Information Sciences: Computing Science: Intelligent Systems: Tractable Deep Learning: Structure vs. Scale in Data
Project runs from 03/28/2019 to 03/27/2022
Many well-established problems in data science, such as data classification and clustering, raise unprecedented challenges in presence of complex and high dimensional data. Our interest is in inference applications, and more generally analysis problems, which in turn often resort to representation theory. Our goal is to build on the many previous and more recent accomplishments in Machine Learning and data science to develop the tradeoff of structure versus deep scale when representing data for either feature characterization and exploitation or inference applications. The vast array of applications in data science we typically encounter, and of interest herein, invariably seek to glean/use an extensive number of features of the data, which may also invoke the scale information whose necessary depth, as in deep learning (DL), remains an open problem We propose to develop an analytical framework for deep structure understanding using the existing Deep Leaning development as a source of inspiration. We proposed to investigate the tradeoffs of structure versus depth, noted above , as well as the known and open challenges of DL, such as Universal Approximation Property, and convergence issues. We expect our resulting algorithmic development to hence present great advantages in relation to the state of the art, with equal or better performance, with predictable behavior.