From few-shot learning to fast adaptation: deep learning-based adaptive wireless networks
File(s)
Author(s)
Wang, Ouya
Type
Thesis
Abstract
Deep learning (DL) shows immense promise for advancing sixth-generation (6G) wireless systems, but its widespread application is hindered by the dynamic nature of wireless environments. Frequent environmental changes require repeated retraining of DL models, which is both data-intensive and time-consuming. This thesis addresses two critical challenges that arise when DL-based wireless systems adapt to new environments: data-intensive retraining and low-latency adaptation.
To address the challenge of data-intensive retraining, this thesis develops data-driven few-shot learning (FSL) techniques that learns transferable experience from multiple known environments to reduce the data demands for adapting to new ones. An initial contribution is a DL-based channel estimator that uses attention mechanisms to learn such experience, enabling strong adaptation to new environments with few pilot blocks. Furthermore, an advanced two-phase FSL scheme is proposed, in which the model is splitted into two parts to learn environment-agnostic and environment-specific features to more precisely capture environmental features. This scheme is solved via both Alternating Direction Method of Multipliers (ADMM)-based optimization and model-structure design, applied to the OFDM receiver.
To achieve low-latency adaptation, this thesis develops fast adaptation techniques that are efficient in data, computation, parameters and storage. Three complementary strategies are developed: (1) Data-driven FSL: an efficient implementation of the two-phase FSL scheme, applied to mmWave beamforming; (2)Wireless domain knowledge: a model-driven multi-user precoding system that leverages the classic WMMSE algorithm as wireless domain knowledge, further enhanced with fixed point theory and meta-learning to reduce training iterations and data needs; (3) Advanced ADMM-based DL optimiers: The DL optimizers that accelerate convergence for general DL applications and reduce training iterations, thereby speeding up the system adaptation.
To address the challenge of data-intensive retraining, this thesis develops data-driven few-shot learning (FSL) techniques that learns transferable experience from multiple known environments to reduce the data demands for adapting to new ones. An initial contribution is a DL-based channel estimator that uses attention mechanisms to learn such experience, enabling strong adaptation to new environments with few pilot blocks. Furthermore, an advanced two-phase FSL scheme is proposed, in which the model is splitted into two parts to learn environment-agnostic and environment-specific features to more precisely capture environmental features. This scheme is solved via both Alternating Direction Method of Multipliers (ADMM)-based optimization and model-structure design, applied to the OFDM receiver.
To achieve low-latency adaptation, this thesis develops fast adaptation techniques that are efficient in data, computation, parameters and storage. Three complementary strategies are developed: (1) Data-driven FSL: an efficient implementation of the two-phase FSL scheme, applied to mmWave beamforming; (2)Wireless domain knowledge: a model-driven multi-user precoding system that leverages the classic WMMSE algorithm as wireless domain knowledge, further enhanced with fixed point theory and meta-learning to reduce training iterations and data needs; (3) Advanced ADMM-based DL optimiers: The DL optimizers that accelerate convergence for general DL applications and reduce training iterations, thereby speeding up the system adaptation.
Version
Open Access
Date Issued
2025-09-20
Date Awarded
2026-03-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Li, Geoffrey
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)