The potential for human neuroimaging to read-out the detailed contents of a person's mental state has yet to be fully explored. For fMRI decoding, it is important to choose an appropriate set of voxels (or features) as inputs to the decoder, since the presence of many irrelevant voxels could lead to poor generalization performance, a problem known as overfitting. We applied ARD-based sparse Bayesian algorithm to solve overfitting in fMRI classification. The simulated data demonstrated that sparse logistic regression can select effective features through the weight parameters for each class, and most of the selected features lied in the class they belong to. We observed that sparse logistic regression over pruned some effective features under the condition of 80 features, yet it had limited negative impact on the performance of prediction. The reason is perhaps that these over pruned features have similar values for all classes. On the other hand, it indicated that the selected features contained enough information for classification. Another interesting result is that as more irrelevant features were added to the training set, the performance of C-SVM dropped sharply. This demonstrated that sparse logistic regression can outperform C-SVM when irrelevant features are far more than relevant ones. Despite sparse sparse logistic regression wrongly selected some effective features, it had very limited impact on the prediction performance using both correct and wrong features.