Computational Mathematics and Scientific Computing Seminar

Time and Location:

Speaker:

Link:

This talk presents recent advances for efficient and accurate computation of approximate low-rank decompositions of tensors. Tensor decompositions are effective tools for a variety of problems in scientific simulation and data analysis. We present a new algorithm for the standard CP tensor decomposition, which performs alternating minimization of Mahalanobis distance to accelerate convergence and improve stability. We also present new developments in inexact optimization of tensor decompositions via randomized sketching and perturbative expansion.