Degree

Master of Science in Engineering (MSE)

Department

Computer Engineering

Document Type

Thesis

Abstract

This thesis presents a set of four Hyperdimensional Computing (HDC) frameworks and their Android application implementations to evaluate efficiency and feasibility on resource-constrained devices. These proposed methods target a range of application domains, including wearable health monitoring, mobile malware detection, and activity recognition utilizing both computer vision and multiple sensor streams as input. The proposed frameworks utilize HDC’s simple, lightweight arithmetic operations to convert raw data into high-dimensional representations for use in both binary and multi-class classification schemes. Each method utilizes unique encoding techniques tailored for each use case, demonstrating the flexible nature and specialization HDC offers as an emerging computing paradigm. Furthermore, experiments highlight the modular nature of HDC’s class hypervector construction, enabling iterative learning, sample unlearning, analysis of sensor contributions, and sensor influence removal. Custom benchmarking applications were created for Android deployment to further evaluate inference latency, memory usage, and energy consumption. The results further support HDC’s efficiency claims, especially when compared to conventional machine learning, demonstrating comparable classification performance while remaining suitable for real-time, edge device deployment. Overall, this work establishes HDC as an adaptable and efficient machine learning technique through practical deployment to real-world smart devices, which has yet to be demonstrated in prior works.

Date

2-6-2026

DOI

https://proquest.com/docview/3347936774

First Committee Chair

Sercan Aygun

First Committee Member

Magdy Bayoumi

Second Committee Member

Martin Margala

Third Committee Member

Michael Totaro

Fourth Committee Member

Nian-Feng Tzeng

Download

Share

COinS