UCLA

Classical approaches to machine learning sought to improve the efficiency and accuracy of prediction but often failed to account for the costs associated with the collection of data and expert labels. This shortcoming is particularly limiting in the smart health setting, where accurate classification often requires an invasive level of information querying. Furthermore, in domains such as medical diagnosis, appropriate data should be collected based on a scientific hypothesis, and ground-truth labels may only be provided by highly trained domain experts. Additionally, in many studies, informative features are not scientifically predetermined, and usually, there are many information sources that can be considered as hypothetical relevant features that including all of them is not practical.

In order to address these issues, we suggest novel end-to-end solutions considering different aspects of a real-world learning system. Specifically, we consider feature acquisition, labeling, model training, and prediction at test-time as different aspects of a system that tries to achieve the goal of making accurate predictions efficiently. In this paradigm, information is acquired incrementally based on the value it provides and the cost that should be paid for acquiring it. In this thesis, we explore dynamic and context-aware information acquisition techniques to collect the right piece of information at the right time. Additionally, as inference using incomplete data is an inevitable part of such methods, we propose a novel approach to not only impute missing values but also to capture prediction uncertainties.