Collaborative Research: III: Small: Efficient and Robust Multi-model Data Analytics for Edge Computing

点击翻译按钮获取中文摘要

Many advanced edge computing applications rely on large-scale data analysis for high-level decision making. Edge computing makes computing faster and more efficient because it takes place near the physical location of either the user or the data rather than sending all the information to the cloud. For example, augmented reality/virtual reality (AR/VR) applications utilize data from high-definition sensors (e.g., cameras, motion sensors, and microphones) to enable accurate and robust human-computer interactions. Drones and electric vehicles perform tracking, adjustments and obstacle recognition and avoidance via analyzing data at the level of the vehicle. However, the current ability to understand and manage various high-dimensional sensing data is obscured by significant knowledge and data gaps due to the heterogeneous edge device and environments, hindering the building of precise models for emerging edge computing applications using data analytics. One important trend in edge computing is utilizing artificial intelligence (AI) to extract complex knowledge from various sensor measurements for precise modeling. However, most edge devices have limited computing and memory resources, making it challenging to perform sophisticated data analytics using AI while satisfying the time requirements of most applications. Therefore, a heuristic data analytic framework is needed to enable efficient and robust edge event prediction using multi-model learning on resource-constrained edge devices. The goal of this project is developing transformative machine learning and data analytics technologies for enabling AI-based applications on resource-constrained edge computing devices (e.g., IoT devices, AR/VR headsets, and drones). The outcome of this project will advance data analytics and machine learning research of deriving and integrating various high-dimensional sensing data from diverse data sources and building robust predictive models for generic edge computing applications. This project addresses two major problems: 1) the gap between the data complexity and limited computing resources on edge devices and 2) the gap between the robust performance requirement and the multi-dimensional data and complex data modeling from heterogeneous edge devices and environments. The project develops an efficient and robust edge computing framework to provide correctness guarantees on heterogeneous edge computing hardware across different environments. In particular, deep neural network acceleration techniques are designed to enable fine-grained data analytics on resource-constrained commercial-off-the-shelf edge devices. Novel multivariate data analytic models are developed to characterize the unique features of the target event based on high-dimensional sensing data. Such models advance the usage of data science in generic edge sensing tasks that usually suffer from long training times, low prediction accuracy, and ineffective parameter selection. Additionally, the project addresses the challenges arising from the heterogeneity in devices and environments by developing environment-transferable features and models, which enable easy deployment of AI-enabled applications across devices and environments. The project seeks to integrate computer science research with graduate and undergraduate curricula and promote female engineering student involvement. The outcomes will be shared through conferences, journals, and website accessibility.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.