Towards realizing the internet-of-things vision : in-body, homes, and farms

Abstract

The Internet-of-things (IoT) enables us to connect our physical and digital worlds by embedding computing devices into our environment. Today, there is a huge interest in IoT systems for smart homes, smart cities, digital healthcare, data-driven agriculture, etc. However, for these IoT systems to deliver their intended vision, we need to address two important challenges: (a) operation under limited resources like power and connectivity, (b) operation in spite of extreme heterogeneity in device deployments. In this thesis, we address both these challenges. We design a new communication primitive that allows inaccessible resource-constrained devices like in-body devices to communicate without requiring them to transmit any power of their own. To address heterogeneity, we present two approaches. First, we build a teacher-student model for IoT systems which allows us to train models that can learn to predict one sensor modality from another. This makes IoT systems more robust to failures, enables more accurate inference, and reduces deployment costs. Second, we build a formal model that embeds contextual information about the environment into the inference process and allows heterogenous devices to perform joint inference that is more accurate and robust than either of the devices alone. We demonstrate the efficacy of our approach through end-to-end systems developed for diverse environments with varying constraints on size, power, communication, and sensing modalities: inside the human body, smart homes, and agricultural farms. We deploy these systems for long-term in real world environments and present our insights from these deployments. Finally, we demonstrate that the techniques developed in this thesis have general applicability beyond the application scenarios themselves, for example, in next generation cellular communications.