Current Research Projects
Details about the ongoing research projects can be found at the webpage of Pitt Intelligent Systems Lab.
Nowadays, billions of embedded devices and sensors are interconnected towards a global Internet of Things, and AI is hence in great need on these devices to provide intelligent, prompt and adaptive decisions with respect to the heterogeneous and dynamic environmental contexts and user behavior patterns. However, the key barrier of enabling such on-device AI lies in the intrinsic heterogeneity, and strict performance and energy constraints of embedded devices, as well as the unprecedented scale of the surrounding data that brings high computational burden to AI models. On the other hand, the increasing complexity of practical embedded systems and applications also call for tailored designs of AI and ML models that can correctly perceive the system behaviors from sensory data and assist in adaptive and intelligent system operations accordingly.
My research addresses these challenges by answering two fundamental research questions: 1) How real-time AI model inference and training can be supported on extremely weak embedded devices? and 2) How on-device AI models should be designed for specific computer and networked system applications? To answer the first question, my research exploits the fine-grained and explainable knowledge about AI model's execution to adaptively decide the most compact but efficient portion of the model for on-device training and inference. To answer the second question, my work adopts the methodology of modular neural networks and applies the domain knowledge of the specific system application to the neural network module design, to ensure that the neural network models can correctly depict the system characteristics. [Read more...]
Recent technical advances of sensing, computation and communication over mobile and embedded platforms, such as smartphones and wearable computing devices, highlights the possibility of pervasive monitoring and unobtrusive diagnostics of various types of acute or chronic diseases, as convenient yet low-cost alternatives of medical-grade methods without any involvement of clinicians. The goal of this research is to fully unleash the potential of today's mobile computing devices towards accurate, efficient yet cost-effective solutions to mobile and connected health.
More specifically, the ongoing research focuses on the following directions:
- Microphones and speakers available at smartphones and wearable computing devices could be used to emulate a high-resolution sonar system, which could be used to monitor the physiological conditions of different parts of patients' bodies. More specifically, real-time measurements about the patients' chest movement could be used as important indications about the patients' likelihood to various asthma symptoms. Similar measurements could also be used to monitor the humans' airway mechanics, as reliable indicators to evaluate COVID-19 infection out of clinic. [Read more...]
- High-resolution cameras at smartphones could potentially be a useful tool for many biomedical imaging applications. In particular, pairing between smartphones' flashlights and cameras leads to a non-invasive spectroscopy system, in which red and green lights emitted from flashlights are firstly differentiated by optical devices and then used for blood diffusions. Such diffused lights can then be used for various physiological testing purposes, ranging from oximeter to brain imaging.
Internet of Things (IoT) exemplifies future large-scale cyber-physical systems which computerize heterogeneous physical objects and integrating them into a unified networked ecosystem, and also produces a unprecedented amount of data at a large number of distributed and autonomous computing objects. My research aims to address this big data challenge in IoT from a system and networking perspective. The resulted system and network designs are critical to efficient processing of such big data, and are also the key enabler of applying the existing techniques of big data analytics to IoT.
Data processing in IoT is completely different from that in current mobile computing systems and raises new challenges. Many embedded computing objects in IoT, such as implantable sensors or microcontrollers, have very limited computation and battery capacities compared to others like smartphones. This limitation prevents data from being locally processed at these devices where data is produced. Instead, data needs to be remotely processed at other stronger IoT devices or the cloud, but the energy consumed by wirelessly transmitting data has to be minimized. Furthermore, most of IoT applications are sensitive to the data processing delay and hence require ultra-low latency for wireless data transmission between IoT devices and the cloud. Minimizing this latency in IoT, however, is challenging when the large population of IoT devices competes for the limited wireless spectrum resources. My research addresses this challenge by answering two fundamental research questions: 1) How data in IoT should be transmitted for remote processing? and 2) What data in IoT should be remotely processed? [Read more...]
Completed Research Projects
Cloud computing has been considered as a viable solution to bridge the gap between the limited capabilities and increasing workloads of mobile devices. Modern cloud computing services hosted by data centers, however, are incapble of efficiently executing mobile applications due to the excessive network latency accessing data centers and the significant overhead of managing a large number of Virtual Machines (VMs). Existing edge cloud solutions reduce the network latency accessing data centers by deploying servers at the network edge, but cannot handle the peak load from mobile users exceeding the capacity of individual servers.
The goal of this research is to design the edge cloud as a tree hierarchy of geodistributed servers, so as to efficiently handle the peak load and satisfy the requirements of remote program execution in two aspects: i) application performance, and ii) execution expense. First, the performance of remote program execution is improved by optimizing the placement of mobile workloads over edge cloud servers. Second, the capital and operating expenses of the cloud on executing mobile programs are reduced through optimized provisioning of edge cloud servers and dynamic migration of mobile workloads. Third, user mobility between edge cloud servers will be well supported to minimize the additional expense and performance degradation due to user mobility. [Read more...]
Mobile Cloud Computing (MCC) has been widely used to address the resource limitation of mobile devices, by migrating expensive local computations to the remote cloud via wireless networks. However, transmitting data between mobile devices and the cloud also consumes energy. The key problem of MCC, then, becomes how to minimize this energy while preserving the mobile application performance. To deal with this problem, researchers suggested to adaptively partition a mobile application and only migrate the portion that benefits the most for remote execution to the cloud. These traditional solutions, however, are limited to static analysis of mobile program binaries, and have minimal explorations into the run-time characteristics of mobile programs, system operations, and wireless networks. Ignorance of these important characteristics is also the major factor hindering practical integration of mobile devices into the cloud.
The goal of this research, instead, is to design MCC schemes from a new dynamic and network-centric perspective, by taking i) various system and network dynamics, and ii) special characteristics of wireless networks into account. First, we exploit the various critical dynamics in mobile clouds that are indispensable to efficient, prompt, and reliable workload migration between local mobile devices and the remote cloud. Second, we incorporate these special characteristics of wireless networks into the MCC design to ensure that each program state is transmitted to the cloud in the most energy-efficient manner. This new design perspective, which is coupled with both analytical modeling of mobile system behaviors and practical mobile system implementation, will fundamentally improve the energy efficiency of remote program execution in practical MCC systems. [Read more...]
In military operations at the tactical edge, adaptability of mobile communication systems is crucial to ensure warfighters' situational awareness and prompt access on the actionable tactical information, but is not well supported by the current communication systems used in the U.S. Army. We envision that such adaptability could be achieved from an unique social-aware perspective based on properly articulated multi-genre network analysis, which exploits the close coupling between mobile communication networks and human social networks at the tactical edge. This analysis interprets warfighters' situational response to the heterogeneous battlefield contexts as their social dynamics, which are defined as the temporal and spatial variations of social relationship among warfighters and are autonomously characterized from warfighters' contact patterns without manual inputs or configurations.
The proposed research will develop stochastic modeling of various essential characteristics of warfighters' behavior patterns, leading to timely and precise prediction of warfighters' communication needs in the future and designs of adaptive mobile networking strategies at the tactical edge. Three major research thrusts are proposed for this project: (1) Adaptive Contact Prediction: improving the accuracy of contact prediction through formulation of the heterogeneous transient characteristics of warfighters' contact patterns in both temporal and spatial dimensions; (2) Characterization of Social Dynamics: characterizing the social dynamics among warfighters by exploring the correspondence of various sociological concepts in DIL network environments; (3) Adaptive Networking Framework: developing social-aware data dissemination and cooperative caching schemes which autonomously adapt to the social dynamics among warfighters. [Read more]
Information about the frequency deviation and phase angle changes, which are critical for stable and reliable microgrid operations, should be efficiently monitored to ensure secure power management and real-time disturbance response. Traditional solutions to Phasor Measurement Units (PMUs), however, are considered unsatisfactory for microgrid monitoring due to their high installation cost and low accessibility. This research aims to bridge the gap between these traditional systems and the unique requirements of microgrid monitoring by fully unleashing the unexploited capabilities of modern embedded and mobile platforms, including the various sensorboards and smartphones. First, we develop a smartphone-based frequency monitoring system which provides a measurement accuracy of 0.1 mHz at the cost of less than $100. Second, we harvest the time signal from the widely available cellular radio for time synchronization, so as to ensure accurate phase angle measurement without relying on continuous GPS reception. [Read more...]