Dr. Martin Mohlenkamp, Associate Professor of Mathematics, presented “Dynamical Systems Analysis of Swamps in ALS” at a mini-symposium at the 2014 Society for Industrial and Applied Mathematics annual meeting on July 7.
The symposium was on Tensor Analysis, Computation and Application.
Abstract: The Alternating Least Squares (ALS) algorithm for approximating a tensor by a low-rank tensor is prone to periods of very slow convergence, known as swamps. Although the existence of swamps is well-known and various fixes have been proposed, the mechanisms that cause swamps are mysterious. I will discuss progress on analyzing and understanding swamps using tools from Dynamical Systems.
Comments