Adaptive Top-K in SGD for Communication-Efficient Distributed Learning in Multi-Robot Collaboration | Synapse