Google DeepMind Now Capable of Complex Mathematics

Google DeepMind has reached a significant milestone by creating an artificial intelligence (AI) system capable of solving complex geometry problems. According to experts, this development represents a substantial stride toward machines possessing reasoning skills more closely aligned with human capabilities.

Challenges in Mathematics for AI Models

According to Technology Review, geometry and mathematics, in a broader context, have long posed challenges for AI researchers. Thang Wang, a co-author of the research published in Nature, notes that compared to text-based AI models, mathematics faces a shortage of training data due to its symbol-driven and domain-specific nature.

Sophisticated Approach With AlphaGeometry

DeepMind's program, AlphaGeometry, innovatively combines a language model with a symbolic engine, a type of AI utilizing symbols and logical rules for deductions. While language models excel in pattern recognition and predicting sequential steps, their logical reasoning often falls short in mathematical problem-solving. The symbolic engine, rooted in formal logic and strict rules, complements the language model by guiding it toward rational decisions.

This collaborative approach mirrors how humans tackle geometry problems—blending existing understanding with explorative experimentation. AlphaGeometry underwent testing on 30 geometry problems at the difficulty level of the International Mathematical Olympiad, completing an impressive 25 within the time limit. In contrast, the previous state-of-the-art system, developed in 1978, managed only 10.

Training AlphaGeometry and Future Prospects

To train AlphaGeometry's language model, researchers generated nearly half a billion random geometric diagrams, feeding them into the symbolic engine. The engine analyzed each diagram, producing statements about their properties, then organized into 100 million synthetic proofs to train the language model.

According to Ft, while AlphaGeometry excels in solving "elementary" mathematics problems, it still grapples with advanced, abstract problems taught at the university level. Researchers aim to extend this approach to broader mathematical fields, emphasizing the potential for AI to engage in deep reasoning beyond geometry.

Roman Yampolskiy, an associate professor of computer science and engineering at the University of Louisville, sees AlphaGeometry's capabilities as a significant step toward more sophisticated, human-like problem-solving skills in machines. Beyond mathematics, the implications extend to various fields relying on geometric problem-solving, such as computer vision, architecture, and theoretical physics.

As the AI community explores new horizons, the goal is to witness AI contribute novel insights to complex problems in research mathematics.

Photo by: Jeswin Thomas/Unsplash