报 告 人：  吴敏   新加坡A*STAR首席科学家

报告地点：天心校区电子楼207

报告时间： 2025年6月13日（周五）上午10:45-12:00

报告题目:  Towards Time Series Foundation Models

个人简介

Dr. Min Wu is currently a Principal Scientist at Institute for Infocomm Research (I2R), Agency for Science, Technology and Research (A*STAR), Singapore. He received his Ph.D. degree in Computer Science from Nanyang Technological University (NTU), Singapore, in 2011 and B.E. degree in Computer Science from University of Science and Technology of China (USTC) in 2006. He received the best paper awards in EMBS Society 2023, IEEE ICIEA 2022, IEEE SmartCity 2022, InCoB 2016 and DASFAA 2015. He also won the CVPR UG2+ challenge in 2021 and the IJCAI competition on repeated buyers prediction in 2015. He has been serving as an Associate Editor for journals like Neurocomputing, Neural Networks and IEEE Transactions on Cognitive and Developmental Systems, as well as conference area chairs of leading AI and machine learning conferences, such as ICLR, NeurIPS, etc. His current research interests focus on AI and machine learning for time series data, graph data, and biological and healthcare data.

报告简介:

Time series analysis is crucial for understanding and predicting sequential patterns in data across various domains, such as finance, healthcare, energy, and manufacturing. With the growing volume of time series data, advanced time series analysis techniques are essential for unlocking actionable insights and driving innovation in data-driven industries. This talk explores three critical aspects of advancing time series analysis: representation learning, domain adaptation, and foundation models. In the first part, we delve into Time-Series Representation Learning, a cornerstone for many downstream tasks. We explore how self-supervised learning methods can capture temporal patterns and relationships without requiring extensive labeled data. Furthermore, we discuss the potential of designing generic architectures to unify and optimize representation learning for diverse time series tasks. The second part transitions to Time-Series Domain Adaptation, focusing on scenarios where labeled data is scarce in the target domain. By leveraging self-supervised pre-training and sensor alignment strategies, we aim to bridge the domain gap and enable robust adaptation. We will also introduce AdaTime, a benchmarking framework for evaluating domain adaptation methods tailored to time series data, showcasing best practices and key challenges. Finally, we discuss the emerging concept of Time-Series Foundation Models, which aim to serve as versatile pre-trained models adaptable to a wide range of applications. Specifically, we highlight PHM-FM, a foundation model tailored for predictive maintenance.