A decoding process for fMRI data is constructed based on Deep Brief Network (DBN) which extracts the feature for classification on each ROI of input fMRI data in order to evaluate robustness for task complexity. The decoding experiment results for hand motion and visual stimulus task show that the results based on DBN in both task can classify the state of subject without the effect of distributions in input voxel values. The decoding process based on the DBN is appropriate for complicate task, which these processes may deal with all voxel values in the selected ROI for each task.