Date of Award

Spring 2021

Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Hu, Wenjun


Deep learning driven intelligent edge has already become a reality, where millions of mobile, wearable, and IoT devices analyze real-time data and transform those into actionable insights on-device. Typical approaches for optimizing deep learning inference mostly focus on accelerating the execution of individual inference tasks, without considering the contextual correlation unique to edge environments and the statistical nature of learning-based computation. Specifically, they treat inference workloads as individual black boxes and apply canonical system optimization techniques, developed over the last few decades, to handle them as yet another type of computation-intensive applications. As a result, deep learning inference on edge devices still face the ever increasing challenges of customization to edge device heterogeneity, fuzzy computation redundancy between inference tasks, and end-to-end deployment at scale. In this thesis, we propose the first framework that automates and scales the end-to-end process of deploying efficient deep learning inference from the cloud to heterogeneous edge devices. The framework consists of a series of service abstractions that handle DNN model tailoring, model indexing and query, and computation reuse for runtime inference respectively. Together, these services bridge the gap between deep learning training and inference, eliminate computation redundancy during inference execution, and further lower the barrier for deep learning algorithm and system co-optimization. To build efficient and scalable services, we take a unique algorithmic approach of harnessing the semantic correlation between the learning-based computation. Rather than viewing individual tasks as isolated black boxes, we optimize them collectively in a white box approach, proposing primitives to formulate the semantics of the deep learning workloads, algorithms to assess their hidden correlation (in terms of the input data, the neural network models, and the deployment trials) and merge common processing steps to minimize redundancy.