Date of Award

2024-08-01

Degree Name

Master of Science

Department

Computer Science

Advisor(s)

Martine Ceberio

Abstract

Neural networks are used in many real-world applications, ranging from classification tasks to medical diagnostics. For each task, a neural network is typically able to make predictions due to its ability to extract meaningful patterns from processing large amounts of data. Thus, given the increases in available data in recent decades, the performance of neural networks in making accurate predictions has greatly increased. However, this data often comes with ingrained uncertainties due to measurement errors or the inherent variability of individual data points. Neural networks can learn despite the errors in the overall data, but what if we want them to consider the uncertainty present in individual data points?

This master's thesis explores whether neural networks can make reliable predictions given inputs with embedded uncertainties. We further expand this idea into two main areas of exploration: the performance of neural networks with uncertain data represented with real values and the design of neural networks that can model and process uncertain data directly throughout their framework. We test the first idea in robot localization scenarios, which produce uncertain data due to the errors present in measurements but nevertheless require reliable results for many applications, such as autonomous vehicles. For the second area, we explore methods for propagating uncertainty throughout the neural network learning process. Such neural networks represent and maintain uncertainties through their computations, potentially creating more accurate and robust predictions under uncertainty. Methods developed in both exploration areas are tested and compared against existing techniques, such as interval computations and traditional dense neural networks, providing mixed results in their overall reliability and accuracy.

Language

en

Provenance

Received from ProQuest

File Size

90 p.

File Format

application/pdf

Rights Holder

Edwin Tomy George

Download

Share

COinS