学术交流

Multiobjective deep learning: theories, systems, and applications

发布者：梁慧丽发布时间：2025-05-08浏览次数：10

报告人

Xiaoyuan Zhang

City University of Hong Kong

时间

2025年4月8日 星期二

下午 13:00-14:00

地点

102 会议室


Abstract


Multi-objective optimization is widely used in machine learning. For instance, in robot control, designing a robot requires balancing multiple criteria, such as maintaining forward speed while optimizing energy efficiency. Similarly, when developing a large language model, several factors must be considered, including safety, helpfulness, and privacy. The underlying structure of these problems includes a Pareto set.
This report explores three approaches to finding Pareto solutions: identifying a specific solution, obtaining a diverse set, and modeling infinite solutions.
The proposed method improves convergence speed for single Pareto solutions, enhances uniformity in solution sets, and pioneers Pareto set learning in post-training of LLMs with billions of parameters.


Biography


Xiaoyuan Zhang is a final-year Ph.D. candidate in the Department of Computer Science at City University of Hong Kong. He previously earned his bachelor's and master's degrees in engineering from Shanghai Jiao Tong University. His research specializes in multi-objective optimization, covering algorithms, theoretical foundations, systems, and applications. He maintains the first gradient-based multiobjective solver, aims to solve problems with millions of parameters and authored over ten publications in leading conferences and journals such as ICML/NeurIPS/ICLR/TEVC on these topics. His research is recognized through an outstanding academic performance award in CityUHK.



上海财经大学 计算经济交叉科学教育部重点实验室©版权所有

搜索
您想要找的