Year: 2025
Type: dissertation
Abstract: Deploying Deep Learning (DL) models on low-power Microcontroller Unit (MCU) devices presents significant challenges due to limited computational resources and storage constraints. This thesis explores novel methods for optimising DL models for edge-based tiny Machine Learning (tinyML) applications, with a focus on Human Activity Recognition (HAR) and stress detection. A multi-model system was implemented on an Arduino Nano 33 BLE Sense, where a HAR model provided contextual information to improve stress detection. The HAR model achieved 98% accuracy in distinguishing between exercise and no exercise, while the stress detection model reached 88% accuracy. Quantisation techniques were employed to reduce memory requirements, allowing both models to fit within 1.3MB of storage. To further optimise deployment, this thesis introduces the tiny Inception Module (tIM), a DL layer-sharing approach designed for tinyML systems. The tIM enables pre-existing network layers to be repurposed across models, reducing redundancy while maintaining accuracy. A case study using a tIM demonstrated a 28.3% storage saving compared to the initial multi-model system and a 47.8% reduction when compared to non-layer-sharing approaches, with only a 4% accuracy sacrifice to the stress model. Additionally, Echo State Networks (ESNs) were investigated as an alternative to traditional DL approaches due to their low training cost. However, ESNs have historically been unsuitable for applications due to their high computational complexity. This thesis presents a novel Multi-Task Learning (MTL)-based ESN approach that shares a single reservoir across multiple datasets, including chaotic systems, time-series, and image classification tasks. The shared ESN reservoir achieved competitive performance with state-of-the-art literature while significantly reducing Floating Point Operations (FLOPs) and storage requirements. In a system classifying ten different tasks, the MTL-based ESN implementation required fewer FLOPs than a single convolutional layer used for MNIST classification, while achieving superior efficiency compared to MobileNetV2 and MCU-Net. Overall, this PhD research demonstrates that both tIM and shared ESN reservoirs offer practical solutions for deploying multiple DL tasks on low-power resource-constrained devices. The proposed methods enable more efficient model deployment, reducing storage and computation costs while maintaining high inference performance, making them promising candidates for future and tinyML applications.
Author: Gibbs, MJ
Institution: Nottingham Trent University
Language: English
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7116598578
