Recent advances in machine learning have enabled a wide range of practical applications including active authentication, autonomous driving, and medical diagnosis. While machine learning algorithms achieve impressive performances for these applications, they have to constantly deal with changing characteristics of input data. Examples of such cases include: recognizing faces under poor lighting conditions and side poses while algorithms are trained on well-illuminated faces at the frontal pose; and detecting and segmenting an organ of interest from low-resolution medical images when available algorithms are instead optimized for high-resolution medical images. This problem is commonly known as domain shift. The accuracies of machine learning systems decrease significantly when domain shifts are present. As a result, users must spend significant amounts of time and money to rebuild machine learning models to work well on new data. This project aims to develop computational methods for automatically detecting the presence of domain shifts, quickly adapting machine learning systems to new data distribution, and intelligently seeking additional information to improve the system's performance. Research outputs of this project, such as software, publications, and best practices will contribute to making a wide range of machine learning systems less vulnerable to perpetual changes of input data, and safer to use in the presence of domain shifts.
To achieve these goals, this project proposes four main thrusts: 1) constructing meta-learning techniques to enable efficient adaptation of classifiers to unseen domains using unlabeled data; 2) developing an optimal reinforcement learning strategy for querying additional information that allows effective generalization when the uncertainty is high; 3) building algorithmic foundations for detecting the presence of domain shifts and preparing machine learning systems for appropriate actions; and 4) validating the proposed approaches with classification and segmentation tasks from large-scale datasets corresponding to autonomous driving and active mobile authentication applications. This project combines recent advances in variational inference, reinforcement learning, and active learning to bring a modern and unique perspective on how to deal with the problem of domain shift.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
