University of Alabama Birmingham
Tensor Decompositions and Rank Approximation of Tensors with Applications
Abstract
dc:description.abstractTensor decompositions are higher-order analogues of matrix decompositions which have applications in data analysis, signal processing, machine learning and data mining. One of the most challenging problems in the tensor decomposition area is to approximate the rank of a given tensor. Unlike the matrix case there is no simple formula to bound the rank of a tensor. In fact, finding the exact rank of a tensor is an NP hard problem. In this thesis we formulate the tensor rank estimation of a tensor as an optimization problem and estimate the rank via \ell1 minimization. We propose a numerical iterative method based on the proximal alternating minimization algorithm and discuss the required conditions for the global convergence of the algorithm. The performance of our algorithm is tested on several types of data such as randomly generated tensors, RGB images and surveillance videos in order to separate the background and foreground of them. In addition, this thesis studies the block sampling of the alternating least squares technique (ALS) and proposes an effective method for the CP decomposition of tensors. The method is tested on randomly generated data as well as real data. The proposed method converges faster that ALS in some cases when there is a presence of swamp. In addition, it requires less computations in each iteration. Also, the application of CP decomposition in image compression is provided.
Degree
thesis:*- Level thesis:degree_level
- Dissertation
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Goudarzi Karim, Ramin
- Contributors dc:contributor
-
- Carmeliza Navasca
- Brendan Ames
- Ian Knowles
- Da Yan
- Guo-Hui Zhang
Subjects
dc:subject × 4Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.library.uab.edu/etd-collection/1771