Advancing mobile and edge computing: Tailored AI acceleration frameworks for future mobility apps
This event is in the past.
1 p.m. to 2 p.m.
Speaker
Xueyu Hou, New Jersey Institute of Technology
Abstract
Over the last ten years, the development of artificial intelligence (AI) has made it increasingly appealing to incorporate AI into both current and new applications across various fields, such as video analytics, autonomous vehicles, unmanned aerial vehicles, and augmented reality. AI holds the promise of enhancing the autonomy, functionality, and quality of these applications. However, the significant costs associated with AI models, due to factors like latency, memory usage, energy consumption, and heat generation, often limit their potential advantages. Traditional methods for accelerating AI models, such as pruning and quantization, involve altering neural network architectures to decrease operational demands, but these adjustments necessitate extensive retraining to regain accuracy and offer only a narrow range of accuracy-latency compromises. In this seminar, I will introduce specialized AI acceleration techniques that are distinct from conventional methods. The first section will focus on AI acceleration frameworks inspired by human vision, tailored for real-time mobile robotics on general mobile edge platforms. By exploring several aspects of human vision perception—such as bifurcated processing, iconic memory, and motion sensitivity, I proposed various effective strategies that mimic the efficient processes of human vision for different robotic platforms, leveraging multi-modal sensors and networked components. The next section will cover AI-enhanced edge-assisted extended reality (XR) systems designed for immersive experiences, which are often hampered by heavy computational demands, leading to excessive power use and overheating. Here, I will demonstrate how edge-assisted XR systems can offer live immersive services by smartly navigating the limitations of computing capacity, network bandwidth, and unpredictable user behaviors. The final section will explore collaborative edge intelligence, focusing on workload management, machine-learning-as-a-service, and trustworthiness, tackling the issues of heterogeneous node capabilities, fluctuating network conditions, and security.
Bio
Xueyu Hou is a final-year Ph.D. student in the Electrical and Computer Engineering Department at the New Jersey Institute of Technology (NJIT), advised by Professor Tao Han. Before coming to NJIT, she obtained her B.S. and M.S. degree in Electrical Engineering from Xi’an Jiaotong University. She was also a student in the Special Class of Gifted Young in Xi’an Jiaotong University. Her current research interests include efficient artificial intelligence, human-centered computing, mobile edge computing, and sustainable computing. Her work has been recognized and published in several prestigious venues across multiple disciplines, including mobile edge computing (e.g., MobiCom, MobiSys, SECON, and InfoCom), distributed processing (e.g., IPDPS and ICPP), and power electronics (e.g., TPE, ECCE, and WiPDA).