Abstract: We tackle optimal sensor placement for Bayesian linear inverse problems by establishing connections to the Column Subset Selection Problem (CSSP). We build on the Golub-Klema-Stewart (GKS) approach which involves computing the truncated Singular Value Decomposition (SVD) followed by a pivoted QR factorization on the right singular vectors. We study the effects of using the Federov exchange rule, greedily swapping sensors while improving the objective, after a GKS-style initialization. Theoretical guarantees on the number of swaps are established. Numerical experiments on model inverse problems demonstrate the performance of our approaches.

Bio: Srinivas Eswar is a postdoctoral fellow in the Mathematics and Computer Science Division at Argonne National Laboratory. He completed his PhD in CS from the School of Computational Science and Engineering at Georgia Tech where he was fortunate to be advised by Rich Vuduc and Haesun Park. His research is on scalable data mining via numerical methods with a focus on matrix and tensor approximation techniques.

See all upcoming talks at https://www.anl.gov/mcs/lans-seminars