Efficient Federated Learning: a Tsetlin Machine-based approach

Abstract

Federated Learning (FL) is a paradigm for collaboratively training a global model across distributed client devices, in which training occurs on-device and only model parameters are communicated to a central server, thereby preserving data privacy. However, its adoption in practice is constrained by communication overhead, computational limitations, and statistical data heterogeneity. Motivated by the logic-based formulation and bit-level representation of Tsetlin Machines (TMs), which enable a low-complexity design, this dissertation addresses FL challenges and proposes a family of TM-based FL frameworks as lightweight, resource-efficient alternatives to conventional deep neural network (DNN)–based FL approaches.The first contribution, FedTM, introduces the first FL framework to capitalise on the bit-level structure and low-complexity learning dynamics of TMs in a federated setting. Building on this, the second contribution, CS-pFedTM, presents a personalised FL framework that addresses statistical heterogeneity through adaptive clause allocation, driven by data heterogeneity and communication constraints. Finally, FedTMOS proposes an efficient one-shot FL framework that eliminates server-side training by clustering local models and constructing a global model in a single communication round via class-adaptive parameter organisation.These methods establish a set of TM-based FL algorithms that are robust to statistical heterogeneity and efficient in terms of communication, computation and memory. Extensive evaluations on image and audio datasets demonstrate competitive performance–resource trade-offs relative to state-of-the-art baselines. While TM research remains at an early stage, this dissertation lays the foundations for TM-based FL and highlights promising directions for future work, including on-device implementations and the exploration of pre-training and transfer learning strategies.

More information

Identifiers

ORCID for Shannon Shi Qi How: orcid.org/0000-0002-5995-2343

ORCID for Geoff Merrett: orcid.org/0000-0003-4980-3894

ORCID for Jonathon Hare: orcid.org/0000-0003-2921-4283

Catalogue record

Export record

Altmetrics