Edge Computing Assisted Deep Learning for Mobile Intelligence

Artificial Intelligence (AI) faces limitations in conventional deployments: cloud-centric inference/training often suffers from latency, while purely mobile solutions are constrained by compute, storage, and energy. Edge computing offers a practical middle ground by placing scalable resources closer to data sources, enabling low-latency, distributed, and more reliable Mobile AI.This thesis investigates how to realize effective collaboration across the device–edge–cloud hierarchy under three recurring bottlenecks: heterogeneous data and capabilities, volatile system dynamics, and costly cross-node tensor transmission. We develop three complementary collaboration paradigms.