NEWS  AI  January 17, 2024  Reading time: 2 Minute(s)

Max (RS editor)

DeepMind, Google's AI research and development lab, has introduced AlphaGeometry. This innovative system, unveiled today, claims to solve geometry problems at a level comparable to an International Mathematical Olympiad gold medalist, showcasing its potential to revolutionize mathematical reasoning and contribute to the development of more advanced and general AI systems.

Geometry as a Gateway to Advanced AI

DeepMind's focus on geometry stems from the belief that mastering mathematical theorems, particularly in geometry, demands both reasoning skills and the ability to navigate through a range of possible solution steps. The unveiling of AlphaGeometry marks a significant milestone, as it successfully solves 25 Olympiad-level geometry problems within the standard time limit, surpassing the previous state-of-the-art system's performance of 10 problems.

Trieu Trinh and Thang Luong, research scientists at Google AI, emphasized in a blog post that conquering Olympiad-level geometry problems represents a crucial step towards achieving deep mathematical reasoning, a key element on the path to more advanced AI systems. They express hope that AlphaGeometry will open up new possibilities across mathematics, science, and AI.

Challenges in Training AI for Geometry Problem Solving

Solving geometry problems using AI presents unique challenges due to the intricacies of translating proofs into machine-readable formats. The scarcity of usable geometry training data adds to the complexity. DeepMind overcame these hurdles by introducing AlphaGeometry as a hybrid system, combining a "neural language" model similar to ChatGPT with a "symbolic deduction engine".

To address the lack of training data, DeepMind generated 100 million synthetic theorems and proofs of varying complexity, training AlphaGeometry from scratch. The neural language model guides the deduction engine through potential answers to geometry problems, showcasing a synergy between intuitive ideas and deliberate, rational decision-making.

The Debate: Symbolic Manipulation vs. Neural Networks

The results published in a study in the journal Nature are expected to reignite the ongoing debate regarding the foundational approach to building AI systems. While neural networks, such as OpenAI's DALL-E 3 and GPT-4, are praised for their ability to emerge intelligent behavior from massive data and compute, supporters of symbolic AI argue for its efficiency in encoding knowledge and providing clear explanations for problem-solving.

AlphaGeometry, positioned as a hybrid symbolic-neural network system akin to DeepMind's previous successes like AlphaFold 2 and AlphaGo, presents itself as a potential bridge between the two approaches. This hybrid model suggests that combining symbol manipulation and neural networks could offer a more comprehensive solution in the quest for generalizable AI.

Looking Ahead: Shaping the Future of AI Knowledge Discovery

While the debate continues, Trinh and Luong underscore DeepMind's long-term goal to build AI systems capable of generalizing across mathematical fields. They envision sophisticated problem-solving and reasoning skills that future AI systems will depend on, extending the frontiers of human knowledge. AlphaGeometry, in this context, is seen as a promising step towards shaping how AI systems discover new knowledge, not just in mathematics but across diverse domains.

 IMAGES CREDITS: DEEPMIND / GOOGLE | COVER IMAGE: FREEPIK / REVIEW SPACE