Machine Learning for 5G/6G

Distributed Learning schemes have emerged as promising alternatives to traditional centralized machine learning, especially in 5G environments where data is generated at the edge. This thesis investigates the impact of non-Independent and Identically Distributed (non-IID) data on the performance of various learning schemes, Federated and Decentralized Learning, using the UCI-HAR dataset as a proof-of-concept. We developed a scalable pipeline which is able to generate different non-IID distributions, We develop a modular pipeline capable of generating multiple non-IID distributions, simulating diverse network topologies, and evaluating model performance under varied hyperparameters and distance metrics. Our results show that when the heterogeneity in the data increases, centralized aggregation methods fail to converge rapidly, and over a certain epoch, they are suboptimal solutions. In contrast, decentralized approaches perform good under non-IID distribution of the data. As we said, in this paper, we developed two main distributed learning schemes, however, some upgraded schemes were planned and can be developed in the future works. One of them is E-Tree Learning, which can be seen as a hierarchal version of Federated Learning, and an extension of Decentralized Learning, named Reduced Decentralized learning, which tries to carefully remove the network links with different metrics, and can reduce communication overhead, latency, and energy consumption, with minimum decrease in the performance. We finish the article by reviewing the obtained results, and pointing the main limitations and challenges, and possible enhancements for future works. The most important enhancements are working with a larger dataset, adding more metrics for measuring the performance of the learning schemes, and trying to reduce the gap between our simulation environment and real-world scenarios.