We develop methods for computers to learn from data. These models are intended to not only extract useful patterns from large sets of data, but also to perform well in practice by inferring how these patterns apply to unseen data. The models we pursue are formalized based on probability and information theory, and learning and inference are related aspects of the approach—learning involves adapting the models while inference or perception entails inverting the models.
My current research focuses on a couple issues: How can these models be used for unsupervised learning, in which the aim is to develop useful internal representations of data in situations in which correct labels are not provided? Practical related questions concern semi-supervised and transfer learning. And second, how can the calibrated, expressive formulations offered by probabilistic models be combined with the learning power of deep neural networks?
I am also interested in a number of other topics, including developing learning approaches to visual scene understanding (with a focus on segmentation and attention), and on ethical issues around the practice of machine learning, such as developing fair and privacy-preserving algorithms.