Abstract
<p>Edge intelligence aims to leverage available computing resources at the network edge, e.g., smart end-devices and edge computing platforms, to deliver high-quality AI service to end-users. This dissertation focuses on model selection schemes to automate the AI applications to identify the best-fit Deep Neural Networks (DNN) for providing better and faster AI capabilities.<br />
We first study automated customization of on-device DNN inference, and the goal is to enhance user QoE by picking the best-fit DNN for mobile users under different usage scenarios. The core of our method is a DNN selection module that learns user QoE patterns online for identifying the best-fit DNN. As the DNN selection module learns user QoE patterns online, it frequently solicits QoE from users which incurs inconvenience. We then design feedback solicitation schemes to mitigate solicitation costs. <br />
We then establish a framework to implement DNN ensemble techniques on edge computing platforms. While the DNN ensemble has been providing state-of-the-art performance for many AI applications, its high computing complexity makes it difficult to be implemented efficiently. A DNN ensemble selection problem is studied to handle these issues. We propose a novel algorithm called Automated DNN Ensemble to learn the in-practice performance of individual DNNs over time and recruit DNN members into the ensemble based on an empirical rule that considers the accuracy and diversity of individual DNNs. <br />
However, it is difficult to guarantee the performance of empirical ensemble selection rules because the interdependency between DNN members is hard to be characterized. We then propose an algorithm called Neural Ensemble to learn the performance of each possible DNN ensemble directly, such that the optimal DNN ensemble can be identified. However, this method is often put off because the decision space of DNN ensembles increases exponentially with the number of candidate DNNs. We, therefore, propose NeuE-S which utilizes the contexts of DNN ensembles and mine their similarity information to handle the scalability issue.</p>