Computer Science and Engineering, Department of
First Advisor
Arman Roohi
Committee Members
Sasitharan Balasubramaniam, Stephen Scott
Date of this Version
12-2024
Document Type
Thesis
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Computer Science
Under the supervision of Professor Arman Roohi
Lincoln, Nebraska, December 2024
Abstract
The rapid proliferation of Internet of Things (IoT) devices has resulted in an unprecedented influx of data generated at the edge by billions of sensors. Traditional approaches relying on cloud-based processing are increasingly inadequate due to constraints in bandwidth, latency, and privacy. Edge computing has emerged as a transformative paradigm, enabling real-time data processing and decision-making by decentralizing computation to the edge. While the integration of deep learning into edge environments—termed edge intelligence—promises autonomous and personalized operations, it is hindered by challenges such as limited computational resources, energy constraints, and data redundancies.
This thesis addresses these challenges by presenting three primary contributions. The first is the introduction of a task-aware informative sampling framework, EnCoDe, designed to intelligently prioritize the most informative or high-utility data samples during training. By leveraging an adversarially trained sampler network, EnCoDe reduces computational, processing, and storage overhead while improving model compactness and generalization.
The second contribution addresses the dynamic nature of edge environments with an adaptive network inference framework tailored for energy-constrained devices, particularly battery-less IoT systems. Featuring an energy-aware scheduler that selects multiple task-specific modules with a shared feature extractor, this framework dynamically adjusts computational pathways based on energy availability and performance requirements, optimizing resource utilization without compromising accuracy.
The third contribution is the development of hardware-aware model compression techniques incorporating 5-bit quantization and weight-sharing strategies. These techniques significantly reduce model size and computational demands, making them suitable for resource-constrained edge devices. To complement these methods, a novel Processing-in-Sensor (PIS) architecture, APRIS, is proposed to facilitate energy-efficient handling of compressed neural network layers.
Collectively, these innovations lay a foundation for efficient, scalable, and privacy-preserving AI deployment in the rapidly expanding IoT ecosystem. Future research directions include the integration of task-aware sampling into federated learning frameworks to enhance privacy, reduce communication overhead, and improve global model efficiency. Additionally, dynamic inference could be extended to real-time complexity evaluation, enabling optimized routing of samples to the most appropriate computational paths.
Adviisor: Arman Roohi
Comments
Copyright 2024, Rebati Gaire. Used by permission