Fine-tuning large pre-trained speech recognition models on resource-constrained edge devices within federated learning frameworks presents significant challenges due to computational limitations, communication overhead, and the complexities of mixed-precision training. To address these issues, we propose FedMP-Head, a two-step method designed to optimize training efficiency while maintaining high accuracy. In the first step, we freeze the encoder of the foundation model and train only a lightweight classifier head before applying mixed-precision quantization. This approach reduces the size of global model updates to approximately 780K parameters and maintains the accuracy during the downstream task of classification. In the second step, we implement budgeted layer-wise gradient scaling and correction with mixed-precision training on the client side. This technique optimizes computational cost and memory footprint on resource-constrained clients by effectively utilizing gradients, even with limited precision and computational capabilities. FedMP-Head accelerates training, reducing convergence time compared to full-precision methods and achieving threshold accuracy in fewer rounds. Experiments on 100 resource-constrained clients highlight its improved performance and suitability for real-world federated learning on edge devices.

Recently, research on federated clustering has been actively studied to improve the performance of federated learning to solve the non-i.i.d issue. Federated clustering makes clusters with members who has similar characteristics of data which is used as inputs of federated learning, and each cluster trains an artificial intelligence model in a federated manner. However, if distances between members of a cluster configured through federated clustering is long in a network, the overhead related to federated learning becomes larger than expected and it may be lose the network cost benefits of federated learning. In this paper, we propose a DTW(Dynamic Time Warping) based federated clustering and MIP(Mixed Integer Programming)-based edge server placement in order to reduce the network overhead of federated learning caused by federated clustering under non-i.i.d setting.

Because of the activation of IoT (Internet of Things) devices due to the rapid development of recent communication technology, network traffic is currently fluctuating and increasing explosively. As existing network resource management policies are not sophisticated enough to cope with network conditions that change constantly, resource utilization can be lowered and costs can be higher. With the recent advances in deep learning techniques, network operators can manage networks intelligently. For the intelligent network, there is a technique which predict the state of network links. However, when the scale of the network increases, overall network management can be complicated. In addition, as the models of link state prediction are affected by the states of adjacent links, it is necessary to consider the spatio-temporal characteristics between links. In this paper, we propose a GCN(Graph Convolutional Neural Network)-GRU(Gated Recurrent Unit) based link state prediction technique. The proposed GCN-GRU model predicts network traffic by considering the spatio-temporal characteristics of each link state such as bandwidth, delay, and packet loss rate. Through extensive experiments on actual network traffic, the proposed GCN-GRU based link state prediction technique has shown to achieve 1.5% lower a mean absolute percentage error (MAPE) compared to a LSTM (Long Short term Memory) based link state prediction technique.