Topic：Advancing Machine Learning and Optimization with Sum-of-Minimum Model

Speaker：Wotao YIN, Director of Decision Intelligence Lab, Alibaba's DAMO Academy

Host：Ming YAN, Associate Professor & Assistant Dean (Undergraduate Affairs), School of Data Science, CUHK-Shenzhen

Date：21 June (Friday), 2024

Time：11:00 AM - 12:00 PM, Beijing Time

Format：Hybrid

Venue：103 Meeting Room, Daoyuan Building

Live on Wechat Channels：

Language：English

Abstract：

We present sum-of-minimum optimization, a novel and powerful framework that jointly optimizes the assignment of data points to expert models and the models' parameters. We demonstrate its effectiveness on various machine learning tasks and in configuring optimization algorithms, exhibiting superior performance over traditional approaches. Although the sum-of-minimum problem is nonconvex and nonsmooth, we provide logarithmic optimality guarantees for initialization and an O(1/t) convergence rate for optimization. We highlight exciting future directions, including extensions to stochastic optimization and applications to large-scale, real-world problems, and argue that sum-of-minimum optimization represents a promising new paradigm for machine learning and optimization.

Collaborators: Lisang Ding, Ziang Chen, and Xinshang Wang

Biography：

Dr. Yin works at Alibaba's DAMO Academy, directing the Decision Intelligence Lab. Prior to joining DAMO Academy, he was a professor in the Department of Mathematics at UCLA, where he made many contributions to the theory and applications of distributed computing, optimization algorithms, machine learning, and image processing. He has received the NSF CAREER Award, the Sloan Research Award, the Morningstar Medal in Applied Mathematics, the DAMO Award, and the Egon Balas Award.