Federated Learning with Exponentially Weighted Moving Average for Real-Time Emotion Classification

Abstract

Federated learning (FL) allows to develop a powerful shared prediction model while preserving the users’ privacy by keeping the data local. In particular it is a useful framework to use resource-constrained edge computing devices, such as mobile and IoT devices (wearable sensors), as local clients. In FL, the local clients update their model by replacing it with the current global model, which may cause performance degradation due to the previous local model information is lost. We propose an exponentially weighted moving average (EWMA) for local model updates at the local client to address this issue and improve model performance. This work focuses mainly on improving our previously developed federated learning-based real-time emotion classification from multi-modal data streams Fed-ReMECS with a local model update by EWMA, called Fed-ReMECS-U. The experiment is carried out with the help of the widely used multi-modal benchmark DEAP dataset for emotion classification. Experiments showed that the accuracy of the proposed method has improved when compared to the former Fed-ReMECS.