Capsule Network Distributed Learning with Multi-Access Edge Computing for the Internet of Vehicles

Abstract

For realizing an intelligent transport system, a vast amount of raw image data is required to train various intelligent applications on the Internet of Vehicles (IoV). A capsule network performs well in the computer vision area with fewer model parameters compared to convolutional neural networks. Due to the small-scale model, multi-access edge computing (MEC) devices can support online training for the whole capsule network model. In this article, we propose a novel framework for MEC-based capsule network (CapsMEC) distributed learning for IoV applications. Capsules in CapsMEC are specially designed to train in a collaborative way, which relieves the network pressure in MEC and the time-consuming training time of the traditional capsule network. Experimental results prove the effectiveness of the proposed framework.