The student network is learned by feature embedding, and relations between data samples are modeled as probabilistic distribution using feature representations of data. The teacher and student probability distributions are matched by knowledge transfer. (Peng et al. (2019)).