Her research interests centre around approximate inference in Bayesian non-parametric models like Gaussian processes and Dirichlet processes. Hierarchical extensions of these models offer greater modelling flexibility, provide robust uncertainty estimation and allow one to capture complex interactions between variables typical in real world data. However, this additional flexibility comes at a price, effective inference in hierarchical models is computationally more intensive and exacerbated by the presence of strong correlations in the free variables at different levels of the hierarchy.
Uncertainty quantification in machine learning predictions is becoming increasingly mainstream as several applications in science and industry require statistical guarantees in the predictions generated by models. Bayesian non-parametrics is currently the only paradigm that allows the user to stipulate a prediction in terms of a probability distribution and allow automatic calibration of model complexity.
She is equally interested in the potential of incorporating machine learning in the science of discovery workflows in high energy physics and astronomy.