Transferring knowledge from a pretrained model by matching internal representations rather than just final outputs.