Neural network quantization in federated learning at the edge

Abstract

The massive amount of data collected in the Internet of Things (IoT) asks for effective, intelligent analytics. A recent trend supporting the use of Artificial Intelligence (AI) solutions in IoT domains is to move the computation closer to the data, i.e., from cloud-based services to edge devices. Federated learning (FL) is the primary approach adopted in this scenario to train AI-based solutions. In this work, we investigate the introduction of quantization techniques in FL to improve the efficiency of data exchange between edge servers and a cloud node. We focus on learning recurrent neural network models fed by edge data producers using the most widely adopted neural networks for time-series prediction. Experiments on public datasets show that the proposed quantization techniques in FL reduces up to 19× the volume of data exchanged between each edge server and a cloud node, with a minimal impact of around 5% on the test loss of the final model.