时 间：2025年10月21日（周二）16：00 – 17：00
地 点：中北理科大楼A1314室
报告人：李勇祥 上海交通大学副教授
主持人：王亚平 华东师范大学教授
摘要：
By providing insights into the complex landscape of systems, uncertainty quantification (UQ) is a critical aspect of modeling and prediction, making it valuable for trustworthy predictions and informed decision making, such as Bayesian optimization. In the context of nonparametric regression, we propose a unified approach to jointly quantify aleatoric and epistemic uncertainties by modeling input and output (response) variables using a Gaussian mixture model and inferring conditional distributions for given inputs. Through a linear dimension reduction model with a projection matrix on the Stiefel manifold, input variables are transformed into a meaningful low-dimensional representation, addressing the challenge of modeling high-dimensional uncertainties. We employ a Riemannian Stochastic Gradient Descent (SGD) algorithm to optimize the projection matrix. Our approach jointly quantifies aleatoric and epistemic uncertainty in a practical, scalable, and computationally efficient manner. The proposed method stands out for its excellent expressiveness to efficiently model both aleatoric and epistemic uncertainty in high-dimensional space without requiring specified model structures. It offers an implicit regression model with limited hyperparameters, enhancing usability and reducing complexity. Numerical examples demonstrate its effectiveness in achieving state-of-the-art results in regression, quantile interval estimation, and Bayesian optimization.
报告人简介：
李勇祥博士现为上海交通大学工业工程与管理系副教授，于2019年在香港城市大学数据学科学院取得哲学博士学位。李勇祥博士从系统工程与数据科学视角，研究复杂系统质量与可靠性不确定性量化，赋能复杂系统的质量设计、质量监测以及质量诊断。研究方向主要包括计算机实验设计与分析、统计与机器学习、统计质量控制、统计信号处理。代表性成果发表在IEEE Trans. On Information Theory、IEEE Trans. on Pattern Analysis and Machine Intelligence、IEEE Trans. on Signal Processing、Technometrics、IISE Transactions等期刊。李勇祥博士先后主持国家自然科学基金青年基金和面上项目，上海市科技创新行动计划自然科学基金面上项目，并于2021年入选上海市浦江（A类）人才计划。