FedTMOS: efficient one-shot federated learning with Tsetlin machine

Abstract

One-Shot Federated Learning (OFL) has emerged as a promising alternative to address the communication bottleneck. OFL restricts communication to a single round, thus minimizing communication errors, cost and reducing the risk of interference caused by iterative updates [1]. Current OFL methods that rely on Knowledge Distillation (KD) and ensemble learning aggregate local models into an ensemble before distilling it into a global model. A key challenge with these methods is their dependence on public datasets, which may be unsuitable for certain tasks [1]. Datafree methods with generative models [2], suffer from additional computational overhead. Neuron matching and model fusion techniques, eliminates the need for server-side training but struggle with performance when models are trained on heterogeneous data distributions due to misalignment of models [3]. On the client side, existing methods rely on Deep Neural Networks (DNNs), which are resource-intensive and impractical for clients with limited computational capabilities, such as edge devices. Therefore, we proposed a solution based on the Tsetlin Machine (TM) for efficient OFL [4], [5].

More information

Identifiers

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

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

Catalogue record

Export record