Year: 2022
Type: dissertation
Abstract: With an increasing number of interconnected devices that produce large amounts of data, the need for distributed machine learning techniques arises. The connections be- tween the client devices often imply that the devices have similar statistical properties which result in the devices forming a clustered similarity graph. Networked federated learning is a machine learning technique that allows the training of personalized learning models over a similarity graph of client devices without moving the local data from the devices to a centralized server. However, the influence of the structure of the clustered similarity graph on the performance of federated learning methods has not been studied in detail. This thesis proposes a fully distributed networked federated learning algorithm (NFLA) which uses the gradient descent method and an adaptive node-wise learning rate for each client device to learn personalized linear models. The NFLA is applied to stochastic block models with synthetic local datasets to study the effect of the cluster structure on the accuracy and the convergence rate of the algorithm. Furthermore, NFLA is compared to a clustered federated learning algorithm that does not require knowledge of the similarity graph to study whether knowing the graph structure improves the accuracy of federated learning methods. The results show that the number of clusters has the largest effect on the accuracy of NFLA, followed by edges between clusters and the edges within clusters had the least influence. In contrast, the type of edge did not affect the convergence rate of the NFLA. The convergence rate was influenced the most by the total number of edges followed by the number of clusters. The comparison experiment showed that using the similarity graph can increase the accuracy and the consistency of federated learning methods.
Source: Aaltodoc (Aalto University)
Author: Lillfors, Joona
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112437701
