报告题目: Orthogonal approximate message passing for signal estimation with rotationally-invariant models
报告人: 马俊杰
单位: 中国科学院数学与系统科学研究院
邀请人:张浩川
时间:2023年11月25日(星期六)10:00-11:00
地点:大学城校区工学二号馆学术报告厅
报告摘要:
Approximate message passing (AMP) algorithms are low-cost iterative algorithms for solving high-dimensional linear regression problems. With independent Gaussian measurements, the performance of AMP can be described by a state evolution recursion in the proportional asymptotic regime. Moreover, for various high-dimensional signal estimation problems, AMP achieves the statistically optimal performance among a wide class of algorithms. In this talk, we will discuss a variant of AMP based on divergence-free nonlinearities. This algorithm, which we call orthogonal AMP, admits simple state evolution characterization for general rotationally-invariant models, without the need of complicated Onsager correction terms tailored to the matrix spectrum. The simple state evolution structure makes it an appealing template for designing efficient and analyzable algorithms for various signal estimation problems, as we will briefly mention in this talk.
报告人简介:
马俊杰,中国科学院数学与系统科学研究院优秀青年副研究员。2010年本科毕业于西安电子科技大学,2015年在香港城市大学取得博士学位。曾于香港城市大学、哥伦比亚大学和哈佛大学从事博士后研究。研究兴趣包括信号处理、无线通信、信息论、机器学习等,近年来主要关注无线通信中的高维信号估计问题。曾入选中科院百人计划,主持自然基金青年项目并参与中科院先导科技专项等科研项目,目前担任中国运筹学会青年工作委员会副秘书长。