Abstract:
While being extensively studied in machine learning community, the problem of improving generalization in Federated Learning (FL) is still in its infancy. The main challenge stems from the heterogeneous nature of client data and the varying computational capacity of clients. Many researchers have recently linked the generalization gap to the sharpness of the landscape of the optimization model. In Foret P. et al, Sun Y. et al, and Qu Z. et al, a Sharpness-Aware Minimization (SAM) framework that seeks flat minima by penalizing sharp regions was introduced. In this thesis, we propose a SAM-like approach for improving generalization in FL settings. Unlike several existing methods that
incorporate SAM when training local models, our proposed framework penalizes the loss of the global function. To motivate our approach, we first provide a counter-example that shows that finding flat minima for local clients does not necessarily result in a flat aggregation for the global model. Furthermore, we develop an efficient sharpness-aware algorithm that adaptively computes global gradient similarity parameters for penalizing sharp regions. Harnessing these similarity parameters, a distinct sharpness penalty parameter is shared with each client. In particular, clients with varying local data distribution receive different penalty terms. We mathematically established the convergence of our suggested algorithm. Then, to demonstrate the efficiency of our algorithm, we perform several experiments on MNIST, FMNIST, and CIFAR datasets. Our results show a significant increase in generalization performance compared to existing approaches.
