students attempt to prove mathematical theorems. The AI was able to solve 25 of those assigned to it.
Machine learning is commonly used to feed AI systems data to help train AIs to complete tasks.
However, human demonstrations available for proving geometry theorems are a limited resource.
Dubbed ‘AlphaGeometry’, Google has developed a new methodology to build and teach the system.
This works by using a language model that can self-teach by processing millions of theorems and proofs, which was then combined with a system designed to identify branching points in similar problems.
The proofs provided by the Ai were understandable to humans, the researchers said.
The findings were published in ‘Solving olympiad geometry without human demonstrations’, which appeared in Nature.