Fast Convergence Federated Learning with Adaptive Gradient: An Application to Mental Healthcare Monitoring System

Nowadays, there is increasing demand for mental health monitoring systems to enable disease diagnoses, such as anxiety and depression. However, the privacy concerns for sensitive data impede its wide adoption. To protect data privacy, federated learning (FL) is proposed to enable decentralized collaborative model learning without sharing sensitive data. Though, FL training process can be slowed with the non-Independent-and-Identically-Distributed (non-IID) datasets across participating clients, causing extra communication costs. In this paper, we propose the FL adaptive gradient optimization method to accelerate the convergence under the context of non-IID training. As the reference direction for parameter update, the gradient has a great impact on the convergence performance throughout the training. By adaptively modifying the local gradients according to the global gradient, we reduce the local parameter divergence to enable robust training and fast convergence. Meanwhile, as an application to our FL optimization algorithm, a novel sleep monitoring system is proposed to detect potential depression. Experiments demonstrate that with our proposed method, faster convergence and higher accuracy can be realized compared to commonly adopted Federated Averaging (FedAVG) and other adaptive optimization methods, which effectively save communication costs.