Dissertation Proposal - Matthew Merris

Tensor-Based Dataset Characterization

Presented by Matthew Merris, Computing PhD Computer Science emphasis

Hybrid Presentation: City Center Plaza 352 or register to attend via Zoom

Abstract

Tensors are algebraic objects that generalize vectors and matrices and are a natural representation for multi-dimensional datasets (network/sensor networks, recommender systems, etc). Modern computational  methods are increasingly dependent on data-driven algorithms in the spaces of machine learning (ML), artificial intelligence (AI), and large-scale data analytics to achieve desired outcomes. The proliferation of
these technologies is being enabled by novel hardware and software systems specifically designed to support the associated computational workloads and the massive volumes of data that accompany them. Despite the impressive performance gains of such systems, understanding of the data component has languished in favor of advancing algorithmic and compute-based solutions. Modern approaches to generating artificial data fail to provide the degrees of realism and is negatively impacting our ability to understand and improve hardware/software system in the modern computational landscape.

The proposed work seeks to improve data understanding through the use of tensors and tensor methods to characterize datasets in terms of statistics, structure, and performance properties. Specifically, we explore statistical characterization of datasets through the use of the Generalized Polyadic tensor decomposition as a tool for probability density estimation, structural characterization through the information captured by Krylov subspaces, and model-based characterization through the analysis of tensorized activations from trained neural networks.

Committee

Dr. Tim Andersen (Chair), Dr. Grady Wright, Dr. Richard Murphy