A Review of Multi-Objective and Multi-Task Federated Learning Approaches

The paper aims to overview two emerging areas of federated learning techniques: Multi-objective federated learning (MOFL) and Multi-task federated learning (MTFL). These two techniques will be described and compared, and their applications will be discussed. In conclusion, the possible future applications of MOFL and MTFL will be addressed. In recent decades, we have witnessed an enormously growing amount of data on many IoT (Internet of Things) devices spread across the network. Data on each device is subject to privacy laws. Hence, using data from many different devices for training by machine learning algorithms in the cloud is nearly impossible and poses new challenges. Federated learning tries to cope with this problem by decentralised learning, where each device (client) owns its copy of the machine learning algorithm (most often neural network), trains its private data on this model and sends the results to aggregation on the server. The server collects model parameters from different clients, creates a common model, and sends new parameters back to each participating client. This work introduces, compares, and gives an overview of work made using two essential techniques for solving optimisation problems in federated learning: multi-object learning and multi-task learning. Both methods were known before, but their application in federated learning is relatively new.