Communication Efficient Decentralized Learning Over D2D Network: Adaptive Relay Selection and Resource Allocation

Device-to-device (D2D)-assisted decentralized learning has been proposed for collaborative model training without the parameter server while protecting the data privacy. However, in such task-oriented network, a large latency would be aroused due to the limited communication resource and straggling D2D links. To tackle these challenges, in this letter, we take into consideration the D2D relay in the decentralized learning system. The device can transmit the local model to others over relay links, thereby alleviating the effect of straggling D2D links. Then, a joint relay selection and spectrum allocation algorithm is proposed to minimize the learning latency while guaranteeing the leaning performance. Finally, thorough tests are carried out to show the effectiveness of the proposed algorithm. The results show that the learning latency can be reduced while maintaining the convergence rate and learning accuracy, as compared against the traditional methods.