Accuracy- and resource-aware framework for resource-constrained mobile computing

Mobile computing is one of the largest untapped reservoirs in today’s pervasive computing world as it has the potential to enable a variety of in-situ, real-time applications. However, the domain of mobile computing suffers from the constraints of limited resources such as device battery, CPU, and memory while at the same time users’ expectations in terms of response times, accuracy, and data rates are increasing at a fast pace. As a result, achieving high energy efficiency while maintaining a high quality of service is a crucial challenge. Many of the mobile applications that are pervasive in our lives–such as localization, object/activity recognition, and mobile gaming to name a few–are expected to perform seamlessly with near-instantaneous responses, but are also affected by the same constraints. Current solutions based on offloading computationally-intensive applications from resource-constrained mobile devices to powerful remote computing platforms (such as the Cloud) or nearby mobile devices, suffer from uncertainty in wireless network connectivity or availability of devices in proximity, respectively. To overcome the limitation of current works, the paradigm of approximate computing emerges as a solution to enable resource-intensive mobile applications in resource-constrained environment. Approximate computing reduces the amount of computation that an application is expected to perform, as a result of which the execution time reduces, which in turn reduces the energy consumption of the application. The gain achieved via reduction in energy consumption, however, comes with a potential loss in the accuracy of the results (within acceptable limits). By leveraging approximate computing, we achieve dynamically a tradeoff between accuracy (or optimality of the results produced by an application) and utilization of the available resources (such as battery, CPU cycles, memory, and I/O data rate). The goal of this thesis is to design new techniques so as to enable real-time computation intensive mobile applications in resource-limited and uncertain environments. In order to achieve this goal, we leverage the paradigm of approximate computing and propose the following three solutions. First, approximation at the application level is introduced by joint optimization of algorithm and parameter space of different tasks in the application and a light-weight algorithm is developed that selects the approximated tasks that should be executed to meet the application deadline under uncertainties encountered at run-time. Second, temporal correlation between the continuous stream of frames obtained from the camera sensors is exploited to learn the application parameters that give acceptable accuracy in each frame of the video with significant savings in time and energy. The problem of selecting the algorithm and input parameters for a video is cast as a Markov Decision Process. Third, to reduce the energy consumption of data-intensive applications in distributed camera networks a novel protocol is proposed to identify the camera nodes in the network with correlated multimedia data. Low-computational-complexity metrics are used to quantify the correlation across cameras nodes by using only local knowledge of the network available to the camera nodes. Furthermore, the effectiveness of the proposed approaches is validated through extensive simulations on publicly available datasets and data collected by building multiple end-to-end computationally-intensive applications from the computer vision domain. The proposed innovations in this research will provide novel solutions to the issue of limited resource availability in mobile devices and will foster the development of mobile research community.Ph.D.Includes bibliographical reference