Matrix Computations and Scientific Computing Seminar: Scalable optimization algorithms for large-scale subspace clustering
Seminar | March 22 | 11 a.m.-12 p.m. | 380 Soda Hall
Daniel Robinson, Johns Hopkins University
I present recent work on the design of scalable optimization algorithms for aiding in the big data task of subspace clustering. In particular, I will describe three approaches that we recently developed to solve optimization problems constructed from the so-called self-expressiveness property of data that lies in the union of low-dimensional subspaces. Sources of such data include multi-class clustering and motion segmentation. Our optimization algorithms achieve scalability by leveraging three features: a rapidly adapting active-set approach, a greedy optimization method, and a divide-and-conquer technique. Numerical results demonstrate the scalability of our approaches.