Distributed training for spatio-temporal neural networks

Abstract

For the training of spatio-temporal sensing data, most of the existing methods are centralized or combined with federated learning, but in some scenarios only fully distributed training can be performed. And the existing methods basically assume that the topology of the agent is fixed, but sometimes the topology changes dynamically. Therefore, this paper combines the (sub)gradient projection method to realize a distributed training method for spatio-temporal sensing data. Our method is not only suitable for scenarios where the agent topology is fixed, but also for scenarios where the topology changes dynamically.