distributed deep learning, communication-efficient, convergence rate, connected dominating set, decentralized learning