High-Performance Static and Streaming Tensor Factorization Algorithms
Abstract
Computations involving unstructured and irregular memory accesses as well as data-dependent and problem-instance specific task dependencies have always represented important but challenging classes of problems for high-performance computing.One such emerging class of problems are techniques for tensor factorization, which are widely used for dimensionality reduction and clustering in machine learning, latent parameter estimation and source separation in signal processing, and numerous other applications in diverse disciplines, such as cyber-security, chemistry and psychology. The size of the datasets that need to be analyzed using tensor decompositions is already measured in several gigabytes and terabytes, and it is expected to grow as different types of data are collectively analyzed and new application areas for tensor analysis emerge. The objective of this project is to develop new theory, algorithms, and data structures for fast and efficient serial tensor factorization as well as scalable high-performance parallel algorithms and open source software tools for tensor factorization on shared-memory, distributed-memory, and emerging parallel architectures. The proposed research will enable a leap in high-performance tensor computations in terms of both the size of data that can be handled and the processing speed, particularly for exploratory factorizations under multiple domain-specific constraints for both static and streaming scenarios. The research puts forth an approach that enables faster and better algorithms, while circumventing intermediate memory and complexity issues that are especially pronounced for big data analytics. At the same time, it will explore emerging high-performance programming abstractions and novel high-performance architectures, which will inform the development of high-performing algorithms and software tools for tensor computations, as well as other unstructured and irregular computations. The proposed development and release of the high-performance tensor factorization software tool will enable researchers and practitioners to scale up the size of data they can analyze. In addition, the high-performance computing insights that this project will generate for emerging programming models and architectures, will be applicable to many other types of problems that involve unstructured and irregular computations.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 08, 2019
- Source ID
- W911NF1810344
Entities
People
- George Karypis
Organizations
- Army Contracting Command
- United States Army
- University of Minnesota