题目：Bayesian scalar-on-image neural network via the soft-thresholded Gaussian process

主讲人：密西根大学生统系 康健副教授

主持人：西南财经大学统计学院 常晋源教授

时间：2019年5月20日（星期一）15:00-16:00

地点：西南财经大学光华校区光华楼1007会议室

报告摘要：

This talk concerns spatial variable selection for scalar-on-image regression. We propose a new class of Bayesian nonparametric models and develop an efficient posterior computational algorithm. The proposed soft-thresholded Gaussian process (STGP) provides large prior support over the class of piecewise-smooth, sparse, and continuous spatially-varying regression coefficient functions.  In addition, under some mild regularity conditions the STGP prior leads to the posterior consistency for parameter estimation and variable selection for scalar-on-image regression, even when the number of predictors is larger than the sample size. As an extension, we apply STGP to a neural network model, which can automatically determine the network architecture.  The proposed method is compared to alternatives via simulation and applied to different imaging datasets. The STGP neural network model may achieve a higher prediction accuracy than other state-of-the-art deep learning methods, especially for the small sample size problem.

主讲人简介：

Jian Kang is an Associate Professor in the Department of Biostatistics at the University of Michigan. He completed his undergraduate studies at Beijing Normal University in 2005 and received MS in mathematics from Tsinghua University in 2007. He obtained PhD in biostatistics at the University of Michigan in 2011.  His main research interests are developing statistical methods and theory for large-scale complex biomedical data analysis with focuses on Bayesian approaches and imaging statistics. He has co-authored over 70 publications in leading statistical journals and medical journals including JASA, Biometrika, NeuroImage and Human Brain Mapping. Dr. Kang was the program chair for statistics in imaging section of the American Statistical Association. He currently serves as the Associate Editor of Biometrics and Statistics in Medicine.