As noted before, the real-time edge intelligence can be achieved by the collaborative learning across different edge nodes. In this chapter, we turn to study the collaborative learning between the edge and the cloud to achieve real-time edge intelligence. More specifically, we propose a distributionally robust optimization (DRO) approach which provides a natural mechanism to enable the synergy between the local data processing and the cloud knowledge transfer. In particular, we construct a distributional uncertainty set to capture the inaccuracy of local data processing, and build another distribution uncertainty model corresponding to the cloud knowledge transfer. Then, we recast the edge learning problem as a DRO problem which takes into account the two distributional uncertainty constraints.