Hellenic Open University
Resource optimization methods for Machine Learning model execution on resource constrained embedded systems
Abstract
dc:descriptionThis dissertation investigates the optimization of machine learning (ML) workloads for execution on resource-constrained microcontroller units (MCUs), addressing both algorithmic and architectural challenges. The rapid emergence of edge Artificial Intelligence (AI) applications necessitates real-time inference under strict energy, memory and latency constraints. This work presents a systematic study of key model compression techniques -quantization, pruning and factorization- and their impact on inference performance and energy consumption on modern embedded platforms. Theoretical discussions in Chapters 1 and 2 explore the evolution of embedded intelligence, the logic of neural networks and fields of optimization such as quantization-aware training and post-training pruning. Chapter 3 presents an extensive experimental evaluation of these techniques across various microcontroller architectures, using rigorous power measurement methodology based on triggered digital multimeters and net energy computation. Chapter 4 focuses on hardware acceleration methods -vectoring, such as Single Instruction Multiple Data (SIMD), hierarchical caching and dedicated neural processing units (NPUs)-providing both architectural analysis and empirical results. Detailed experiments on the ESP32 (LX6), ESP32-S3 (LX7) and Cortex-M85 show significant performance and energy gains from vectoring and caching. In parallel, experiments on the Alif Semiconductor E7 (Cortex-M55 + Ethos-U55 NPU) demonstrate the transformative role of NPUs, achieving up to 143× energy savings and 125× latency reduction on complex models like TinyYolo. Chapter 5 synthesizes findings and proposes future directions, emphasizing hardware-software co-design, dynamic accelerator utilization and full-system optimization. It concludes that modern MCUs, when equipped with domain-specific accelerators and matched with optimized models, can deliver real-time AI inference at ultra-low power budgets, enabling scalable deployment of intelligent edge systems.
Degree
thesis:*- Grantor dc:publisher
- Hellenic Open University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Fanariotis, Anastasios
- Φαναριώτης, Αναστάσιος
Subjects
dc:subject × 26- Eνσωματωμένη Μηχανική Μάθηση
- Τεχνητή Νοημοσύνη σε Μικροελεγκτές
- Ενεργειακά Αποδοτικά Νευρωνικά Δίκτυα
- Μέτρηση Ενέργειας Ανά Inference
- Συστήματα Περιορισμένων Πόρων
- Ενεργειακά Αποδοτικό ΑΙ
- Embedded Machine Learning
- Microcontroller-Based AI
- Edge AI Inference
- Energy-Efficient Neural Networks
- Per-Inference Energy Measurement
- Resource-Constrained Systems
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Πληροφοριακά συστήματα
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Information Systems
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60520