Google DeepMind used artificial intelligence (AI) to almost match the geometry problem solving skills of some of the brightest mathematicians in the world. This is a major step forward in the quest to apply this fast-growing technology for complex mathematics.
According to a Wednesday Nature article, AlphaGeometry is the system of the tech giant that correctly answered 25 out of 30 questions in the International Mathematical Olympiad for high school students.
AI’s maths proficiency is growing, but there are still obstacles to overcome. Complex maths presents a challenge to reasoning and learning, making it a crucial test for the creation of an artificial general intellect (AGI) capable of equaling or surpassing humans.
Quoc Le, a DeepMind research scientist, said: “This is an important step towards building AGI.” This is yet another example of how AI can be used to advance science and help us better understand how the world functions.
AlphaGeometry, also known as a neuro-symbolic method, combines language learning with deductive reasoning. The company compares its hybrid method to the phrase “Thinking, Fast and Slow” coined by Daniel Kahneman, which describes the power of harnessing rapid pattern recognition with more deliberative logic thinking.
Trieu Trinh, a member of DeepMind’s research team, described the approach as combining “the best of two worlds” in solving geometry problems. This field is familiar to us all at the level of everyday observation of shapes and spaces, but is also underpinned by a complex mathematical framework.
Researchers built up a database of 100mn synthetic geometry examples as a training set for the system. The system’s performance of 25 out of 30 was not far behind the benchmark of 25.9 achieved by human Olympiad winners from 2000-2022, and it was well ahead of its previous automated system which scored 10.
AlphaGeometry found some problems difficult and others confusing. The program was unable solve the conundrum of intersecting circle that was solved by Le Ba Khanh Trinh, a Vietnamese mathematician who inspired some of the researchers.
DeepMind’s and other researchers’ ultimate goal is to develop AI systems capable of solving maths problems beyond the human brain.
Mikhail Burtsev is a Landau AI fellow at the London Institute for Mathematical Sciences. He said that DeepMind’s work was an important step forward, but only “within the limits set by the challenge”.
He said, “The steeper challenges remain.” It is to see if AI can solve a problem that has not yet been solved.
It is still a long way off that an AI maths program will be able to defeat a human opponent, like the Deep Blue computer did in 1997 with Garry Kasparov, world chess champion.
DeepMind has no immediate plans to compete in the International Mathematical Olympiad, but the company is not ruling it out. The company is pushing ever more into the demanding world of mathematics.