Transductive Zero-Shot Learning with Reliability-Based Pseudo-Label Integration

Zero-shot Learning (ZSL) involves classifying samples from unseen classes using semantic similarities with seen classes. However, ZSL models are often subject to bias toward seen classes, especially in a generalized setting where test samples could be part of either seen or unseen classes. Although sample labels are not available during the training phase, applying transductive learning, which fine-tunes the classifier using unlabeled data, can enhance ZSL model performance in the absence of training labels for unseen classes. The transductive approach can mitigate bias by integrating pseudo-labeled unseen samples during the training phase. Two reliability criteria are proposed in this study for selecting unseen samples to be integrated into the training set with associated pseudo-labels: raw score and margin score. Utilizing a standard ZSL base model, we evaluate both reliability measures across various sample selection strategies, including balanced and unbalanced selections, with differing integration ratios. Our results demonstrate that reliability-based sample selection significantly improves model performance.