Google’s DeepMind builds hybrid AI system to unravel complex geometry problems

Researchers at DeepMind, the synthetic intelligence research division of Alphabet Inc., have created software that’s capable of solve difficult geometry proofs which might be often used to check the brightest highschool students within the International Mathematical Olympiad.

The brand new system, which was outlined within the scientific journal Nature, is alleged to be a major advance over earlier AI algorithms, which have previously struggled to copy the mathematical reasoning needed to tackle geometry problems.

AI researchers, including teams at DeepMind’s rivals Anthropic PBC and OpenAI, have been striving to enhance the reasoning and planning abilities of generative AI systems, as these are seen as crucial to creating algorithms that may match the capabilities of humans. It’s believed that if AI systems might be endowed with such skills, they won’t only give you the chance to match humans, but even surpass them and make recent scientific discoveries of their very own.

OpenAI made headlines in November when it was reported that its researchers had made a key breakthrough in creating an AI system that might solve grade school-level math problems it hadn’t come across before. It was a modest achievement, and OpenAI didn’t even confirm it officially, however it created plenty of excitement within the research community anyway.

Now, DeepMind has gone one step further, showcasing a brand new system that may solve problems at a level comparable to a human gold medalist within the IMO, which is a prestigious competition for prime school students.

DeepMind’s geometry-solving AI system combines two different techniques. One component of the software, called AlphaGeometry, is a neural network that’s based loosely on the human brain. Neural networks have been credited with a number of the biggest advances made by AI systems, but they alone weren’t capable of solve essentially the most advanced geometry problems.

Nevertheless, DeepMind paired AlphaGeometry with a symbolic AI engine, which uses a series of human-coded rules around methods to represent data equivalent to symbols, after which manipulate those symbols to reason. Symbolic AI is a comparatively old-school technique that was surpassed by neural networks over a decade ago.

DeepMind’s researchers explained that the system uses AlphaGeometry to develop an intuition about what could be one of the best approach to solving a geometry problem. This intuition is then used to guide the symbolic AI engine and give you solutions. Based on DeepMind, the brand new system was capable of achieve results which might be on a par with gold medal-winning highschool students who compete within the annual IMU challenge.

All told, it was tested on 30 geometry problems, completing 25 inside the required cut-off date. The previous state-of-the-art AI system, developed way back within the Nineteen Seventies, solved only 10 problems.

Based on the researchers, AlphaGeometry’s proofs weren’t quite as elegant as those created by humans, and they often took significantly more steps to unravel each problem than most students do. Nevertheless, in addition they identified that AlphaGeometry developed some unique approaches which will result in the invention of geometric theorems that were previously unknown to mathematics. They plan to undertake additional research to find out if that is true.

Certainly one of the predominant challenges of teaching AI systems to unravel mathematical problems has at all times been the shortage of coaching data. DeepMind got around this by taking geometry questions utilized in the IMO and synthetically generating 100 million similar, but not similar, examples. They then used this dataset to coach AlphaGeometry’s neural network, and their success highlights the potential of synthetic data for use to coach different kinds of AI systems where the shortage of coaching data has caused difficulties for researchers.

DeepMind’s researchers said the hybrid neural network/symbolic AI approach may additionally hold promise for AI in other difficult domains, equivalent to physics and finance. In those areas, problems might be solved using a mixture of explicit rules and a more intuitive sense of how those rules ought to be applied. To encourage further exploration of this idea, it’s open-sourcing AlphaGeometry’s code and training data.

Image: Freepik