北京雁栖湖应用数学研究院

邬荣领

蔡云峰

2024年06月15日 10:30 至 12:30

A6-101

Zoom 637 734 0280 (BIMSA)

Automated theorem proving of Olympiad-level geometry problems is considered a notable milestone in human-level automated reasoning. AlphaGeometry (by DeepMind), a neuro- symbolic model trained with 100 million synthetic samples, marked a major breakthrough. It solved 25 of 30 International Mathematical Olympiad (IMO) problems. Combining AlphaGeometry with Wu’s method, it solves 27 out of 30 problems, which outperforms an IMO gold medalist. This talk delves into the core principles of AlphaGeometry, illustrating how machine learning algorithms, particularly deep learning, are leveraged to address complex geometric problems that have traditionally been challenging for classical methods. We will explore the architecture of AlphaGeometry, highlighting its capability to autonomously learn and generalize from geometric data. Attendees will gain insights into the technical underpinnings of AlphaGeometry, including its algorithmic framework, training paradigms, and performance benchmarks. The talk will also discuss the future directions and potential expansions of this innovative approach. Join us to discover how AlphaGeometry is not only advancing the state-of-the-art in geometry processing but also paving the way for new breakthroughs across multiple disciplines.

蔡云峰从2000年9月到2004年7月在中国科学技术大学数学系学习。2004年9月，他进入北京大学数学学院开始研究生的学习，并于2009年1月获得理学博士学位。在2009年1月到2012年6月期间，他先后在中科院数学所与美国加州大学戴维斯分校从事博士后工作。2012年9月，他加入北京大学数学学院，任特聘研究员。2018年9月他加入百度研究院，从事人工智能的研究工作。2024年6月，他加入BIMSA，获聘教授。