Prediction consistency regularization for Generalized Category Discovery

Yu Duan, Junzhi He, Runxin Zhang, Rong Wang, Xuelong Li, Feiping Nie

Abstract: Highlights•PCR combines the advantages of representation learning and a linear classifier.•PCR ensure consistent predictions in the label space and representation space.•We employ the EM framework to optimize the model with theoretical guarantees.•The experimental results demonstrate the superiority and efficiency of PCR.