Resource-aware distributed edge intelligence

Distributed edge intelligence is emerging as a key enabler for artificial intelligence (AI)-native wireless networks, where learning, adaptation, and personalization are performed collaboratively across a large number of devices without sharing their raw data. However, practical deployments face multiple resource constraints, including dynamic wireless connectivity, limited bandwidth, heterogeneous computation resources, and memory budgets, which tightly couple the learning performance with communication and computation overhead. This thesis develops a unified learning–communication–computation co-design framework for distributed edge intelligence, and proposes scalable algorithms that jointly optimize the learning performance and efficiency. First, we study decentralized federated learning (DFL) over device-to-device (D2D) wireless networks and propose a graph neural network (GNN)-based framework that jointly determines the neighbor selection and resource allocation, enabling distributed implementation and improving efficiency under dynamic wireless channel conditions. Second, we develop an adaptive split federated learning (ASFL) framework that exploits the computation resources of the central server to train part of the model and enables adaptive model splitting as well as resource allocation. Then, we design an online optimization algorithm to address the coupling of the model splitting and resource allocation decisions. Third, to enable federated fine-tuning of large language models (LLMs) under memory constraints, we propose a zeroth-order federated fine-tuning framework with heterogeneous block activation, together with a multi-objective optimization approach that characterizes the trade-off between the convergence rate and memory usage. Finally, we improve the efficiency and stability of federated parameter-efficient fine-tuning (PEFT) by proposing low-rank Gram-matrix aggregation with Procrustes alignment, which mitigates ill-posed aggregation under low-rank parameterizations and substantially reduces client–server communication. Extensive experiments and simulations demonstrate that the proposed frameworks achieve favorable trade-offs among accuracy, delay, energy consumption, memory usage, and communication overhead across a variety of vision and language tasks.