Fast Search-based Algorithm for Smooth and Nonsmooth Optimization, with Some Applications in Imaging
主 题: Fast Search-based Algorithm for Smooth and Nonsmooth Optimization, with Some Applications in Imaging
报告人: Prof. Xiaojing Ye (Georgia State University)
时 间: 2014-06-26 15:00-16:00
地 点: 镜春园82号院甲乙丙楼82J04教室（主持人：文再文）
We introduce several variations of classical optimization methods using adaptive searches of step sizes, and show that these can significantly improve the efficiency of traditional approaches on large-scale nonconvex or nonsmooth inverse problems such as those in imaging applications. For instance, when the Barzilai-Borwein step size selection method and a properly designed line search strategy are adopted, we can solve a large class of nonsmooth optimization problems with low per-iteration cost and obtain extensively accelerated convergence. The promising performance is further demonstrated using large real-world imaging data sets.