[Tensor Learning Team Seminar] Talk by Dr. Maolin Wang (City University of Hong Kong)

Note that this is an online talk.

Language: English

Date and Time: Wednesday, December 17th, 14:00 - 15:00 (JST)

Location: Zoom

Title: Modern Model Compression Applications at Alipay

Speaker: Dr. Maolin Wang, City University of Hong Kong, China.

Abstract:This talk will present two groundbreaking works on model compression developed through collaboration between City University of Hong Kong and Ant Group. The first work, published at WWW 2024, introduces a multi-stage compression framework combining iterative pruning and pair-wise distillation to address the computational and energy challenges of deploying large multimodal models, achieving significant latency reduction in Ant Group's advertisement auditing system. The second work, published at KDD 2025, proposes a novel cross-distillation method that enables teacher models to dynamically adapt to student models' learning capabilities, achieving an unprecedented 1.91MB BERT-based model that has been successfully deployed in Alipay's edge recommendation system serving millions of daily active devices. Together, these works demonstrate a comprehensive approach to making AI models accessible on resource-constrained devices while preserving essential capabilities for real-world applications.

Bio: Dr. Maolin Wang is a Research Assistant Professor at Hong Kong Institute of AI for Science, City University of Hong Kong. His research focuses on efficient and effective AI, model compression, LLMs and agents, and recommendation systems. He received his Ph.D. in Data Science from City University of Hong Kong, and his M.Phil. and B.E. degrees from the University of Electronic Science and Technology of China. His work has achieved significant real-world impact, with deployments at major tech companies including Baidu, Alibaba, and Ant Group, resulting in substantial performance improvements in production systems. He has published extensively in top-tier conferences including KDD, AAAI, WWW, and SIGIR, winning the KDD Best Paper Award Runner-Up.