In this paper, we make an important step towards the black-box machine
teaching by considering the cross-space machine teaching, where the teacher and
the learner use different feature representations and the teacher can not fully
observe the learner's model. In such scenario, we study