Generalized Zero-Shot Learning with Noisy Labeled Data

Generalized zero-shot learning (GZSL) is a challenging task that aims to classify samples with both seen and unseen categories. In practice, noisy labels can significantly degrade the performance of GZSL classifiers, which has received little attention in previous research. When noisy labeled data exists, the class-level semantic representation may fail to accurately describe some samples of the corresponding class. At the same time, noisy data can also disturb the original distributions of the corresponding classes, leading to estimation errors in modeling the sample distributions. To address these issues, we propose a novel method that aims to alleviate the influence of noisy samples in GZSL. Specifically, we propose a sample-level semantic generation method to ensure an accurate description of the corresponding sample. Furthermore, we introduce an unbalanced learning framework to address the sample distribution estimation with noisy labels to make the estimation error on each class and dimension balanced and dynamically mitigate the negative effects of the error distributions from multiple classes. Experimental results on benchmark datasets demonstrate that our approach effectively mitigates the influence of noisy samples and outperforms other advanced methods.