Gradient Scale Monitoring for Federated Learning Systems

As the computational and communicational capabilities of edge and IoT devices grow, so do the opportunities for novel Machine Learning solutions. This leads to an increase in popularity of Federated Learning (FL), especially in cross-device settings. However, while there is a multitude of ongoing research works analyzing various aspects of the FL process, most of them do not focus on issues of operationalization and monitoring. For instance, there is a noticeable lack of research in the topic of effective problem diagnosis in FL systems. This work begins with a case study, in which we have intended to compare the performance of four selected approaches to the topology of FL systems. For this purpose, we have constructed and executed simulations of their training process in a controlled environment. We have analyzed the obtained results and encountered concerning periodic drops in the accuracy for some of the scenarios. We have performed a successful reexamination of the experiments, which led us to diagnose the problem as caused by exploding gradients. In view of those findings, we have formulated a potential new method for the continuous monitoring of the FL training process. The method would hinge on regular local computation of a handpicked metric - the gradient scale coefficient (GSC). We then extend our prior research to include a preliminary analysis of the effectiveness of GSC and average gradients per layer as potentially suitable for FL diagnostics metrics. In order to perform a more thorough examination of their usefulness in different FL scenarios, we simulate the occurrence of the exploding gradient problem, vanishing gradient problem and stable gradient serving as a baseline. We then evaluate the resulting visualizations based on their clarity and computational requirements. We introduce a gradient monitoring suite for the FL training process based on our results.