The rapid progression of the Internet-of-Things (IoT) has equipped the physical world with capabilities for sensing, computation, and communication, thereby altering human-centered interactions and fostering promising applications. The recent advancements in deep learning have revolutionized numerous fields and are now being adopted in IoT sensing applications. Nonetheless, there still exist some unresolved issues with regards to ubiquitous IoT applications. We identify four primary challenges encountered across IoT systems. Firstly, there is the problem of data heterogeneity due to diverse data collection contexts. Secondly, the labeling burden can be time-consuming. Thirdly, there are privacy concerns surrounding the distributed collection of data. Lastly, there are limited computatio...[ Read more ]

The rapid progression of the Internet-of-Things (IoT) has equipped the physical world with capabilities for sensing, computation, and communication, thereby altering human-centered interactions and fostering promising applications. The recent advancements in deep learning have revolutionized numerous fields and are now being adopted in IoT sensing applications. Nonetheless, there still exist some unresolved issues with regards to ubiquitous IoT applications. We identify four primary challenges encountered across IoT systems. Firstly, there is the problem of data heterogeneity due to diverse data collection contexts. Secondly, the labeling burden can be time-consuming. Thirdly, there are privacy concerns surrounding the distributed collection of data. Lastly, there are limited computation and communication resources for IoT devices. This dissertation elaborates on these four issues and their potential solutions to take a step towards ubiquitous IoT applications.

For the first two issues, we consider how to swiftly adapt the model to unlabeled data with different distributions via unsupervised domain adaptation and how to generalize the model for different data distributions via domain generalization. We present innovative approaches for two specific applications, namely wireless-based gesture recognition and human activity recognition with partial sensor sets. Regarding the privacy issue of distributed collected data and limited resources of IoT devices, one work proposes an efficient and privacy-preserving way to adapt the well pre-trained model to the target data on the edge device without access to the original training data. Another work adopts a federated learning scheme to alleviate computation and communication overhead while safeguarding privacy. The last work considers the network layer for efficient data communication. We design deep learning models for wireless channel prediction, leveraging the stripe features present in the CSI matrix to reduce bandwidth overheads caused by the transmission of large downlink CSI matrix.