I. Introduction
Artificial intelligence is perceived as the core engine of science innovation, economic development, and industrial reformation and it has triggered the recent evolution to future networks. As the workhorse of artificial intelligence, deep neural networks (DNN) have become the first choice for computer vision applications due to their prominent performance and flexibility. However, the harsh requirement for precisely-annotated data has largely constrained the wide application of deep learning in future mission-critical industrial networks. It is generally known that the collection and annotation of large-scale data are extremely costly and time-consuming in some professional industries (e.g. healthcare, finance, and manufacture). Therefore, an important research trend in the field of artificial intelligence is semi-supervised learning (SSL) [1]–[4].