Collaborative Research: CNS Core: Small: Towards Automated and QoE-driven Machine Learning Model Selection for Edge Inference

手机、无人机和机器人等边缘设备已经成为深度神经网络(DNN)推理的一个越来越重要的平台。对于边缘设备来说，从众多可能性中选择最优的DNN模型对于最大化用户的体验质量(QOE)至关重要，但边缘设备的高度异构性和不断变化的使用场景对此提出了重大挑战。目前的做法通常是为许多或所有边缘设备选择单一的DNN模型，最多只能为一小部分用户提供满意的QOE。或者，特定于设备的DNN模型优化非常耗时，并且不能扩展到大量不同的边缘设备。此外，现有的方法侧重于优化用于边缘推理的特定客观度量，这可能不能转化为对用户实际QOE的改善。通过利用机器学习的预测能力，并将用户保持在循环中，该项目提出了一个自动化的、可扩展的设备级DNN模型选择引擎，用于QOE-最优边缘推理。具体地说，该项目包括两个方面：第一，它利用在线学习来预测每个边缘设备的QOE，使部署阶段的DNN模型选择自动化；第二，它构建一个运行时QOE预测器，并根据运行时上下文信息自动选择最优的DNN模型。它可以将启用DNN的智能的好处带给更多资源受限的边缘设备，并提供最佳的QOE。此外，它为边缘推理提供了新的观察、见解和原理，催化了DNN模型设计向以用户为中心的新范式的转变。这个项目还提供了新的机会来改进课程设计，吸引学生，特别是代表不足的少数族裔，从事科学、技术、工程和数学领域。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Edge devices, such as mobile phones, drones and robots, have been emerging as an increasingly more important platform for deep neural network (DNN) inference. For an edge device, selecting an optimal DNN model out of many possibilities is crucial for maximizing the user’s quality of experience (QoE), but this is significantly challenged by the high degree of heterogeneity in edge devices and constant-changing usage scenarios. The current practice commonly selects a single DNN model for many or all edge devices, which can only provide a satisfactory QoE for a small fraction of users at best. Alternatively, device-specific DNN model optimization is time-consuming and not scalable to a large diversity of edge devices. Moreover, the existing approaches focus on optimizing a certain objective metric for edge inference, which may not translate into improvement of the actual QoE for users. By leveraging the predictive power of machine learning and keeping users in a loop, this project proposes an automated and scalable device-level DNN model selection engine for QoE-optimal edge inference. Specifically, this project includes two thrusts: first, it exploits online learning to predict QoE for each edge device, automating deployment-stage DNN model selection; and second, it builds a runtime QoE predictor and automatically selects an optimal DNN model given runtime contextual information.This project represents an important departure from and an essential complement to the current practices in DNN model optimization. It can bring the benefits of DNN-enabled intelligence to many more resource-constrained edge devices with an optimal QoE. Additionally, it provides novel observations, insights and principles for edge inference, catalyzing the transformation of the design of DNN models into a new user-centric paradigm. This project also enables new opportunities to improve curriculum design and attract students, especially under-represented minorities, to engage in science, technology, engineering, and mathematics fields.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.