Deep learning enabled by neural networks has been proven to be an effective Artificial Intelligence (AI) algorithm in sophisticated applications. The algorithm is normally divided into two phases: learning phase and inference phase. In this research, we assume the learning phase is already accomplished offline and focus on expediting the inference phase by replacing the centralized processing of Cloud with the distributed processing of Fog. In our approach, inference algorithms in AI are distributed to multiple layers of Fog networking, constructed from oneM2M Middle Nodes. We verify the performance improvement of our proposed distributed AI/Fog system by comparing it against a Cloud-centric system based on a use case of smart shopping mall.