Survey of edge–edge collaborative training for edge intelligence

With the rapid arrival of the Internet of Everything era, massive data resources are generated on edge sides, causing problems such as large network load, high energy consumption, and privacy security in traditional distributed training based on cloud computing. Edge computing sinks computing power resources to the edge side, forming a collaborative computing system that integrates “cloud, edge, and end,” which can meet the basic needs of real-time operations, intelligence, security, and privacy protection. With the help of edge computing capabilities, edge intelligence effectively promotes the intelligent development of the edge side, which has become a popular topic. Through our research, we found that edge collaborative intelligence is currently in a stage of rapid development. At this stage, several deep learning models are combined with edge computing, and many edge collaborative intelligent processing solutions have exploded, such as distributed training in edge computing scenarios, federated learning, and distributed collaborative reasoning based on technologies such as model cutting and early exit. The combination of a shallow breadth learning system and virtualization technology allows for quick implementation of edge intelligence, which considerably improves service quality and user experience and makes services more intelligent. As a key link of edge intelligence, edge intelligence collaborative training aims to assist or implement the distributed training of machine learning models on the edge side. However, in an edge computing scenario, the distributed training of the model must coordinate several edge nodes, and many challenges remain. Therefore, by fully investigating the existing research foundation of edge intelligent collaborative training, we focus on the challenges and solutions of edge intelligent collaborative training in edge scenarios such as equipment heterogeneity, limited equipment resources, and unstable network environments. This paper introduces and summarizes the overall architecture and core modules of edge intelligent collaborative training. The overall architecture mainly focuses on the interaction framework between edge devices. In terms of whether there is a central server role, it can be divided into two categories: parameter server centralized architecture and fully decentralized parallel architecture. The core module of edge intelligent collaborative training mainly focuses on the problem of collaborative training of a large number of edge devices for neural network models to update parameters. In terms of the role of parallel computing in model training, it is divided into data parallelism and model parallelism. Finally, the many challenges and prospects of edge collaborative training are analyzed and summarized.