Uncertainty-Aware Federated Learning Model for Distributed Scientific Computing

Abstract

Federated learning has become an enticing paradigm of distributed scientific computing by allowing collaborative training of models without sharing central information. Nevertheless, the majority of existing methods emphasize mainly on predictive performance and fail to consider the essential factor of uncertainty estimation that is vital in good decision-making in the sciences. This ambiguity constrains the reliability of models that are used in data sensitive settings with high complexity. To overcome this difficulty, this paper will present an uncertainty-conscious federated learning model that incorporates the use of probabilistic reasoning in the distributed training. The framework proposed integrates the principles of Bayesian learning to include epistemic and aleatoric uncertainty to take stronger and interpretable predictions among heterogeneous clients. The model can be used effectively when the data is non-IID, which is a usual situation in real-life scientific computing problems. To maintain uncertainty information when updating the world, a modified strategy of aggregation is presented. Core metrics like the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and accuracy, are used to measure the performance of the proposed approach, and a comprehensive evaluation of the predictive ability is made. Experimental evidence shows that the proposed model is not only competitive with respect to the traditional federated learning approaches but also yields better reliability with a well-calibrated uncertainty estimate. These results underscore its possible application in improving the decision-making in distributed scientific setups.