Knowledge distillation based on decision boundary instances generated by DBI-GAN

Ziqi Zhu, Xi Liu, Chunhua Deng, Jing Liu, Jixin Zou

Abstract: Highlights•A novel DBI-GAN model is designed to generate high quality augment samples for knowledge distillation.•A new knowledge distillation scheme is proposed based on the DBI-GAN.•A new loss function is proposed for the student network optimization based on the train-set consisting of original samples and the synthesized DBIs.•Our scheme achieves better performance compared with some state-of-the-art methods on benchmark datasets.