CAREER: Pyramidal Intelligence for Ultra-low-power Wearable Massive-sensor Computers

About This Grant

The era of big data is prompting a large-scale deployment of on-body monitors towards wearable massive-sensor computers. These massive sensors have promising and broad prospects to greatly advance big data-driven precision health, through comprehensively capturing behavioral, physiological or biological signals from the human body. Nevertheless, because of the big data volume brought by massive sensors, the system is very power-hungry and thus it is very pressing to innovate an ultra-low-power architecture. Targeting this crucial challenge, this project aims to develop new design methodologies, techniques and implementations to achieve a generalizable ultra-low-power architecture for wearable massive-sensor computers. Concretely, this project seeks to leverage novel deep learning approaches to minimize the power consumption of the system. Firstly, the data characteristics of sensor streams will be learned by deep learning to analyze, evaluate, and measure the redundancy in the data, which will then be used to activate just-enough sensors. The deep learning will learn the signal dynamics to intelligently determine the sensor activation schemes. Besides, the data on the activated sensors will be further analyzed and compressed to minimize the power consumption. The signal fluctuations and patterns will be learned by efficient deep learning models and then be encoded to sparsified representations. Real-world experiments will also be conducted to evaluate and validate the effectiveness of the proposed ultra-low-power architecture. This project will develop a new ultra-low-power architecture to enable energy-efficient wearable massive-sensor computers, and thus greatly advance their real-world deployment. This new architecture will dramatically boost the battery life, enhance the usability, and improve the long-term data capturing capability of the wearable sensors. This is essential for big data-driven precision health. The achieved human big data will effectively contribute to the study of time-varying, nonlinear, and unknown dynamics of the human body, and broadly benefit many areas like fitness and lifestyle management, medical decision support, disease model establishment, individualized treatment plan, and population-level big data mining. The research findings from this project will be broadly disseminated to the scientific communities, medical areas and other communities. The broad impact of this proposal also stems from the educational program for students from K-12 to undergraduate and graduate levels, through efforts like attracting undergraduate, women and underrepresented students to research, training students in real-world problem solving, and broader research training of high school students and outreach to K-12. This systematic plan of integrating education to research aims to train the next generation of professional STEM researchers and engineers. 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.