Title: Consistent variational Bayes neural networks classification with application to an Alzheimer disease study
Authors: Shrijita Bhattacharya - Michigan State University (United States) [presenting]
Zihuan Liu - Michigan State University (United States)
Taps Maiti - Michigan State University (United States)
Abstract: Bayesian neural networks models (BNN) have re-surged recently due to the advancement of scalable computations and its utility in solving complex prediction problems in applications such as medical image analysis and computer vision tasks. However, the conventional Markov Chain Monte Carlo (MCMC) based implementation suffers from various issues such as computational costs, finding suitable proposal densities, etc. which limit the use of this powerful technique in large scale studies. The variational Bayesinference has become a viable alternative to circumvent some of the computational issues. Although the approach is popular in machine learning, its application in statistics is somewhat limited. A variational BNN estimation methodology and related tactical theory are developed. The numerical algorithms and their practical aspects are discussed in detail. The theory for posterior consistency, a desirable property in nonparametric Bayesian statistics, is also developed. The theory provides an assessment of prediction accuracy and guidelines for characterizing the prior distributions and variational family. The loss of using a variational posterior over the true posterior has also been quantified. The development is motivated by an important application in biomedical engineering, namely building predictive tools for the transition from mild cognitive impairment (MCI) to Alzheimer disease (AD) and emphasizing clinical aspects of the field.