XPS: FULL: DSD: A Parallel Tensor Infrastructure (ParTI!) for Data Analysis

点击翻译按钮获取中文摘要

This project concerns efficient parallel algorithms and software foremerging and future data analysis and mining applications, based on anemerging class of techniques known as tensor networks. Tensors, whichare higher-dimensional generalizations of matrices, are findingapplications in signal and image processing, computer vision,healthcare analytics, and neuroscience, to name just a few. Yetdespite this demand, there is no comprehensive, high-performancesoftware infrastructure targeting server systems that may have manyparallel processors. Thus, the overarching research goal of thisproject is to design the first such infrastructure. The resultingprototype will be an open-source package, called the Parallel TensorInfrastructure, or ParTI! The broader impact of the ParTI! project isto make the use of tensors, in a variety of data processing domains,much easier to do and more widespread.The ParTI! project will focus specifically on algorithmic and softwaresupport for sparse tensors on single-node multi- and many-coreaccelerated platforms. The technical approach relies on a specific wayof representing tensors, referred to as tensor networks. A tensornetwork is an efficient approach for representing the structure of ahigh-order tensor or tensor factorization. It can used by the dataanalyst as a simple, high-level way to express the specific structureor relationships he or she seeks in the data that the tensorrepresents. However, a tensor network is not just a tool for theanalyst; it is also an abstract intermediate form, from which it ispossible to derive algorithms, express and manage parallelism, andsemi-automatically generate tensor processing software. This insight,combined with well-known data layout and communication-avoidingparallelization techniques from high-performance sparse linearalgebra, is what will enable a ParTI! for tensor-based dataanalysis. The project will show the utility of this approach byevaluating the ParTI! prototype on real data sets and systems, throughcollaborations with government research laboratory and industrypartners.For further information see the project web site at: parti-project.org