主题：Improving Training Efficiency and Quality in Federated Learning
嘉宾：何黎刚 英国华威大学计算机系 Reader
时间：2023年4月18日 上午10:00 – 11:00
In this talk, three piece of work we have conducted in FL will be presented. First, a semi-asynchronous FL protocol called SAFA is presented to improve the training efficiency of FL. Second, A FL scheme called hybridFL is presented. HybridFL further enhances the efficiency of SAFA by taking the reliability of FL clients into account. Moreover, hybridFL extends SAFA from a two-layer (i.e., client/server) FL scheme to a three-layer one in mobile-edge-cloud systems, enabling the support of even larger-scale FL training. Finally, in hybridFL, the clients are randomly selected to participate in FL training. If some clients have low quality data, the fact that they have equal opportunities to contribute to the final model may hurt the model quality. To address this issue, a selective FL scheme is proposed, in which data quality can be quantified and the clients with lower-quality data have fewer chances to be selected for training.