Decoupled Optimization for Teacher-Student Semi-Supervised Learning via a Pioneer Student
The Pioneer Student (PiS) proposes an auxiliary branch to decouple optimization issues in Teacher-Student semi-supervised learning. It operates in a separate parameter space to simultaneously solve the Fitting–Generalization Dilemma and the LabelDriven Trap. By periodically transferring its robust knowledge back, PiS improves performance, especially in low-label regimes.