Appendix: FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling

As introduced in the paper, CPL has its ability of improving performance on those hard-to-learn classes by taking into consider the model's learning status. There are 1000 iterations between every two checkpoints. The results in Table 4 show that our CPL method can dramatically improve the performance of existing SSL algorithms and the FlexMatch achieves the best accuracy. These conclusions are in consistency with the results of Table1 in the main text, showing the effectiveness of our proposed CPL algorithm. As shown in Table 5, we see that in addition to the reduced error rates, CPL also has the best performance on precision, recall, F1 score, and AUC.