SPARSE BAYESIAN LEARNING USING VARIATIONAL BAYES INFERENCE BASED ON A GREEDY-BASED CRITERION.

Abstract

We study the problem of finding the sparse signal from a set of compressively sensed measurements using variational Bayes inference. The main focus of this paper is to show that the estimated solution is sensitive to the selection of the parameters of the hyperprior on learning the supports of the solution in our modeling. Selection of such hyperparameters should be made with care, otherwise the solution suffers from the overfitting issues as the number of measurements becomes small. To tackle this issue, we add a greedy criterion which filters out a subset of the estimated supports based on the number of measurements compared to the dimension of the signal of interest.