主题二：Recent progress on oblique random forests and deep learning: Selected works and perspectives
主讲人：新加坡国立大学夏应存教授
主持人：统计与数据科学学院林华珍教授
时间：2026年10月22日（周四）下午14:00-15:00
地点：柳林校区弘远楼408会议室
主办单位：统计与数据科学学院和统计研究中心 科研处
主讲人简介：
夏应存（Yingcun Xia），新加坡国立大学统计与数据科学系教授。1999 年获香港大学统计学博士学位。2000–2001 年任伦敦政治经济学院统计系研究助理，2000–2003 年任英国剑桥大学动物学系研究助理。2003 年起任职于新加坡国立大学，历任助理教授（2003–2005）、副教授（2006–2009）、教授（2009 至今）。现任 International Statistical Review、Electronic Journal of Statistics、Computational Statistics and Data Analysis 副主编，曾任 Annals of Statistics（2008–2010；2013–2021）和 Computational Statistics（2006–2021）副主编。研究方向包括高维数据分析、机器学习方法与健康数据统计推断等。研究成果发表于 Annals of Statistics、Journal of the American Statistical Association、Biometrika、Journal of Econometrics、PNAS 等国际期刊。
内容提要：
Machine learning methods such as random forests and deep neural networks have demonstrated remarkable empirical success, particularly in high-dimensional settings. However, their theoretical foundations, especially in regimes involving overparameterization and massive model complexity, remain only partially understood.
This talk offers a personal, statistics-oriented perspective on modern machine learning, drawing on recent advances in the theory of large neural networks, ensemble learning, and high-dimensional regression. A central message is that model size alone does not determine generalization performance. Instead, appropriate regularization, whether explicit (e.g., path-norm penalties or spectral constraints) or implicit (through algorithmic structure or ensemble averaging), plays a decisive role in ensuring statistical consistency. In particular, we present recent results demonstrating that overparameterized deep neural networks can remain consistent even as their width grows arbitrarily large, provided suitable regularization is imposed.
We also discuss ensemble methods, including projection pursuit with feature bagging and random forests and their variants, which combine the strengths of classical statistical modeling and modern machine learning. These approaches can significantly improve empirical performance while maintaining strong theoretical guarantees.
