A semi-supervised classification approach based on restricted Boltzmann machine for fMRI data

Abstract

Due to the limitation in the acquiring and labeling human fMRI data, it is not easy to train an efficient and robust supervised-learning classifier for fMRI data. Semi-supervised learning uses unlabeled data for auxiliary training and combines them into labeled data to construct a better classification model. However, the accuracy of pseudo-labels in conventional semi-supervised approaches may affect the performance of the classification model, which depends on the feature selection of labeled data and the initial classifier trained by labeled data. In this study, we proposed a semi-supervised approach combined unsupervised representation learning and supervised classification. The unsupervised representation learning model was constructed by restricted Boltzmann machine (RBM), and feature parameters from the unsupervised model were provided as constraints in the semi-supervised training of classification model. The results showed a satisfactory feature representation capability and a better performance on multiple classification tasks in fMRI data. The effectiveness of our proposed semi-supervised classification approach may be beneficial to identify complicated brain state under different stimuluses or tasks.