Statistics in Deep Learning

In view of the emerging big and possibly heterogeneous data from real world problems and the rapid development of artificial intelligence methods, the institute has formed a new research group of “Robust statistical inference for high dimensional data and deep models” starting this year. The aim of this newly formed group is to develop general-applicable as well as application-specific novel methods and theory for solving real world problems and providing explainable and reproducible results. We will establish robust statistical methods and computationally efficient algorithms of deep models and high-dimensional data.

In methodology, we will develop statistical methods to enhance deep learning models, such as interpretable methods, robust training methods, the use of supervised methods, learning algorithms guided cluster analysis, etc. In practice, we will use real-world big data to test these statistical methods. For example, in marketing science, we can use market big data analysis to formulate the best dynamic pricing strategy, and in medical data, we can use gene and medical imaging big data to increase the clinical significance of biology and medicine. Through robust statistical derivation, it will help the correct use and interpretation of information, and even affect the value of data and future trends in different areas of scientific studies.