Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Parallel optimization algorithms for very large tensor decompositions

Papagiannakos Ioannis-Marios

Full record


URI: http://purl.tuc.gr/dl/dias/0069F4A3-9C49-47B7-A69D-11B96FBD36EA
Year 2019
Type of Item Diploma Work
License
Details
Bibliographic Citation Ioannis-Marios Papagiannakos, "Parallel optimization algorithms for very large tensor decompositions", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2019 https://doi.org/10.26233/heallink.tuc.83411
Appears in Collections

Summary

Tensors are generalizations of matrices to higher dimensions and are very powerful tools that can model a wide variety of multi-way data dependencies. As a result, tensor decompositions can extract useful information out of multi-aspect data tensors and have witnessed increasing popularity in various fields, such as data mining, social network analysis, biomedical applications, machine learning etc. Many decompositions have been proposed, but in this thesis we focus on Tensor Rank Decomposition or Canonical Polyadic Decomposition (CPD) using Alternating Least Squares (ALS). The main goal of the CPD is to decompose tensors into a sum of rank-1 terms, a procedure more difficult than its matrix counterpart, especially for large-scale tensors. CP decomposition via ALS consists of computationally expensive operations which cause performance bottlenecks. In order to accelerate this method and overcome these obstacles, we developed two parallel versions of the ALS that implement the CPD. The first one uses the full tensor and runs in parallel on heterogeneous & shared memory systems (CPUs and GPUs). The second one decomposes the tensor in parallel using small random block samples and runs on homogeneous & shared memory systems (CPUs).

Available Files

Services

Statistics