(a) Time-Critical Edge Applications in Autonomous Vehicles (Dr. Kecheng Yang)
With the advances in robotics, autonomous systems have become increasingly pervasive and are able to provide considerable computing power at the edge. Meanwhile, a significant portion of data and applications in such autonomous systems are time-sensitive. This project focuses on the infrastructures and the schedulers in particular in such systems to improve their abilities to support time-critical applications. The Robot Operating System (ROS) has been widely used to develop autonomous systems, and its successor ROS2 was proposed for enhanced real-time support and performance. Students will experience how the applications are implemented on top of ROS or ROS2, investigate the computing resource allocation schemes, and perform system-level optimization to improve the real-time performance of the applications. The outcomes will be evaluated by both simulation and actual implementation on autonomous model cars.
(b) Wearable Sensing System for Personalized Fall Detection at the Edge (Dr. Anne H.H. Ngu)
Fall is a major public health problem leading to loss of life independence, premature mortality, and major hospitalization cost. Wearable devices, especially smartwatches that pair with smartphones, are increasingly a platform of choice for deploying digital health applications. This project focuses on creating a personalized fall detection system on low-cost smartwatches that can detect both “hard” and “soft” falls. In particular, students will experiment with different machine learning personalization strategies, different model validation strategies and scalable mechanism for data collection for personalization. Students will be exposed to skills such as Android programming, machine learning with Python, non-SQL database and edge computing using smartwatches and phones.
(c) Energy Efficient NFV Resource Allocation in Edge Computing Environment (Dr. Xiao Chen)
To reduce cost, deliver flexibility, and accelerate time-to-market, Network Function Virtualization (NFV) has been supported by many network managers. NFV decouples network functions from proprietary hardware and runs them as software applications on general-purpose hardware. Currently, most studies on NFV are done in the cloud environment where resources are substantial. With the development of IoT and 5G, a recent trend is to shift the computing paradigm from the centralized cloud computing to edge computing. Edge computing pushes mobile computing, network control and storage to distributed devices at the network edge to provide server resources, data analysis and artificial intelligence closer to data collection sources. It offers faster data processing, generates less network traffic, and costs less than cloud computing. It is very important in the realization of physical computing, smart cities, ubiquitous computing and IoT. Different from the cloud environment, the physical resources (PRs) at the edge are not abundant. The main challenge for the deployment of NFV at the edge is the NFV-Resource Allocation (NFV-RA) problem. That is, how to allocate PRs to satisfy a set of virtual network function requests (VNFRs). The current NFV-RA solutions seek to minimize the total placement cost, achieve load balancing, and support QoS. NFV is still in its early stage. There are still many important issues that need to be investigated to efficiently manage and allocate resources in NFV-based edge computing. One aspect that was not discussed much is the provisioning of energy-aware strategies to find efficient NFV-RA solutions within reasonable time. In this project, our goal is to study energy-efficient strategies for the NFV-RA problem at the edge.
(d) High-Performance Edge Computing (Dr. Tanzima Islam)
Traditionally, High-Performance Computing (HPC) systems provide many computing cycles to a single simulation that needs a large amount of memory or many cores to solve a significant problem, e.g., COVID-19 vaccine design. However, with the recent surge of edge devices and the volume of data collected, intelligent decision-making on edge depends on quick number crunching and pattern detection, which the computational cycles and high-bandwidth networks of a large-scale system such as HPC can facilitate. The rise of heterogeneous and many-core platforms and the high-bandwidth, low-latency nature of the networks make HPC highly beneficial to the edge computing scenarios. In contrast, the traditional approach of using cloud computing backends suffer from slow data movement problems. While the applicability of HPC to Edge Computing is apparent, there exists little work for analyzing the performance of edge computing workloads on heterogeneous platforms and understanding the performance and throughput bottlenecks of these applications and optimizing the performance of critical edge workflows. During the summer of 2022, we plan to thoroughly investigate different edge computing scenarios using Edge Computing benchmarks on heterogeneous platforms to understand and overcome these performance bottlenecks. The benchmark suite such as Edge AIBench models four typical application scenarios: ICU Patient Monitor, Surveillance Camera, Smart Home, and Autonomous Vehicle.
(e) Semi-supervised Labeling of Sensor-Generated Time-Series Data (Dr. Vangelis Metsis)
The existence of labeled data is paramount for the training of machine learning models. In many cases, data generated by edge computing devices and sensors appear in the form of sequential signals, which are difficult for humans to interpret and accurately label visually. This project aims to create a machine learning-assisted platform that enables the accurate labeling of time-series data with minimal input from humans. Time-series data is any dataset where points are indexed by time. These types of datasets have many important applications, such as human emotion recognition and activity recognition. Researchers in human activity recognition, human-computer interaction, and other related fields often collect live time-series data from multiple sensors to conduct new experiments but run into these common problems: (1) Need to manually label previously collected data at a later stage. (2) Labeling errors mistakenly entering their dataset. (3) The difficulty for humans to visually recognize patterns in signals without proper feature extraction and visualization tools. (4) The significant manual effort required for the manual labeling of even small datasets, which does not scale well as the need for larger amounts of training data increases. The goal of this project is to tackle and address these problems.