Dr. Seung Jun Kim, Department of Computer Science and Electrical Engineering.
Dictionary learning (DL) is a method that approximately decomposes a signal into a linear combination of spare elements. We have seen great success of DL recently. The applications of DL to the multi-subject fMRI data sets estimate sparse neural activation components and the components’ coefficients. The single-task DL takes into consideration a single data set. By exploiting the label subjects of an fMRI data set, the single-task DL can capture the neural activation components that tell the difference between two groups of subjects (healthy control and patients) and those common across groups. In this paper, we propose to apply the supervised multi-task dictionary learning (sMDL) to multiple fMRI data sets under (prior) joint sparse code to force collaborations of the data sets. sMDL not only reveals the difference between groups but also outputs the components commonly shared across data sets and those that are distinct across data sets. The results of the simulation show that the proposed method
works as expected. The results of real fMRI data sets show the method achieves higher prediction accuracy than single-task DL, and more common components are estimated than relative works, such as DS-ICA.