PhD Oral Exam - Weiyang Qian, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The advent of Mobile Vision Analytics (MVA) is transforming applications from intelligent surveillance to immersive Augmented Reality (AR) in next-generation wireless networks. These systems, however, are constrained by the limited computation and energy of user devices and the stringent latency requirements of real-time video analytics. Edge computing provides a promising solution but introduces new challenges in wireless resource allocation, distributed task orchestration, and multi-tenant GPU scheduling. This thesis presents a comprehensive study on the modeling, design, and optimization of edge-assisted MVA systems.

First, we develop mathematical frameworks that capture the end-to-end pipeline of edge-assisted video analytics, integrating wireless transmission, task offloading, and concurrent DNN execution. Using queuing theory and stochastic modeling, we quantify trade-offs among latency, accuracy, energy, and frame drop rates. Next, we design load-aware orchestrators that balance workloads across multiple edge servers, dynamically scaling processing instances to maintain quality of service under fluctuating traffic. Building upon this, we propose cross-layer algorithms for joint allocation of Physical Resource Blocks (PRBs) and computing resources, ensuring fair and efficient offloading under dynamic wireless conditions.

To support multi-edge deployments with heterogeneous workloads, we develop a reinforcement learning based task orchestration framework that simultaneously determines frame admission, server placement, and migration, leveraging queue-aware state representations and reward functions that encode accuracy-latency tradeoffs. Finally, recognizing the increasing prevalence of concurrent multi-model inference at the edge, we propose GPU-aware scheduling and hardware partitioning strategies that enhance throughput and reduce interference among deep neural networks.

Extensive numerical evaluations and simulations validate the proposed methods, showing significant improvements in end-to-end latency, load balancing, and resource utilization. This thesis establishes a unified foundation for the design and optimization of future real-time edge-assisted mobile vision.