Adaptive machine learning techniques for edge computing
File(s)
Author(s)
Huang, Yushan
Type
Thesis
Abstract
Edge computing, by deploying computational capabilities to terminal devices near data sources (e.g., sensors, wearables, microcontrollers), significantly reduces latency and privacy risks, providing critical support for real-time sensing scenarios such as healthcare monitoring and industrial IoT. However, the widespread distribution of IoT devices and their resource constraints pose multidimensional challenges to the adaptation of edge computing: these edge devices are typically deployed in complex environments, resulting in data that may be state-varying and noisy; the local tasks can also be heterogeneous, requiring personalized solutions; edge computing tasks are generally energy-sensitive and often constrained by limited computational capabilities.
This thesis explores to improve adaptation for edge computing from three perspectives: data, model, and system. At the data aspect, we propose an information-theory-based feature extraction framework for multivariate, multi-source, state-varying, and noisy time-series data, achieving highly robust and privacy-preserving edge data representation. At the model aspect, we aim to address data drift issues and achieve efficient on-device personalization. We first utilize mobile devices (such as Raspberry Pi) as the experimental devices, design a lightweight targeted fine-tuning method that dynamically maps data drift types to model blocks. We then utilize more resource-constrained MCUs as the experimental devices, propose MicroT, which employs techniques including self-supervised learning, knowledge distillation, and MCU-designed early-exiting. Finally, at the system aspect, we further explore the MCU with AI accelerators, establish an energy and power characterization for related operations involved in real-world applications.
Our research covers low-to-mid-performance edge devices, including mobile devices (e.g., Raspberry Pi), ultra-resource-constrained MCUs, and MCUs equipped with AI accelerators. The experimental datasets include image and time-series data. The research theoretically and practically validates the necessity of improving the adaptation for edge computing from data, model, and system aspects, establishing a methodological foundation and technical benchmarks for building efficient, robust, and energy-sustainable edge computing.
This thesis explores to improve adaptation for edge computing from three perspectives: data, model, and system. At the data aspect, we propose an information-theory-based feature extraction framework for multivariate, multi-source, state-varying, and noisy time-series data, achieving highly robust and privacy-preserving edge data representation. At the model aspect, we aim to address data drift issues and achieve efficient on-device personalization. We first utilize mobile devices (such as Raspberry Pi) as the experimental devices, design a lightweight targeted fine-tuning method that dynamically maps data drift types to model blocks. We then utilize more resource-constrained MCUs as the experimental devices, propose MicroT, which employs techniques including self-supervised learning, knowledge distillation, and MCU-designed early-exiting. Finally, at the system aspect, we further explore the MCU with AI accelerators, establish an energy and power characterization for related operations involved in real-world applications.
Our research covers low-to-mid-performance edge devices, including mobile devices (e.g., Raspberry Pi), ultra-resource-constrained MCUs, and MCUs equipped with AI accelerators. The experimental datasets include image and time-series data. The research theoretically and practically validates the necessity of improving the adaptation for edge computing from data, model, and system aspects, establishing a methodological foundation and technical benchmarks for building efficient, robust, and energy-sustainable edge computing.
Version
Open Access
Date Issued
2025-04-09
Date Awarded
2025-07-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Haddadi, Hamed
Sponsor
Engineering and Physical Sciences Research Council
Medical Research Council (Great Britain)
Alzheimer's Society
Great Ormond Street Hospital (London, England)
China Scholarship Council
Grant Number
EP/W031892/1
EP/W005271/1
UKDRI7002
21PP30
202106160008
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)