FedSFA: Federating Spikes Fired, Approximately

Over the last decade rapid advancements have been made in the domain of Artificial Neural Networks. A less known variant of these computing architectures are Spiking Neural Networks, an exceptionally energy efficient approach, which aims to simulate the firing mechanisms of biological neural systems.
Federated Learning is a distributed network training paradigm which has increasingly been gaining popularity, due to its decentralized model training approach, allowing training data to remain in its native device's storage.
The intersection of these two topics has previously been explored, with some works existing on Spiking Neural Networks being trained utilizing Federated Learning. This paper proposes FedSFA, a Federated Spiking Transformer for image classification tasks, accelerated by Spike Firing Approximation Training, attempting to tackle tasks with high degrees of data heterogeneity across participants in a federation. Moreover, a scheduler for decaying the amount of local epochs in each training round is proposed. FedSFA presents similar to marginal improvements on finetuning E-Spikeformer on the CIFAR-10/100 when compared to centralized scenarios. End-to-end training was conducted in SVHN(+1.43\%/5.18\%), and NMNIST(+1.58\%) datasets, while having a smaller resource and temporal footprint.