Research Interests

My research interests lie at the intersection of information theory and data science. Data science is the science of learning from data (Donoho 2015). Information theory is concerned with the fundamental performance limits of data, i.e., data processing, data communications, and data storage. Combined with applications to healthcare, transportation, and wireless networks, my research thrusts span many disciplines, from mathematics to statistics to computing to engineering to medicine.

Information Theory: In the past fifty years, information theory has been critical in characterizing the fundamental limits of communications and driving innovations in physical layer processing. Moving up the protocol stack, networking research for wireline networks is fairly mature but wireless network research has advanced in a somewhat ad-hoc manner. Just as it did for point-to-point communications, information theory in a network setting would likely reveal the true capability of wireless networks and suggest mechanisms for efficient operation. I work on both network information theory problems and algorithms & protocols for wireless ad-hoc and sensor networks.

Machine Learning: We are working on efficient techniques machine learning algorithms for decision making in healthcare scenarios. In healthcare problems, it is critical to determine what patient data is important in making medical decisions (i.e., this process is called feature selection). We have developed competitive feature selection algorithms based on deep learning that effectively learn compact representations of patient vital signs in an unsupervised manner. This research is in collaboration with clinicians from the National University Hospital (NUH).

We are also investigating how information theoretic ideas can help to understand the fundamental limits and operation of machine learning and deep learning algorithms. I predict that information theory can offer a deeper understanding of why these algorithms work and how we can make them better. Our preliminary work has highlighted the role of mutual information in feature selection.