Abstract

The main focus of this doctoral thesis is to study the problem of robust and scalable data representation and analysis. The success of any machine learning and signal processing framework relies on how the data is represented and analyzed. Thus, in this work, we focus on three closely related problems: (i) supervised representation learning, (ii) unsupervised representation learning, and (iii) fault tolerant data analysis. For the first task, we put forward new theoretical results on why a certain family of neural networks can become extremely deep and how we can improve this scalability property in a mathematically sound manner. We further investigate how we can employ them to generate data representations that are robust to outliers and to retrieve representative subsets of huge datasets. For the second task, we will discuss two different methods, namely compressive sensing (CS) and nonnegative matrix factorization (NMF). We show that we can employ prior knowledge, such as slow variation in time, to introduce an unsupervised learning component to the traditional CS framework and to learn better compressed representations. Furthermore, we show that prior knowledge and sparsity constraint can be used in the context of NMF, not to find sparse hidden factors, but to enforce other structures, such as piece-wise continuity. Finally, for the third task, we investigate how a data analysis framework can become robust to faulty data and faulty data processors. We employ Bayesian inference and propose a scheme that can solve the CS recovery problem in an asynchronous parallel manner. Furthermore, we show how sparsity can be used to make an optimization problem robust to faulty data measurements. The methods investigated in this work have applications in different practical problems such as resource allocation in wireless networks, source localization, image/video classification, and search engines. A detailed discussion of these practical applications will be presented for each method.

Notes

If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu

Graduation Date

2021

Semester

Spring

Advisor

Rahnavard, Nazanin

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Electrical and Computer Engineering

Degree Program

Electrical Engineering

Format

application/pdf

Identifier

CFE0008939; DP0026218

URL

https://purls.library.ucf.edu/go/DP0026218

Language

English

Release Date

November 2021

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Download

Share

COinS