Deepmind, Google’s AI laboratory, has just unveiled AlphaGeometry, an artificial intelligence model capable of complex mathematical reasoning. The result is truly astonishing.
The Google Deepmind unit has been busy since its inception in the middle of last year. She has just given birth to a brand new baby, AlphaGeometry, an AI already capable of competing with the gold medalists of the International Mathematics Olympiad in geometry. Chrome already had strong math capabilities, but AlphaGeometry goes way beyond that.
In an article published on the Deepmind blog, Trieu Trinh and Thang Luong, AI researchers at Google, explain: “ solving geometry problems at Olympic level is an essential step in developing deep mathematical reasoning, leading to more advanced and general AI systems “. Google thinks big, very big!
A virtuoso of geometry enabled by the neuro-symbolic approach
AlphaGeometry far outperformed the previous AI system in solving Olympic-level geometry problems. It solved 25 in the allotted time, well ahead of the other model, which had solved “only” 10. This is truly impressive, especially since the AlphaGeometry code was made public on January 17.
To achieve such results, AlphaGeometry works in a rather specific way. It combines a neural language model, analogous to ChatGPT, with a “symbolic deduction engine”. It is a type of computer system that relies on symbols (representations of objects, concepts or entities) and formal rules in order to generate logical reasoning. In this context, this allows the model to rely on predefined logical rules to manipulate symbols efficiently.
Thanks to this combination, AlphaGeometry is therefore able to guide its deduction engine quickly in order to solve geometric problems. His training was not easy, since DeepMind created 100 million synthetic theorems (artificially formed mathematical statements) to strengthen the game of his model.
What are the implications for the future of AI?
These excellent results obtained by AlphaGeometry thus relaunch a central debate in the development of AI systems: the use of neural networks or an established approach to symbolic manipulation. AlphaGeometry is therefore a hybrid system, located at the intersection of the two; this merger could thus be part of the answer. Trinh and Luong said: “ our long-term goal is to build AI systems capable of generalizing across different mathematical domains, developing sophisticated problem-solving and reasoning skills, while expanding the frontiers of human knowledge “.
AlphaGeometry is not just a simple show of force, but goes well beyond that. Its very design could very well influence future creations of artificial intelligence systems, and the methodologies on which they will rely to operate.
Sources: Tech Crunch, New Scientist, DeepMind
7