The proliferation of IoT devices heralds the emergence of intelligent embedded ecosystems that can collectively learn and that interact with humans in a human-like fashion. Recent advances in deep learning revolutionized related fields, such as vision and speech recognition, but the existing techniques remain far from efficient for resource-constrained embedded systems. This dissertation pioneers a broad research agenda on Deep Learning for IoT. By bridging state-of-the-art IoT and deep learning concepts, I hope to enable a future sensor-rich world that is smarter, more dependable, and friendlier, drawing on foundations borrowed from areas as diverse as sensing, embedded systems, machine learning, data mining, and real-time computing. Collectively, this dissertation addresses five research questions related to architecture, performance, predictability and implementation. First, are current deep neural networks fundamentally well-suited for learning from time-series data collected from physical processes, characteristic to IoT applications? If not, what architectural solutions and foundational building blocks are needed? Second, how to reduce the resource consumption of deep learning models such that they can be efficiently deployed on IoT devices or edge servers? Third, how to minimize the human cost of employing deep learning (namely, the cost of data labeling in IoT applications)? Fourth, how to predict uncertainty in deep learning outputs? Finally, how to design deep learning services that meet responsiveness and quality needed for IoT systems? This dissertation elaborates on these core problems and their emerging solutions to help lay a foundation for building IoT systems enriched with effective, efficient, and reliable deep learning models....
Authors: Uma Sai Chaitanya Khandavalli, Ginnaram Nishanth Goud, Shruti Bhargava Choubey.