Bilateral Transformation of Biased Pseudo-Labels under Distribution Inconsistency

Ruibing Hou, Hong Chang, Minyang Hu, Bingpeng Ma, Shiguang Shan, Xilin Chen

Abstract: Typically self-training methods assign pseudo-labels by a source model and subsequently iteratively train the model with these pseudo-labeled samples. These approaches are widely employed to leverage vast reserves of unlabeled data, e.g., in semi-supervised learning (SSL) and unsupervised finetuning (UNF). However, our theoretical analysis reveals that the distribution inconsistency between source and unlabeled data could lead to a significant generation error bound for self-training methods. Motivated by this theoretical insight, we present a Bilateral Transformation Self-Training (BTST) learning approach to mitigate the distribution discrepancy and improve the generalization of self-training methods. Firstly, Representation Transformation Module (RTM) is designed to reduce representation distribution discrepancy by bidirectionally transforming high-level representations between source and unlabeled samples. Secondly, Logit Transformation Module (LTM) is designed to reduce class distribution discrepancy by aligning classifier’s predictions with the unlabeled class distribution. Also, LTM incorporates a self-supervised regularization term to estimate unlabeled distribution, theoretically proven to effectively reduce estimation error bound. The two modules work complementary to reduce the generalization bound, ultimately achieving a more generalizable self-training model. Extensive experiments demonstrate that BTST can seamlessly integrate with self-training methods, improving their generalization across various SSL and UNF settings.

External IDs:dblp:journals/ijcv/HouCHMSC26