Brain-machine interfaces (BMIs) generate control-commands for specific operations by discriminating signals representing brain functions. In most of current BMIs, machine learning methods are employed as classifier. Their drawback is strong dependency on the individual characteristics of the datasets for each subject. To expand usability of BMIs as more conventional tools, it is desirable that the discriminator is valid even for functional brain data obtained from arbitrary subjects who are not included in the training sample. We proposed Multi-ROI 3D-CNNs such a “generalized” discriminator to classify the motor imagery status using task based fMRI data. As the result, it was suggested that our 3D-CNNs work for unknown subjects' data and finish training faster than our previous classifier (linear Support Vector Machine) did.