Federated Learning with Spiking Neural Networks in Heterogeneous Systems

With the advances in IoT and edge-computing, Federated Learning is ever more popular as it offers data privacy. Low-power spiking neural networks (SNN) are ideal candidates for local nodes in such federated setup. Most prior works assume that the participating nodes have uniform compute resources, which may not be practical. In this work, we propose a federated SNN learning framework for a realistic heterogeneous environment, consisting of nodes with diverse memory-compute capabilities through activation-checkpointing and time-skipping that offers ~$4\times$ reduction in effective memory requirement for low-memory nodes while improving the accuracy upto 10% for non-independent and identically-distributed data.