Recursive Confidence Training for Pseudo-Labeling Calibration in Semi-Supervised Few-Shot Learning

Semi-Supervised Few-Shot Learning (SSFSL) aims to address the data scarcity in few-shot learning by leveraging both a few labeled support data and abundant unlabeled data. In SSFSL, a classifier trained on scarce support data is often biased and thus assigns inaccurate pseudo-labels to the unlabeled data, which will mislead downstream learning tasks. To combat this issue, we introduce a novel method called Certainty-Aware Recursive Confidence Training (CARCT). CARCT hinges on the insight that selecting pseudo-labeled data based on confidence levels can yield more informative support data, which is crucial for retraining an unbiased classifier to achieve accurate pseudo-labeling-a process we term pseudo-labeling calibration. We observe that accurate pseudo-labels typically exhibit smaller certainty entropy, indicating high-confidence pseudo-labeling compared to those of inaccurate pseudo-labels. Accordingly, CARCT constructs a joint double-Gaussian model to fit the certainty entropies collected across numerous SSFSL tasks. Thereby, A semi-supervised Prior Confidence Distribution (ssPCD) is learned to aid in distinguishing between high-confidence and low-confidence pseudo-labels. During an SSFSL task, ssPCD guides the selection of both high-confidence and low-confidence pseudo-labeled data to retrain the classifier that then assigns more accurate pseudo-labels to the low-confidence pseudo-labeled data. Such recursive confidence training continues until the low-confidence ones are exhausted, terminating the pseudo-labeling calibration. The unlabeled data all receive accurate pseudo-labels to expand the few support data to generalize the downstream learning task, which in return meta-refines the classifier, named self-training, to boost the pseudo-labeling in subsequent tasks. Extensive experiments on basic and extended SSFSL setups showcase the superiority of CARCT versus state-of-the-art methods, and comprehensive ablation studies and visualizations justify our insight. The source code is available at https://github.com/Klein-JING/CARCT.