Heterogeneous Resource Management for Services based on Artificial Intelligence

Machine-to-machine communication (M2M) is becoming the most significant share of wireless traffic, largely due to emerging applications in the Internet of Things (IoT) including those for smart cities and intelligent transportation systems. A large number of such applications leverage artificial intelligence (AI) through machine learning (ML) and have heterogeneous resource requirements. To this end, several novel computing and communication paradigms have been proposed, including cloud, fog, edge, and network slicing. This dissertation addresses heterogeneous resource management for AI-based services with a focus on distributed processing and IoT scenarios. Specifically, we leverage the fog and edge computing paradigms for efficient management of resources including processing, communication, and AI knowledge. First, we consider how to achieve fast and scalable deep neural network (DNN) inference involving IoT devices. Accordingly, we propose distributed techniques that collaboratively partition and offload computation under dynamic network conditions to minimize DNN inference time. Second, we develop a tool to improve the DNN inference time through fast sparse matrix-vector multiplication (SpMV), which is a major computing operation for pruned DNNs. The related data structures and algorithms are selected through a rigorous analysis of sparsity and prediction of the related performance. Next, we focus on efficient network resource utilization while providing a target service quality. In detail, we leverage slicing and Fog-RANs to improve resource utilization for generic services in 5G networks with multiple service providers. To this end, we propose a hierarchical resource scheduling mechanism named 2L-MRA to jointly allocate multiple Fog-RAN resources to network slices in two stages. Finally, we target improving the accuracy of the DNNs by developing an economic market that incentivizes different service providers to trade and combine their existing knowledge for higher model accuracy. Specifically, we devise a model based on Fisher's market for optimal knowledge sharing through transfer learning and a weight fusion technique to merge the acquired knowledge.