Abstract:
|
A recent trend of big data problems is to develop computationally efficient inferences that embed computational thinking into traditional uncertainty quantification methods. A particular focus of this talk involves two new classes of nonparametric testing that scales well with massive data. One class is based on randomized sketches which can be implemented in one computer, while another class requires parallel computing. Besides introducing these two new methods, our theoretical contribution is to characterize the minimal computational cost that is needed to achieve the minimax optimal testing power.
|