Semisupervised Distributed Learning With Non-IID Data for AIoT Service Platform

Thanks to the advances in wireless communication and machine learning technologies, we can envision a novel AIoT (AI + IoT) service platform that collects video data from the individuals’ edge devices. Then, it transforms the video data into useful information, providing services to IoT or smart city applications. However, collecting raw video data directly to the cloud server is merely possible due to network bandwidth limitations and data privacy concerns. One possible solution is to adopt federated learning, which enables edge devices to collaboratively train a shared model without sending the raw data to the cloud. Unfortunately, this scheme cannot directly be applied to the targeted scenario since it assumes labeled data for training, and only at the cloud, we have the human power and time to label the video data. Thus, to tackle those issues, we propose an edge learning system based on semisupervised learning and federated learning technologies. The system trains AI models at edge devices using an improved semisupervised learning scheme and periodically uploads the training results to the cloud server to form a single model by adapting the federated learning technology. Then, we observe that in the real world, the data on the end devices are nonindependent and identically distributed (non-IID) such that it may cause weight divergence during training and result in a considerable decrease in the model performance. Therefore, we propose a new operation called federated swapping (FedSwap) to replace partial federated learning operations based on a few shared data during federated training to alleviate the adverse impact of weight divergence. We evaluate our system on both image classification using the state-of-the-art benchmark data and object detection using real-world video data. The experimental results show that the proposed system can have up to 5.9% higher accuracy of object detection for the video analysis applications by fully utilizing unlabeled data, compared with the situation that only labeled data are used. Moreover, the proposed FedSwap can improve the accuracy of image classification by 3.8% and the object detection task by 1.1%.