Resource-Aware Distributed Machine Learning for Edge Intelligence

The rapid growth of intelligent edge applications—including industrial automation, autonomous systems, smart healthcare, immersive multimedia, and large-scale sensing—has created unprecedented demand for deploying advanced machine learning models close to data sources. Modern deep learning models provide remarkable predictive and decision-making capabilities, but their high requirements in computation, memory, communication bandwidth, and energy consumption make their deployment challenging in resource-constrained edge environments. Edge devices, mobile systems, and embedded sensing platforms often operate under strict limitations in battery capacity, processing capability, storage, and wireless connectivity, making conventional cloud-centric machine learning approaches insufficient for many latency-critical and privacy-sensitive applications.

Traditionally, machine learning models are trained and executed in centralized cloud infrastructures, where virtually unlimited computational resources enable scalable model training and complex analytics. However, cloud-centric intelligence suffers from several fundamental limitations, including high end-to-end latency, increased backhaul bandwidth consumption, vulnerability to network disruptions, and potential exposure of sensitive user data during transmission and remote processing. These limitations become particularly critical in real-time applications such as autonomous mobility, industrial control, healthcare monitoring, and immersive communication systems.

Recent advances in distributed edge intelligence aim to overcome these limitations by distributing learning and inference tasks across end devices, edge servers, and cloud infrastructures. Emerging paradigms such as Federated Learning, split learning, collaborative inference, model compression, adaptive pruning, and knowledge distillation enable machine learning models to operate efficiently under heterogeneous resource constraints while preserving privacy and reducing communication overhead. In parallel, learning-based optimization methods, including reinforcement learning and multi-agent learning, provide adaptive mechanisms for dynamic resource allocation, model partitioning, client selection, and service orchestration in highly dynamic edge environments.

This research area investigates how machine learning tasks can be efficiently distributed across heterogeneous edge infrastructures while jointly optimizing learning accuracy, energy consumption, latency, communication overhead, scalability, and privacy. Representative research topics include resource-aware federated learning, personalized distributed learning, split inference, collaborative training across heterogeneous GPU clusters, adaptive model compression, and AI-driven orchestration for next-generation edge and 6G systems.

Related Literature:

[1] Jia, Ninghui, Zhihao Qu, Baoliu Ye, Yanyan Wang, Shihong Hu, and Song Guo. "A comprehensive survey on communication-efficient federated learning in mobile edge environments."  IEEE Communications Surveys & Tutorials  27, no. 6 (2025): 3710-3741.
[2] Tan, Y., Tan, C., Mi, Z. and Chen, H., 2025, March. Pipellm: Fast and confidential large language model services with speculative pipelined encryption. In Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems.
[3] Zhou, Zhi, Xu Chen, En Li, Liekang Zeng, Ke Luo, and Junshan Zhang. "Edge intelligence: Paving the last mile of artificial intelligence with edge computing." Proceedings of the IEEE 107, no. 8 (2019): 1738-1762.
[4] Samikwa, Eric, Antonio Di Maio, and Torsten Braun. "Disnet: Distributed micro-split deep learning in heterogeneous dynamic IoT." IEEE internet of things journal 11, no. 4 (2023): 6199-6216.
[5] Samikwa, Eric, Antonio Di Maio, and Torsten Braun. "DFL: Dynamic federated split learning in heterogeneous IoT."  IEEE transactions on machine learning in communications and networking  2 (2024): 733-752.