Revisiting Unbiased Risk Estimation
数学专题报告
报告题目(Title)：Revisiting Unbiased Risk Estimation
报告人(Speaker)：王超（南方科技大学）
地点(Place)：后主楼1124
时间(Time)：2026年9月14日（周一）16:00-17:00
邀请人(Inviter)：刘君
报告摘要
Unbiased risk estimation enables image restoration without clean reference images by estimating the expected mean squared reconstruction error from noisy observations. This talk revisits this principle through multiplicative noise removal, where signal-dependent fluctuations complicate conventional risk estimation. We consider two complementary aspects. First, we investigate neural network representations, using their expressive capacity to model image structure while developing practical risk-based objectives for single-image training and model selection. Second, we explore nonlocal linear representations, exploiting similarities among image patches to obtain a computationally efficient restoration framework. Under moment and independence assumptions, this approach accommodates multiple multiplicative noise distributions and yields closed-form denoising weights. Together, these perspectives show how the choice of representation shapes the statistical guarantees, computational efficiency, and restoration capabilities of unbiased risk estimation.
主讲人简介
王超，南方科技大学统计与数据科学系副研究员，博导，其研究方向主要为图像处理、科学计算与交叉学科的数据科学。以第一作者或通讯作者身份在Cell子刊、SIAM系列、IEEE汇刊等权威期刊及CCF-A会议发表论文。入选广东省青年珠江学者，获CVPR研讨会最佳论文奖，CSIAM年会学生论文奖，以及SIAM差旅奖。主持国家自然科学基金项目两项。
