报告题目 (Title)：探索利用时间序列预测可学习的机器学习模型
报告人 (Speaker)： Hai Xiang Lin （林海翔） 教授（Delft University of Technology, Netherlands,荷兰代尔夫特理工大学）
报告时间 (Time)：2025年11 月08日 ( 周六 ) 14:00-16：00
报告地点 (Place)：校本部F619
邀请人(Inviter)：白延琴 教授
报告摘要：Recently, machine learning, particularly, deep neural networks (NNs) have been very successful in natural language processing (LLMs), computer visions, etc. Their skills even outperform human beings in some application areas. Time series prediction is a special type of problems that differ from language processing and image processing, it is dynamic, time dependent and often influenced by external factors. Times series predictions are important real-life applications, currently, research of using neural networks to learn from data are actively ongoing. Promising results have been reported, however, do these models really learn the underlying dynamics or physics? Does a small MSE error in the prediction means the model better learns or understands the underlying dynamics? In this talk, we discuss the learnability of NNs using times series examples, and consider the limitations of current models while emphasizing the need of incorporating physical information, such as physics-informed neural networks (PINNs).