Efficient Algorithms and Hardware Accelerators for Tensor Decomposition and Their Applications to Multidimensional Data Analysis

Professor Lee, Victor Ho Fun   (Co-Principal Investigator (Co-PI) (for projects led by other university))

Professor Yan Hong   (Project coordinator)

36

2016-06-01

2019-05-31

600000

Efficient Algorithms and Hardware Accelerators for Tensor Decomposition and Their Applications to Multidimensional Data Analysis

Hardware accelerators, Numerical analysis, Parallel processors, Tensor computing, Tensor decomposition

Computing HardwareApplied Mathematics

C1007-15G

Collaborative Research Fund (CRF) - Group Research Project

2015

Completed

1. To investigate robust mathematical models for tensor decomposition useful for the analysis of multidimensional data. Project deliverables will include: discovery and proof of theorems, and development of new concepts for tensor models related to eigne-analysis, super-modes and super-compatibilities in high-order structures, existence, orthogonality, uniqueness, and optimality issues. 2. To develop efficient computer algorithms for tensor decomposition, especially those suitable for parallel processors. Project deliverables will include: robust methods for search and optimization, regularization and special tensor structures, generalized projection methods, optimal block structures, scheduling and ordering blocks as input data to parallel processors, combined Jacobi and Gauss-Seidel/SOR iterations, convergence, speed, stability and error bound analysis. 3. To design system architecture and computer software for hardware accelerators for fast tensor computation. Project deliverables will include: mapping tensor computing algorithms to GPU, FPGA and MIC processors using the OpenCL platform, processor specific optimizations, data traffic control and memory management strategies for ""Big Data"" processing. 4. To apply tensor computing to the analysis of massive multidimensional data, such as those for imaging, video and biomolecular applications. Project deliverables will include: image and video processing and object segmentation and tracking using tensor models, with potential to be realized on low-power and portable devices for commercialization by industry targeting a mass market, extraction of coherent gene expression patterns, and identification of the most affected pathway in lung cancer treatment to combat drug resistance, with a concrete step forward towards personalized medicine.