SpikeFed: Federated Training of Spiking Neural Networks for Event-Based Vision

In this paper, we investigate the distributed training of spiking neural networks (SNNs) for event-based vision data. We present SpikeFlower, a federated learning (FL) framework for training SNNs, which is implemented using the Flower FL platform and integrates event-based vision models to support collaborative training across multiple clients without centralized data sharing. We evaluate the framework on the CIFAR10-DVS dataset, focusing on convergence behavior under varying hyperparameter configurations, including the number of clients, local training epochs, and different batch sizes. Our results show that SNNs can be effectively trained in decentralized settings. We further observe that the number of local training epochs introduces a trade-off between convergence speed and stability in federated scenarios, highlighting the balance between local computation and global aggregation. This work provides a practical foundation for studying SNNs in distributed learning environments and opens new directions at the intersection of federated learning and neuromorphic computing.